Text Mining and Visualization: Case Studies Using Open - X

Text Mining and Visualization: Case Studies Using Open - X
Computer Science & Engineering
“The timing of this book could not be better. It focuses on text mining, text being one of the data sources still to be truly harvested, and on open-source tools
for the analysis and visualization of textual data. … Markus and Andrew have
done an outstanding job bringing together this volume of both introductory and
advanced material about text mining using modern open-source technology in a
highly accessible way.”
—From the Foreword by Professor Dr. Michael Berthold, University of Konstanz,
Germany
Text Mining and Visualization: Case Studies Using Open-Source Tools provides an introduction to text mining using some of the most popular and powerful
open-source tools: KNIME, RapidMiner, Weka, R, and Python.
The contributors—all highly experienced with text mining and open-source software—explain how text data are gathered and processed from a wide variety of
sources, including books, server access logs, websites, social media sites, and
message boards. Each chapter presents a case study that you can follow as part
of a step-by-step, reproducible example. You can also easily apply and extend
the techniques to other problems. All the examples are available on a supplementary website.
The book shows you how to exploit your text data, offering successful application examples and blueprints for you to tackle your text mining tasks and benefit
from open and freely available tools. It gets you up to date on the latest and most
powerful tools, the data mining process, and specific text mining activities.
T ext M ining and V isualization
Chapman & Hall/CRC
Data Mining and Knowledge Discovery Series
Chapman & Hall/CRC
Data Mining and Knowledge Discovery Series
T ext M ining and
V isualization
Case Studies Using
Open-Source Tools
Hofmann
Chisholm
Edited by
Markus Hofmann
Andrew Chisholm
K23176
w w w. c rc p r e s s . c o m
K23176_cover.indd 1
11/3/15 9:18 AM
T ext M ining and
V isualization
Case Studies Using
Open-Source Tools
Chapman & Hall/CRC
Data Mining and Knowledge Discovery Series
SERIES EDITOR
Vipin Kumar
University of Minnesota
Department of Computer Science and Engineering
Minneapolis, Minnesota, U.S.A.
AIMS AND SCOPE
This series aims to capture new developments and applications in data mining and knowledge
discovery, while summarizing the computational tools and techniques useful in data analysis. This
series encourages the integration of mathematical, statistical, and computational methods and
techniques through the publication of a broad range of textbooks, reference works, and handbooks. The inclusion of concrete examples and applications is highly encouraged. The scope of the
series includes, but is not limited to, titles in the areas of data mining and knowledge discovery
methods and applications, modeling, algorithms, theory and foundations, data and knowledge
visualization, data mining systems and tools, and privacy and security issues.
PUBLISHED TITLES
ACCELERATING DISCOVERY : MINING UNSTRUCTURED INFORMATION FOR
HYPOTHESIS GENERATION
Scott Spangler
ADVANCES IN MACHINE LEARNING AND DATA MINING FOR ASTRONOMY
Michael J. Way, Jeffrey D. Scargle, Kamal M. Ali, and Ashok N. Srivastava
BIOLOGICAL DATA MINING
Jake Y. Chen and Stefano Lonardi
COMPUTATIONAL BUSINESS ANALYTICS
Subrata Das
COMPUTATIONAL INTELLIGENT DATA ANALYSIS FOR SUSTAINABLE
DEVELOPMENT
Ting Yu, Nitesh V. Chawla, and Simeon Simoff
COMPUTATIONAL METHODS OF FEATURE SELECTION
Huan Liu and Hiroshi Motoda
CONSTRAINED CLUSTERING: ADVANCES IN ALGORITHMS, THEORY,
AND APPLICATIONS
Sugato Basu, Ian Davidson, and Kiri L. Wagstaff
CONTRAST DATA MINING: CONCEPTS, ALGORITHMS, AND APPLICATIONS
Guozhu Dong and James Bailey
DATA CLASSIFICATION: ALGORITHMS AND APPLICATIONS
Charu C. Aggarawal
DATA CLUSTERING: ALGORITHMS AND APPLICATIONS
Charu C. Aggarawal and Chandan K. Reddy
DATA CLUSTERING IN C++: AN OBJECT-ORIENTED APPROACH
Guojun Gan
DATA MINING FOR DESIGN AND MARKETING
Yukio Ohsawa and Katsutoshi Yada
DATA MINING WITH R: LEARNING WITH CASE STUDIES
Luís Torgo
EVENT MINING: ALGORITHMS AND APPLICATIONS
Tao Li
FOUNDATIONS OF PREDICTIVE ANALYTICS
James Wu and Stephen Coggeshall
GEOGRAPHIC DATA MINING AND KNOWLEDGE DISCOVERY,
SECOND EDITION
Harvey J. Miller and Jiawei Han
GRAPH-BASED SOCIAL MEDIA ANALYSIS
Ioannis Pitas
HANDBOOK OF EDUCATIONAL DATA MINING
Cristóbal Romero, Sebastian Ventura, Mykola Pechenizkiy, and Ryan S.J.d. Baker
HEALTHCARE DATA ANALYTICS
Chandan K. Reddy and Charu C. Aggarwal
INFORMATION DISCOVERY ON ELECTRONIC HEALTH RECORDS
Vagelis Hristidis
INTELLIGENT TECHNOLOGIES FOR WEB APPLICATIONS
Priti Srinivas Sajja and Rajendra Akerkar
INTRODUCTION TO PRIVACY-PRESERVING DATA PUBLISHING: CONCEPTS
AND TECHNIQUES
Benjamin C. M. Fung, Ke Wang, Ada Wai-Chee Fu, and Philip S. Yu
KNOWLEDGE DISCOVERY FOR COUNTERTERRORISM AND
LAW ENFORCEMENT
David Skillicorn
KNOWLEDGE DISCOVERY FROM DATA STREAMS
João Gama
MACHINE LEARNING AND KNOWLEDGE DISCOVERY FOR
ENGINEERING SYSTEMS HEALTH MANAGEMENT
Ashok N. Srivastava and Jiawei Han
MINING SOFTWARE SPECIFICATIONS: METHODOLOGIES AND APPLICATIONS
David Lo, Siau-Cheng Khoo, Jiawei Han, and Chao Liu
MULTIMEDIA DATA MINING: A SYSTEMATIC INTRODUCTION TO
CONCEPTS AND THEORY
Zhongfei Zhang and Ruofei Zhang
MUSIC DATA MINING
Tao Li, Mitsunori Ogihara, and George Tzanetakis
NEXT GENERATION OF DATA MINING
Hillol Kargupta, Jiawei Han, Philip S. Yu, Rajeev Motwani, and Vipin Kumar
RAPIDMINER: DATA MINING USE CASES AND BUSINESS ANALYTICS
APPLICATIONS
Markus Hofmann and Ralf Klinkenberg
RELATIONAL DATA CLUSTERING: MODELS, ALGORITHMS,
AND APPLICATIONS
Bo Long, Zhongfei Zhang, and Philip S. Yu
SERVICE-ORIENTED DISTRIBUTED KNOWLEDGE DISCOVERY
Domenico Talia and Paolo Trunfio
SPECTRAL FEATURE SELECTION FOR DATA MINING
Zheng Alan Zhao and Huan Liu
STATISTICAL DATA MINING USING SAS APPLICATIONS, SECOND EDITION
George Fernandez
SUPPORT VECTOR MACHINES: OPTIMIZATION BASED THEORY,
ALGORITHMS, AND EXTENSIONS
Naiyang Deng, Yingjie Tian, and Chunhua Zhang
TEMPORAL DATA MINING
Theophano Mitsa
TEXT MINING: CLASSIFICATION, CLUSTERING, AND APPLICATIONS
Ashok N. Srivastava and Mehran Sahami
TEXT MINING AND VISUALIZATION: CASE STUDIES USING OPEN-SOURCE
TOOLS
Markus Hofmann and Andrew Chisholm
THE TOP TEN ALGORITHMS IN DATA MINING
Xindong Wu and Vipin Kumar
UNDERSTANDING COMPLEX DATASETS: DATA MINING WITH MATRIX
DECOMPOSITIONS
David Skillicorn
T ext M ining and
V isualization
Case Studies Using
Open-Source Tools
Edited by
Markus Hofmann
Andrew Chisholm
CRC Press
Taylor & Francis Group
6000 Broken Sound Parkway NW, Suite 300
Boca Raton, FL 33487-2742
© 2016 by Taylor & Francis Group, LLC
CRC Press is an imprint of Taylor & Francis Group, an Informa business
No claim to original U.S. Government works
Version Date: 20151105
International Standard Book Number-13: 978-1-4822-3758-0 (eBook - PDF)
This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been
made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright
holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this
form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may
rectify in any future reprint.
Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the
publishers.
For permission to photocopy or use material electronically from this work, please access www.copyright.com (http://
www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923,
978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For
organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged.
Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for
identification and explanation without intent to infringe.
Visit the Taylor & Francis Web site at
http://www.taylorandfrancis.com
and the CRC Press Web site at
http://www.crcpress.com
Dedication - Widmung
Für meine Großeltern, Luise and Matthias Hofmann - Danke für ALLES!
Euer Enkel, Markus
To Jennie
Andrew
vii
This page intentionally left blank
Contents
I
RapidMiner
1
1 RapidMiner for Text Analytic Fundamentals
John Ryan
1.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2
Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1 Education Objective . . . . . . . . . . . . . . . . .
1.2.2 Text Analysis Task Objective . . . . . . . . . . . .
1.3
Tools Used . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4
First Procedure: Building the Corpus . . . . . . . . . . .
1.4.1 Overview . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2 Data Source . . . . . . . . . . . . . . . . . . . . . .
1.4.3 Creating Your Repository . . . . . . . . . . . . . .
1.4.3.1 Download Information from the Internet .
1.5
Second Procedure: Build a Token Repository . . . . . . .
1.5.1 Overview . . . . . . . . . . . . . . . . . . . . . . . .
1.5.2 Retrieving and Extracting Text Information . . . .
1.5.3 Summary . . . . . . . . . . . . . . . . . . . . . . .
1.6
Third Procedure: Analyzing the Corpus . . . . . . . . . .
1.6.1 Overview . . . . . . . . . . . . . . . . . . . . . . . .
1.6.2 Mining Your Repository — Frequency of Words . .
1.6.3 Mining Your Repository — Frequency of N-Grams
1.6.4 Summary . . . . . . . . . . . . . . . . . . . . . . .
1.7
Fourth Procedure: Visualization . . . . . . . . . . . . . .
1.7.1 Overview . . . . . . . . . . . . . . . . . . . . . . . .
1.7.2 Generating word clouds . . . . . . . . . . . . . . . .
1.7.2.1 Visualize Your Data . . . . . . . . . . . .
1.7.3 Summary . . . . . . . . . . . . . . . . . . . . . . .
1.8
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Empirical Zipf-Mandelbrot Variation for Sequential
Documents
Andrew Chisholm
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . .
2.2
Structure of This Chapter . . . . . . . . . . . . . . . .
2.3
Rank–Frequency Distributions . . . . . . . . . . . . .
2.3.1 Heaps’ Law . . . . . . . . . . . . . . . . . . . .
2.3.2 Zipf’s Law . . . . . . . . . . . . . . . . . . . . .
2.3.3 Zipf-Mandelbrot . . . . . . . . . . . . . . . . . .
2.4
Sampling . . . . . . . . . . . . . . . . . . . . . . . . .
2.5
RapidMiner . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1 Creating Rank–Frequency Distributions . . . .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3
4
4
4
4
5
5
5
5
5
11
11
11
17
20
20
20
25
27
27
27
27
27
35
35
Windows within
37
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
37
38
39
39
39
40
41
42
43
ix
x
Contents
2.5.1.1
2.5.1.2
2.6
2.7
2.8
II
Read Document and Create Sequential Windows . . . . .
Create Rank–Frequency Distribution for Whole Document
and Most Common Word List . . . . . . . . . . . . . . . .
2.5.1.3 Calculate Rank–Frequency Distributions for Most Common
Words within Sequential Windows . . . . . . . . . . . . .
2.5.1.4 Combine Whole Document with Sequential Windows . . .
2.5.2 Fitting Zipf-Mandelbrot to a Distribution: Iterate . . . . . . . . . .
2.5.3 Fitting Zipf-Mandelbrot to a Distribution: Fitting . . . . . . . . . .
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.2 Starting Values for Parameters . . . . . . . . . . . . . . . . . . . .
2.6.3 Variation of Zipf-Mandelbrot Parameters by Distance through
Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
KNIME
3 Introduction to the KNIME Text Processing Extension
Kilian Thiel
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Installation . . . . . . . . . . . . . . . . . . . . .
3.2
Philosophy . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Reading Textual Data . . . . . . . . . . . . . . .
3.2.2 Enrichment and Tagging . . . . . . . . . . . . . .
3.2.2.1 Unmodifiabililty . . . . . . . . . . . . .
3.2.2.2 Concurrency . . . . . . . . . . . . . . .
3.2.2.3 Tagger Conflicts . . . . . . . . . . . . .
3.2.3 Preprocessing . . . . . . . . . . . . . . . . . . . .
3.2.3.1 Preprocessing Dialog . . . . . . . . . . .
3.2.4 Frequencies . . . . . . . . . . . . . . . . . . . . .
3.2.5 Transformation . . . . . . . . . . . . . . . . . . .
3.3
Data Types . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Document Cell . . . . . . . . . . . . . . . . . . .
3.3.2 Term Cell . . . . . . . . . . . . . . . . . . . . . .
3.4
Data Table Structures . . . . . . . . . . . . . . . . . . .
3.5
Example Application: Sentiment Classification . . . . .
3.5.1 Reading Textual Data . . . . . . . . . . . . . . .
3.5.2 Preprocessing . . . . . . . . . . . . . . . . . . . .
3.5.3 Transformation . . . . . . . . . . . . . . . . . . .
3.5.4 Classification . . . . . . . . . . . . . . . . . . . .
3.6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . .
43
46
47
47
47
48
52
52
52
53
58
59
61
63
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Social Media Analysis — Text Mining Meets Network Mining
Kilian Thiel, Tobias Kötter, Rosaria Silipo, and Phil Winters
4.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2
The Slashdot Data Set . . . . . . . . . . . . . . . . . . . . . . . .
4.3
Text Mining the Slashdot Data . . . . . . . . . . . . . . . . . . .
4.4
Network Mining the Slashdot Data . . . . . . . . . . . . . . . . .
4.5
Combining Text and Network Mining . . . . . . . . . . . . . . .
4.6
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
63
64
64
66
66
68
68
68
70
70
72
72
73
73
74
74
75
76
76
77
77
79
81
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
81
82
82
87
88
91
Contents
III
xi
Python
93
5 Mining Unstructured User Reviews with Python
Brian Carter
5.1
Introduction . . . . . . . . . . . . . . . . . . . . .
5.1.1 Workflow of the Chapter . . . . . . . . . .
5.1.2 Scope of Instructions . . . . . . . . . . . .
5.1.3 Technical Stack Setup . . . . . . . . . . .
5.1.3.1 Python . . . . . . . . . . . . . .
5.1.3.2 MongoDB . . . . . . . . . . . . .
5.1.4 Python for Text Mining . . . . . . . . . .
5.1.4.1 Sparse Matrices . . . . . . . . . .
5.1.4.2 Text Representation in Python .
5.1.4.3 Character Encoding in Python .
5.2
Web Scraping . . . . . . . . . . . . . . . . . . . .
5.2.1 Regular Expression in Python . . . . . . .
5.2.2 PillReports.net Web Scrape . . . . . . . .
5.3
Data Cleansing . . . . . . . . . . . . . . . . . . .
5.4
Data Visualization and Exploration . . . . . . .
5.4.1 Python:Matplotlib . . . . . . . . . . . . .
5.4.2 Pillreports.net Data Exploration . . . . . .
5.5
Classification . . . . . . . . . . . . . . . . . . . .
5.5.1 Classification with sklearn . . . . . . . . .
5.5.2 Classification Results . . . . . . . . . . . .
5.6
Clustering & PCA . . . . . . . . . . . . . . . . .
5.6.1 Word Cloud . . . . . . . . . . . . . . . . .
5.7
Summary . . . . . . . . . . . . . . . . . . . . . .
95
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
6 Sentiment Classification and Visualization of Product
Alexander Piazza and Pavlina Davcheva
6.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . .
6.2
Process . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1 Step 1: Preparation and Loading of Data Set . .
6.2.2 Step 2: Preprocessing and Feature Extraction .
6.2.3 Step 3: Sentiment Classification . . . . . . . . .
6.2.4 Step 4: Evaluation of the Trained Classifier . .
6.2.5 Step 5: Visualization of Review Data . . . . . .
6.3
Summary . . . . . . . . . . . . . . . . . . . . . . . . .
7 Mining Search Logs for Usage Patterns
Tony Russell-Rose and Paul Clough
7.1
Introduction . . . . . . . . . . . . . . .
7.2
Getting Started . . . . . . . . . . . . .
7.3
Using Clustering to Find Patterns . .
7.3.1 Step 1: Prepare the Data . . . .
7.3.2 Step 2: Cluster Using Weka . .
7.3.3 Step 3: Visualise the Output . .
7.4
Replication and Validation . . . . . .
7.5
Summary . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Review Data
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
95
97
97
99
99
100
100
100
103
105
106
108
108
111
114
115
115
119
120
122
124
128
130
133
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
134
136
137
138
142
143
145
151
153
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
153
154
158
159
159
161
164
168
xii
Contents
8 Temporally Aware Online News Mining and Visualization with
Python
Kyle Goslin
8.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.1 Section Outline . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.2 Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2
What You Need . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.1 Local HTTP Host . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.2 Python Interpreter . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.3 PIP Installer . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.4 lxml and Scrapy . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.5 SigmaJS Visualization . . . . . . . . . . . . . . . . . . . . . . .
8.2.6 News Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.7 Python Code Text Editor . . . . . . . . . . . . . . . . . . . . .
8.3
Crawling and Scraping — What Is the Difference? . . . . . . . . . . .
8.4
What Is Time? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.5
Analyzing Input Sources . . . . . . . . . . . . . . . . . . . . . . . . . .
8.5.1 Parameters for Webpages . . . . . . . . . . . . . . . . . . . . . .
8.5.2 Processing Sections of Webpages vs. Processing Full Webpages .
8.6
Scraping the Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.1 Pseudocode for the Crawler . . . . . . . . . . . . . . . . . . . .
8.6.2 Creating a Basic Scraping Tool . . . . . . . . . . . . . . . . . .
8.6.3 Running a Scraping Tool . . . . . . . . . . . . . . . . . . . . . .
8.6.4 News Story Object . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.5 Custom Parsing Function . . . . . . . . . . . . . . . . . . . . .
8.6.6 Processing an Individual News Story . . . . . . . . . . . . . . .
8.6.7 Python Imports and Global Variables . . . . . . . . . . . . . . .
8.7
Generating a Visual Representation . . . . . . . . . . . . . . . . . . .
8.7.1 Generating JSON Data . . . . . . . . . . . . . . . . . . . . . . .
8.7.2 Writing a JSON File . . . . . . . . . . . . . . . . . . . . . . . .
8.8
Viewing the Visualization . . . . . . . . . . . . . . . . . . . . . . . . .
8.9
Additional Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.9.1 Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.9.2 Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.10 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9 Text Classification Using Python
David Colton
9.1
Introduction . . . . . . . . . . . . . . . . . . . . .
9.1.1 Python . . . . . . . . . . . . . . . . . . . .
9.1.2 The Natural Language Toolkit . . . . . . .
9.1.3 scikit-learn . . . . . . . . . . . . . . . . . .
9.1.4 Verifying Your Environment . . . . . . . .
9.1.5 Movie Review Data Set . . . . . . . . . . .
9.1.6 Precision, Recall, and Accuracy . . . . . .
9.1.7 G-Performance . . . . . . . . . . . . . . .
9.2
Modelling with the Natural Language Toolkit . .
9.2.1 Movie Review Corpus Data Review . . . .
9.2.2 Developing a NLTK Naı̈ve Bayes Classifier
9.2.3 N-Grams . . . . . . . . . . . . . . . . . . .
9.2.4 Stop Words . . . . . . . . . . . . . . . . .
173
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
174
174
174
175
175
175
176
177
177
178
178
179
179
180
182
182
183
184
184
184
185
185
186
188
188
189
191
193
194
196
196
197
199
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
199
200
200
200
201
201
202
203
204
204
205
208
209
Contents
9.2.5 Other Things to Consider . . . . . . . . . . . . .
Modelling with scikit-learn . . . . . . . . . . . . . . . .
9.3.1 Developing a Naı̈ve Bayes scikit-learn Classifier .
9.3.2 Developing a Support Vector Machine scikit-learn
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . .
9.3
9.4
IV
xiii
. . . . . .
. . . . . .
. . . . . .
Classifier
. . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
R
10 Sentiment Analysis of Stock Market Behavior
the R Tool
Nuno Oliveira, Paulo Cortez, and Nelson Areal
10.1 Introduction . . . . . . . . . . . . . . . . . . .
10.2 Methods . . . . . . . . . . . . . . . . . . . . .
10.3 Data Collection . . . . . . . . . . . . . . . . .
10.4 Data Preprocessing . . . . . . . . . . . . . . .
10.5 Sentiment Analysis . . . . . . . . . . . . . . .
10.6 Evaluation . . . . . . . . . . . . . . . . . . .
10.7 Inclusion of Emoticons and Hashtags Features
10.8 Conclusion . . . . . . . . . . . . . . . . . . .
210
212
213
218
219
221
from Twitter Using
223
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
11 Topic Modeling
Patrick Buckley
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.1.1 Latent Dirichlet Allocation (LDA) . . . . . . . . . . . .
11.2 Aims of This Chapter . . . . . . . . . . . . . . . . . . . . . . .
11.2.1 The Data Set . . . . . . . . . . . . . . . . . . . . . . . .
11.2.2 R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3 Loading the Corpus . . . . . . . . . . . . . . . . . . . . . . . .
11.3.1 Import from Local Directory . . . . . . . . . . . . . . . .
11.3.2 Database Connection (RODBC) . . . . . . . . . . . . . .
11.4 Preprocessing the Corpus . . . . . . . . . . . . . . . . . . . . .
11.5 Document Term Matrix (DTM) . . . . . . . . . . . . . . . . . .
11.6 Creating LDA Topic Models . . . . . . . . . . . . . . . . . . . .
11.6.1 Topicmodels Package . . . . . . . . . . . . . . . . . . . .
11.6.2 Determining the Optimum Number of Topics . . . . . .
11.6.3 LDA Models with Gibbs Sampling . . . . . . . . . . . .
11.6.4 Applying New Data to the Model . . . . . . . . . . . . .
11.6.5 Variational Expectations Maximization (VEM) Inference
11.6.6 Comparison of Inference Techniques . . . . . . . . . . . .
11.6.7 Removing Names and Creating an LDA Gibbs Model . .
11.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12 Empirical Analysis of the Stack Overflow Tags Network
Christos Iraklis Tsatsoulis
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . .
12.2 Data Acquisition and Summary Statistics . . . . . . . . .
12.3 Full Graph — Construction and Limited Analysis . . . .
12.4 Reduced Graph — Construction and Macroscopic Analysis
12.5 Node Importance: Centrality Measures . . . . . . . . . . .
12.6 Community Detection . . . . . . . . . . . . . . . . . . . .
12.7 Visualization . . . . . . . . . . . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
224
225
226
227
230
234
236
238
241
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
. . . . . .
Technique
. . . . . .
. . . . . .
. . . . . .
241
242
243
243
243
244
244
245
246
247
249
250
251
255
259
259
261
263
263
265
. . . .
. . . .
. . . .
. . .
. . . .
. . . .
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
265
267
268
272
273
277
282
xiv
Contents
12.8
12.9
Index
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix: Data Acquisition & Parsing . . . . . . . . . . . . . . . . . . . .
285
290
295
Foreword
Data analysis has received a lot of attention in recent years and the newly coined data
scientist is on everybody’s radar. However, in addition to the inherent crop of new buzz
words, two fundamental things have changed. Data analysis now relies on more complex
and heterogeneous data sources; users are no longer content with analyzing a few numbers.
They want to integrate data from different sources, scrutinizing data of diverse types. Almost
more importantly, tool providers and users have realized that no single proprietary software
vendor can provide the wealth of tools required for the job. This has sparked a huge increase
in open-source software used for professional data analysis.
The timing of this book could not be better. It focuses on text mining, text being one
of the data sources still to be truly harvested, and on open-source tools for the analysis and
visualization of textual data. It explores the top-two representatives of two very different
types of tools: programming languages and visual workflow editing environments. R and
Python are now in widespread use and allow experts to program highly versatile code for
sophisticated analytical tasks. At the other end of the spectrum are visual workflow tools
that enable even nonexperts to use predefined templates (or blueprints) and modify analyses.
Using a visual workflow has the added benefit that intuitive documentation and guidance
through the process is created implicitly. RapidMiner (version 5.3, which is still open source)
and KNIME are examples of these types of tools. It is worth noting that especially the
latter stands on the shoulders of giants: KNIME integrates not only R and Python but also
various libraries. (Stanford’s NLP package and the Apache openNLP project, among others,
are examined more closely in the book.) These enable the use of state-of-the-art methods
via an easy-to-use graphical workflow editor.
In a way, the four parts of this book could therefore be read front to back. The reader
starts with a visual workbench, assembling complex analytical workflows. But when a certain
method is missing, the user can draw on the preferred analytical scripting language to access
bleeding-edge technology that has not yet been exposed natively as a visual component. The
reverse order also works. Expert coders can continue to work the way they like to work by
quickly writing efficient code, and at the same time they can wrap their code into visual
components and make that wisdom accessible to nonexperts as well!
Markus and Andrew have done an outstanding job bringing together this volume of
both introductory and advanced material about text mining using modern open source
technology in a highly accessible way.
Prof. Dr. Michael Berthold (University Konstanz, Germany)
This page intentionally left blank
Preface
When people communicate, they do it in lots of ways. They write books and articles, create
blogs and webpages, interact by sending messages in many different ways, and of course
they speak to one another. When this happens electronically, these text data become very
accessible and represent a significant and increasing resource that has tremendous potential
value to a wide range of organisations. This is because text data represent what people are
thinking or feeling and with whom they are interacting, and thus can be used to predict
what people will do, how they are feeling about a particular product or issue, and also who
else in their social group could be similar. The process of extracting value from text data,
known as text mining, is the subject of this book.
There are challenges, of course. In recent years, there has been an undeniable explosion
of text data being produced from a multitude of sources in large volumes and at great speed.
This is within the context of the general huge increases in all forms of data. This volume and
variety require new techniques to be applied to the text data to deal with them effectively.
It is also true that text data by their nature tend to be unstructured, which requires specific
techniques to be adopted to clean and restructure them. Interactions between people leads
to the formation of networks, and to understand and exploit these requires an understanding
of some potentially complex techniques.
It remains true that organisations wishing to exploit text data need new ways of working
to stay ahead and to take advantage of what is available. These include general knowledge of
the latest and most powerful tools, understanding the data mining process, understanding
specific text mining activities, and simply getting an overview of what possibilities there
are.
This book provides an introduction to text mining using some of the most popular and
powerful open-source tools, KNIME, RapidMiner, Weka, R, and Python. In addition, the
Many Eyes website is used to help visualise results. The chapters show text data being
gathered and processed from a wide variety of sources, including books, server-access logs,
websites, social media sites, and message boards. Each chapter within the book is presented
as an example use-case that the reader can follow as part of a step-by-step reproducible
example. In the real world, no two problems are the same, and it would be impossible to
produce a use case example for every one. However, the techniques, once learned, can easily
be applied to other problems and extended. All the examples are downloadable from the
website that accompanies this book and the use of open-source tools ensures that they are
readily accessible. The book’s website is
http://www.text-mining-book.com
Text mining is a subcategory within data mining as a whole, and therefore the chapters
illustrate a number of data mining techniques including supervised learning using classifiers
such as naı̈ve Bayes and support vector machines; cross-validation to estimate model perxvii
xviii
Preface
formance using a variety of performance measures; and unsupervised clustering to partition
data into clusters.
Data mining requires significant preprocessing activities such as cleaning, restructuring, and handling missing values. Text mining also requires these activities particularly
when text data is extracted from webpages. Text mining also introduces new preprocessing
techniques such as tokenizing, stemming, and generation of n-grams. These techniques are
amply illustrated in many of the chapters. In addition some novel techniques for applying
network methods to text data gathered in the context of message websites are shown.
What Is the Structure of This Book, and Which Chapters Should I
Read?
The book consists of four main parts corresponding to the main tools used: RapidMiner,
KNIME, Python, and R.
Part 1 about RapidMiner usage contains two chapters. Chapter 1 is titled “RapidMiner for Text Analytic Fundamentals” and is a practical introduction to the use of various
open-source tools to perform the basic but important preprocessing steps that are usually
necessary when performing any type of text mining exercise. RapidMiner is given particular
focus, but the MySQL database and Many Eyes visualisation website are also used. The specific text corpus that is used consists of the inaugural speeches made by US presidents, and
the objective of the chapter is to preprocess and import these sufficiently to give visibility
to some of the features within them. The speeches themselves are available on the Internet,
and the chapter illustrates how to use RapidMiner to access their locations to download the
content as well as to parse it so that only the text is used. The chapter illustrates storing
the speeches in a database and goes on to show how RapidMiner can be used to perform
tasks like tokenising to eliminate punctuation, numbers, and white space as part of building
a word vector. Stop word removal using both standard English and a custom dictionary
is shown. Creation of word n-grams is also shown as well as techniques for filtering them.
The final part of the chapter shows how the Many Eyes online service can take the output
from the process to visualise it using a word cloud. At all stages, readers are encouraged to
recreate and modify the processes for themselves.
Chapter 2 is more advanced and is titled “Empirical Zipf-Mandelbrot Variation for
Sequential Windows within Documents”. It relates to the important area of authorship
attribution within text mining. This technique is used to determine the author of a piece
of text or sometimes who the author is not. Many attribution techniques exist, and some
are based to a certain extent on departures from Zipf’s law. This law states that the rank
and frequency of common words when multiplied together yield a constant. Clearly this
is a simplification, and the deviations from this for a particular author may reveal a style
representative of the author. Modifications to Zipf’s law have been proposed, one of which
is the Zipf-Mandelbrot law. The deviations from this law may reveal similarities for works
produced by the same author. This chapter uses an advanced RapidMiner process to fit,
using a genetic algorithm approach, works by different authors to Zipf-Mandelbrot models
and determines the deviations to visualize what similarities there are between authors.
Preface
xix
Additionally, an author’s work is randomised to produce a random sampling to determine
how different the actual works are from a random book to show whether the order of words
in a book contributes to an author’s style. The results are visualised using R and show some
evidence that different authors have similarities of style that is not random.
Part 2 of the book describes the use of the Konstanz Information Miner (KNIME)
and again contains two chapters. Chapter 3 introduces the text processing capabilities of
KNIME and is titled “Introduction to the KNIME Text Processing Extension”. KNIME is
a popular open-source platform that uses a visual paradigm to allow processes to be rapidly
assembled and executed to allow all data processing, analysis, and mining problems to be
addressed. The platform has a plug-in architecture that allows extensions to be installed,
and one such is the text processing feature. This chapter describes the installation and use
of this extension as part of a text mining process to predict sentiment of movie reviews. The
aim of the chapter is to give a good introduction to the use of KNIME in the context of this
overall classification process, and readers can use the ideas and techniques for themselves.
The chapter gives more background details about the important preprocessing activities
that are typically undertaken when dealing with text. These include entity recognition
such as the identification of names or other domain-specific items, and tagging parts of
speech to identify nouns, verbs, and so on. An important point that is especially relevant as
data volumes increase is the possibility to perform processing activities in parallel to take
advantage of available processing power, and to reduce the total time to process. Common
preprocessing activities such as stemming, number removal, punctuation, handling small
and stop words that are described in other chapters with other tools can also be performed
with KNIME. The concepts of documents and the bag of words representation are described
and the different types of word or document vectors that can be produced are explained.
These include term frequencies but can use inverse document frequencies if the problem at
hand requires it. Having described the background, the chapter then uses the techniques to
build a classifier to predict positive or negative movie reviews based on available training
data. This shows use of other parts of KNIME to build a classifier on training data, to apply
it to test data, and to observe the accuracy of the prediction.
Chapter 4 is titled “Social Media Analysis — Text Mining Meets Network Mining” and
presents a more advanced use of KNIME with a novel way to combine sentiment of users
with how they are perceived as influencers in the Slashdot online forum. The approach is
motivated by the marketing needs that companies have to identify users with certain traits
and find ways to influence them or address the root causes of their views. With the ever
increasing volume and types of online data, this is a challenge in its own right, which makes
finding something actionable in these fast-moving data sources difficult. The chapter has
two parts that combine to produce the result. First, a process is described that gathers
user reviews from the Slashdot forum to yield an attitude score for each user. This score
is the difference between positive and negative words, which is derived from a lexicon, the
MPQA subjectivity lexicon in this case, although others could be substituted as the domain
problem dictates. As part of an exploratory confirmation, a tag cloud of words used by an
individual user is also drawn where negative and positive words are rendered in different
colours. The second part of the chapter uses network analysis to find users who are termed
leaders and those who are followers. A leader is one whose published articles gain more
comments from others, whereas a follower is one who tends to comment more. This is done
in KNIME by using the HITS algorithm often used to rate webpages. In this case, users take
the place of websites, and authorities become equivalent to leaders and hubs followers. The
two different views are then combined to determine the characteristics of leaders compared
with followers from an attitude perspective. The result is that leaders tend to score more
xx
Preface
highly on attitude; that is, they are more positive. This contradicts the normal marketing
wisdom that negative sentiment tends to be more important.
Part 3 contains five chapters that focus on a wide variety of use cases. Chapter 5 is titled
“Mining Unstructured User Reviews with Python” and gives a detailed worked example of
mining another social media site where reviews of drugs are posted by users. The site,
pillreports.com, does not condone the use of drugs but provides a service to alert users to
potentially life-threatening problems found by real users. The reviews are generally short
text entries and are often tagged with a good or bad review. This allows for classification
models to be built to try and predict the review in cases where none is provided. In addition,
an exploratory clustering is performed on the review data to determine if there are features
of interest. The chapter is intended to be illustrative of the techniques and tools that can be
used and starts with the process of gathering the data from the Pill Reports website. Python
is used to navigate and select the relevant text for storage in a MongoDb datastore. It is
the nature of Web scraping that it is very specific to a site and can be fairly involved; the
techniques shown will therefore be applicable to other sites. The cleaning and restructuring
activities that are required are illustrated with worked examples using Python, including
reformatting dates, removing white space, stripping out HTML tags, renaming columns, and
generation of n-grams. As a precursor to the classification task to aid understanding of the
data, certain visualisation and exploration activities are described. The Python Matplotlib
package is used to visualise results, and examples are given. The importance of restructuring
the data using grouping and aggregation techniques to get the best out of the visualisations
is stressed with details to help. Moving on to the classification step, simple classifiers are
built to predict the positive or negative reviews. The initial results are improved through
feature selection, and the top terms that predict the class are shown. This is very typical of
the sorts of activities that are undertaken during text mining and classification in general,
and the techniques will therefore be reusable in other contexts. The final step is to cluster
the reviews to determine if there is some unseen structure of interest. This is done using a
combination of k-means clustering and principal component analysis. Visualising the results
allows a user to see if there are patterns of interest.
Chapter 6 titled “Sentiment Classification and Visualization of Product Review Data” is
about using text data gathered from website consumer reviews of products to build a model
that can predict sentiment. The difficult problem of obtaining training data is addressed
by using the star ratings generally given to products as a proxy for whether the product
is good or bad. The motivation for this is to allow companies to assess how well particular
products are being received in the market. The chapter aims to give worked examples
with a focus on illustrating the end-to-end process rather than the specific accuracy of
the techniques tried. Having said that, however, accuracies in excess of 80 percent are
achieved for certain product categories. The chapter makes extensive use of Python with
the NumPy, NLTK, and Scipy packages, and includes detailed worked examples. As with
all data mining activities, extensive data preparation is required, and the chapter illustrates
the important steps required. These include, importing correctly from webpages to ensure
only valid text is used, tokenizing to find words used in unigrams or bigrams, removal of stop
words and punctuation, and stemming and changing emoticons to text form. The chapter
then illustrates production of classification models to determine if the extracted features
can predict the sentiment expressed from the star rating. The classification models produce
interesting results, but to go further and understand what contributes to the positive and
negative sentiment, the chapter also gives examples using the open-source Many Eyes tool
to show different visualisations and perspectives on the data. This would be valuable for
product vendors wanting to gain insight into the reviews of their products.
Preface
xxi
Chapter 7 “Mining Search Logs for Usage Patterns” is about mining transaction logs
containing information about the details of searches users have performed and shows how
unsupervised clustering can be performed to identify different types of user. The insights
could help to drive services and applications of the future. Given the assumption that what
a user searches for is a good indication of his or her intent, the chapter draws together
some of the important contributions in this area and proceeds with an example process
to show this working in a real context. The specific data that are processed are search
transaction data from AOL, and the starting point is to extract a small number of features
of interest. These are suggested from similar works, and the first step is to process the
logs to represent the data with these features. This is done using Python, and examples
are given. The open-source tool Weka is then used to perform an unsupervised clustering
using expectation maximization to yield a candidate “best” clustering. As with all clustering
techniques and validity measures, the presented answer is not necessarily the best in terms
of fit to the problem domain. However, there is value because it allows the user to focus
and use intelligent reasoning to understand what the result is showing and what additional
steps would be needed to improve the model. This is done in the chapter where results are
considered, alternative features are considered and different processing is performed with
the end result that a more convincing case is made for the final answer. On the way, the
importance of visualising the results, repeating to check that the results are repeatable, and
being sceptical are underlined. The particular end result is of interest, but more importantly,
it is the process that has been followed that gives the result more power. Generally speaking,
this chapter supports the view that a process approach that is iterative in nature is the way
to achieve strong results.
Chapter 8, “Temporally Aware Online News Mining and Visualization with Python”,
discusses how some sources of text data such as newsfeeds or reviews can have more significance if the information is more recent. With this in mind, this chapter introduces time into
text mining. The chapter contains very detailed instructions on how to crawl and scrape
data from the Google news aggregation service. This is a well-structured website containing
time-tagged news items. All sites are different, and the specific instructions for different
sites would naturally be different; the instructions in the chapter would need to be varied
for these. Detailed instructions for the Google site are given, and this, of necessity, drills
into detail about the structure of HTML pages and how to navigate through them. The
heavy lifting is done using the Python packages “scrapy” and “BeautifulSoup”, but some
details relating to use of XPath are also covered. There are many different ways to store
timestamp information. This is a problem, and the chapter describes how conversion to a
common format can be achieved. Visualizing results is key, and the use of the open-source
SigmaJS package is described.
Chapter 9, “Text Classification Using Python”, uses Python together with a number of
packages to show how these can be used to classify movie reviews using different classification models. The Natural Language Toolkit (NLTK) package provides libraries to perform
various processing activities such as parsing, tokenising, and stemming of text data. This is
used in conjunction with the Scikit package, which provides more advanced text processing
capabilities such as TF-IDF to create word vectors from movie review data. The data set
contains positive and negative reviews, and supervised models are built and their performance checked using library capabilities from the Scikit learn package. Having performed
an initial basic analysis, a more sophisticated approach using word n-grams is adopted to
yield improvements in performance. Further improvements are seen with the removal of
stop words. The general approach taken is illustrative of the normal method adopted when
performing such investigations.
xxii
Preface
Part 4 contains three chapters using R. Chapter 10, titled “Sentiment Analysis of Stock
Market Behavior from Twitter Using the R Tool”, describes sentiment analysis of Twitter
messages applied to the prediction of stock market behaviour. The chapter compares how
well manually labelled data is predicted using various unsupervised lexical-based sentiment
models or by using supervised machine learning techniques. The conclusion is that supervised techniques are superior, but in the absence of labelled training data, which is generally
difficult to obtain, the unsupervised techniques have a part to play. The chapter uses R and
well illustrates how most data mining is about cleaning and restructuring data. The chapter
includes practical examples that are normally seen during text mining, including removal
of numbers, removal of punctuation, stemming, forcing to lowercase, elimination of stop
words, and pruning to remove frequent terms.
Chapter 11, titled “Topic Modeling”, relates to topic modeling as a way to understand
the essential characteristics of some text data. Mining text documents usually causes vast
amounts of data to be created. When representing many documents as rows, it is not unusual
to have tens of thousands of dimensions corresponding to words. When considering bigrams,
the number of dimensions can rise even more significantly. Such huge data sets can present
considerable challenges in terms of time to process. Clearly, there is value in anything that
can reduce the number of dimensions to a significantly smaller number while retaining the
essential characteristics of it so that it can be used in typical data mining activities. This
chapter is about topic modeling, which is one relatively new technique that shows promise
to address this issue. The basic assumption behind this technique is that documents contain
a probabilistic mixture of topics, and each topic itself contains a distribution of words. The
generation of a document can be conceived of as the selection of a topic from one of the
available ones and from there randomly select a word. Proceed word by word until the
document is complete. The reverse process, namely, finding the optimum topics based on a
document, is what this chapter concerns itself with. The chapter makes extensive use of R
and in particular the “topicmodels” package and has ‘worked examples to allow the reader
to replicate the details. As with many text mining activities, the first step is to read and
preprocess the data. This involves stemming, stop word removal, removal of numbers and
punctuation, and forcing to lowercase. Determination of the optimum number of topics is a
trial and error process and an important consideration is the amount of pruning necessary to
strike a balance between frequent and rare words. The chapter then proceeds with the detail
of finding topic models, and advanced techniques are shown based on use of the topicmodels
package. The determination of the optimum number of topics still requires trial and error,
and visualisation approaches are shown to facilitate this.
Chapter 12 titled “Empirical Analysis of the Stack Overflow Tags Network”, presents
a new angle on exploring text data using network graphs where a graph in this context
means the mathematical construct of vertices connected with edges. The specific text data
to be explored is from Stack Overflow. This website contains questions and answers tagged
with mandatory topics. The approach within the chapter is to use the mandatory topic
tags as vertices on a graph and to connect these with edges to represent whether the tags
appear in the same question. The more often pairs of tags appear in questions, the larger
the weight of the edge between the vertices corresponding to the tags. This seemingly
simple approach leads to new insights into how tags relate to one another. The chapter
uses worked R examples with the igraph package and gives a good introductory overview
of some important concepts in graph exploration that this package provides. These include
whether the graph is globally connected, what clusters it contains, node degree as a proxy
for importance, and various clustering coefficients and path lengths to show that the graph
differs from random and therefore contains significant information. The chapter goes on to
Preface
xxiii
show how to reduce the graph while trying to retain interesting information and using certain
node importance measures such as betweenness and closeness to give insights into tags. The
interesting problem of community detection is also illustrated. Methods to visualise the data
are also shown since these, too, can give new insights. The aim of the chapter is to expose
the reader to the whole area of graphs and to give ideas for their use in other domains.
The worked examples using Stack Overflow data serve as an easy-to-understand domain to
make the explanations easier to follow.
This page intentionally left blank
About the Editors
Markus Hofmann
Dr. Markus Hofmann is currently a lecturer at the Institute of Technology Blanchardstown, Ireland, where he focuses on the areas of data mining, text mining, data exploration and visualisation, and business intelligence. He holds a PhD from Trinity College
Dublin, an MSc in Computing (Information Technology for Strategic Management) from
the Dublin Institute of Technology, and a BA in Information Management Systems. He
has taught extensively at the undergraduate and postgraduate levels in the fields of data
mining, information retrieval, text/web mining, data mining applications, data preprocessing and exploration, and databases. Dr. Hofmann has published widely at national as well
as international level and specialised in recent years in the areas of data mining, learning
object creation, and virtual learning environments. Further, he has strong connections to
the business intelligence and data mining sectors, on both academic and industry levels.
Dr. Hofmann has worked as a technology expert together with 20 different organisations in
recent years for companies such as Intel. Most of his involvement was on the innovation side
of technology services and for products where his contributions had significant impact on
the success of such projects. He is a member of the Register of Expert Panellists of the Irish
Higher Education and Training Awards council, external examiner to two other third-level
institutes, and a specialist in undergraduate and postgraduate course development. He has
been an internal and external examiner of postgraduate thesis submissions. He also has
been a local and technical chair of national and international conferences.
Andrew Chisholm
Andrew Chisholm holds an MA in Physics from Oxford University and over a long career has been a software developer, systems integrator, project manager, solution architect,
customer-facing presales consultant, and strategic consultant. Most recently, he has been a
product manager creating profitable test and measurement solutions for communication service providers. A lifelong interest in data came to fruition with the completion of a masters
degree in business intelligence and data mining from the Institute of Technology, Blanchardstown, Ireland. Since then he has become a certified RapidMiner Master (with official
number 7, which pads nicely to 007) and has published papers, a book chapter relating to
the practical use of RapidMiner for unsupervised clustering and has authored a book titled
Exploring Data with RapidMiner. Recently, he has collaborated with Dr. Hofmann to create
both basic and advanced RapidMiner video training content for RapidMinerResources.com.
In his current role, he is now combining domain knowledge of the telecommunications inxxv
xxvi
About the Editors
dustry with data science principles and practical hands-on work to help customers exploit
the data produced by their solutions. He fully expects data to be where the fun will be.
List of Contributors
Editors
• Markus Hofmann, Institute of Technology Blanchardstown, Ireland
• Andrew Chisholm, Information Gain Ltd., UK
Chapter Authors
• Nelson Areal, Department of Management, University of Minho, Braga, Portugal
• Patrick Buckley, Institute of Technology, Blanchardstown, Ireland
• Brian Carter, IBM Analytics, Dublin, Ireland
• Andrew Chisholm, Information Gain Ltd., UK
• David Colton, IBM, Dublin, Ireland
• Paul Clough, Information School, University of Sheffield, UK
• Paulo Cortez, ALGORITMI Research Centre/Department of Information Systems,
University of Minho, Guimarães, Portugal
• Pavlina Davcheva, Chair of Information Systems II, Institute of Information Systems, Friedrich-Alexander-University Erlangen-Nuremberg, Germany
• Kyle Goslin, Department of Computer Science, College of Computing Technology,
Dublin, Ireland
• Tobias Kötter, KNIME.com, Berlin, Germany
• Nuno Oliveira, ALGORITMI Research Centre, University of Minho, Guimarães,
Portugal
• Alexander Piazza, Chair of Information Systems II, Institute of Information Systems, Friedrich-Alexander-University Erlangen-Nuremberg, Germany
• Tony Russell-Rose, UXLabs, UK
• John Ryan, Blanchardstown Institute of Technology, Dublin, Ireland
• Rosaria Silipo, KNIME.com, Zurich, Switzerland
• Kilian Thiel, KNIME.com, Berlin, Germany
• Christos Iraklis Tsatsoulis, Nodalpoint Systems, Athens, Greece
• Phil Winters, KNIME.com, Zurich, Switzerland
xxvii
This page intentionally left blank
Acknowledgments
Many people have contributed to making this book and the underlying open-source software
solutions a reality. We are thankful to all of you.
We would like to thank the contributing authors of this book, who shared their experience in the chapters and who thereby enable others to have a quick and successful
text mining start with open-source tools, providing successful application examples and
blueprints for the readers to tackle their text mining tasks and benefit from the strength of
using open and freely available tools.
Many thanks to Dr. Brian Nolan, Head of School of Informatics, Institute of Technology
Blanchardstown (ITB); and Dr. Anthony Keane, Head of Department of Informatics, ITB
for continuously supporting projects such as this one.
Many thanks also to our families. MH: A special thanks goes to Glenda, Killian, Darragh, Daniel, SiSi, and Judy for making my life fun; My parents, Gertrud and Karl-Heinz
Hofmann, for continuously supporting my endeavours. Also a huge thank you to Hans
Trautwein and Heidi Krauss for introducing me to computers and my first data related
application, MultiPlan, in 1986. AC: To my parents for making it possible and to my wife
for keeping it possible.
The entire team of the Taylor & Francis Group was very professional, responsive, and
always helpful in guiding us through this project. Should any of you readers consider publishing a book, we can highly recommend this publisher.
Open-source projects grow strong with their community. We are thankful to all contributors, particularly, text analysis — related open source-tools and all supporters of these
open-source projects. We are grateful not only for source code contributions, community
support in the forum, and bug reports and fixes but also for those who spread the word
with their blogs, videos, and word of mouth.
With best regards and appreciation to all contributors,
Dr. Markus Hofmann, Institute of Technology Blanchardstown, Dublin, Ireland
Andrew Chisholm, Information Gain Ltd., UK
xxix
This page intentionally left blank
List of Figures
1.1
Creating your repository: Overall process. . . . . . . . . . . . . . . . . . .
6
1.2
Creating your repository: Step A – Get Page operator. . . . . . . . . . . .
7
1.3
Creating your repository: Step B. . . . . . . . . . . . . . . . . . . . . . . .
8
1.4
Creating your repository: Step B – vector window.
. . . . . . . . . . . . .
8
1.5
Creating your repository: Step B – Cut Document operator. . . . . . . . .
9
1.6
Creating your repository: Step B – Extract Information operator. . . . . .
9
1.7
Creating your repository: Step C – create attributes. . . . . . . . . . . . .
9
1.8
Creating your repository: Step C – attribute name. . . . . . . . . . . . . .
10
1.9
Creating your repository: Step D. . . . . . . . . . . . . . . . . . . . . . . .
10
1.10
Creating your repository: Step E – Write Excel operator. . . . . . . . . . .
10
1.11
Build a token repository: Process I. . . . . . . . . . . . . . . . . . . . . . .
11
1.12
Build a token repository: Process I Step A – Read Excel operator. . . . . .
12
1.13
Build a token repository: Process I Step B – Get Pages operator. . . . . .
12
1.14
Build a token repository: Process I Step C – Process Documents operator.
13
1.15
Build a token repository: Process I Step C – vector window. . . . . . . . .
13
1.16
Build a token repository: Process I Step C – Extract Information operator.
13
1.17
Build a token repository: Process I Step C – extracting date. . . . . . . . .
14
1.18
Build a token repository: Process I Step C – extracting president’s name. .
14
1.19
Build a token repository: Process I Step C – extracting regular expression.
15
1.20
Build a token repository: Process I Step C – regular region. . . . . . . . .
16
1.21
Build a token repository: Process I Step C – cutting speech content. . . . .
16
1.22
Build a token repository: Process I Step C – cut document nodes. . . . . .
16
xxxi
xxxii
List of Figures
1.23
Build a token repository: Process I Step D – store repository.
. . . . . . .
16
1.24
Build a token repository: Process II. . . . . . . . . . . . . . . . . . . . . . .
17
1.25
Build a token repository: Process II Step A – Retrieve operator. . . . . . .
17
1.26
Build a token repository: Process II Step B – Rename operator. . . . . . .
18
1.27
Build a token repository: Process II Step B – additional attributes. . . . .
18
1.28
Build a token repository: Process II Step C – Select Attributes operator. .
18
1.29
Build a token repository: Process II Step C – subset attributes. . . . . . .
19
1.30
Build a token repository: Process II Step D – Write Database operator. . .
19
1.31
Analyzing the corpus: Process I. . . . . . . . . . . . . . . . . . . . . . . . .
20
1.32
Analyzing the corpus: Process I Step A – Read Database operator. . . . .
21
1.33
Analyzing the corpus: Process I Step A – SQL. . . . . . . . . . . . . . . .
21
1.34
Analyzing the corpus: Process I Step B – term occurrences. . . . . . . . .
22
1.35
Analyzing the corpus: Process I Step B – vector creation window. . . . . .
22
1.36
Analyzing the corpus: Process I Step B – Extract Content operator. . . . .
23
1.37
Analyzing the corpus: Process I Step B – Tokenize operator. . . . . . . . .
23
1.38
Analyzing the corpus: Process I Step B – Filter Stopwords (English)
operator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
1.39
Analyzing the corpus: Process I Step B – custom dictionary. . . . . . . . .
24
1.40
Analyzing the corpus: Process I Step C – Write Excel operator. . . . . . .
24
1.41
Analyzing the corpus: Process I Step C – transposed report. . . . . . . . .
25
1.42
Analyzing the corpus: Process II Step B. . . . . . . . . . . . . . . . . . . .
26
1.43
Analyzing the corpus: Process II Step B – Generate n-Grams (Terms)
operator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
Analyzing the corpus: Process II Step B – Filter Tokens (by Content)
operator. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
26
1.45
Analyzing the corpus: Process II Step C. . . . . . . . . . . . . . . . . . . .
26
1.46
Visualization: Layout of transposed worksheet. . . . . . . . . . . . . . . . .
28
1.47
Visualization: Wordle menu. . . . . . . . . . . . . . . . . . . . . . . . . . .
29
1.48
Visualization: Copy and paste data into create section. . . . . . . . . . . .
30
1.44
List of Figures
xxxiii
1.49
Visualization: Speech represented as a word cloud. . . . . . . . . . . . . . .
30
1.50
Visualization: Word cloud layout. . . . . . . . . . . . . . . . . . . . . . . .
31
1.51
Visualization: Word cloud colour. . . . . . . . . . . . . . . . . . . . . . . .
31
1.52
Visualization: Word cloud token limit. . . . . . . . . . . . . . . . . . . . .
32
1.53
Visualization: Word cloud font types. . . . . . . . . . . . . . . . . . . . . .
32
1.54
Visualization: Word cloud remove tokens (filtered to 20 tokens). . . . . . .
33
1.55
Visualization: Word cloud filter option (remove token). . . . . . . . . . . .
33
1.56
Visualization: Word cloud filtered (specific tokens removed). . . . . . . . .
34
1.57
Visualization: Word cloud options (print, save, new window, randomize). .
34
1.58
Visualization: Layout of transposed bigram worksheet. . . . . . . . . . . .
34
1.59
Visualization: Word cloud bigrams. . . . . . . . . . . . . . . . . . . . . . .
35
1.60
Visualization: Word cloud bigrams filtered (removal of tokens). . . . . . . .
35
2.1
Observed variation for the word “the” for consecutive 5,000-word windows
within the novel Moby Dick. . . . . . . . . . . . . . . . . . . . . . . . . . .
42
2.2
RapidMiner process to calculate word frequencies. . . . . . . . . . . . . . .
43
2.3
Process Section A within Figure 2.2. . . . . . . . . . . . . . . . . . . . . .
44
2.4
Process Section B within Figure 2.2. . . . . . . . . . . . . . . . . . . . . .
44
2.5
Process Section C within Figure 2.2. . . . . . . . . . . . . . . . . . . . . .
45
2.6
Process Section D within Figure 2.2. . . . . . . . . . . . . . . . . . . . . .
45
2.7
RapidMiner process to execute process for all attributes to fit ZipfMandelbrot distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
48
2.8
Detail of RapidMiner process to execute Zipf-Mandelbrot distribution fit. .
48
2.9
RapidMiner process to fit Zipf-Mandelbrot distribution. . . . . . . . . . . .
49
2.10
Configuration for Optimize Parameters (Evolutionary) operator. . . . . . .
50
2.11
Details for Optimize Parameters (Evolutionary) operator. . . . . . . . . . .
50
2.12
Details for macro-generation workaround to pass numerical parameters to
Optimize Parameters operator. . . . . . . . . . . . . . . . . . . . . . . . . .
51
Calculation of Zipf-Mandelbrot probability and error from known
probability. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
51
2.13
xxxiv
2.14
2.15
2.16
2.17
2.18
2.19
2.20
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
List of Figures
Log of probability and estimated probability as a function of log rank for
the 100 most common words within all of Pride and Prejudice . . . . . . .
53
Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Moby Dick . . . . . . . . . . . . . . . . . .
54
Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within The Piazza Tales . . . . . . . . . . . . . . .
55
Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Sense and Sensibility . . . . . . . . . . . .
56
Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Mansfield Park . . . . . . . . . . . . . . . .
56
Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within The Return of Sherlock Holmes . . . . . . .
57
Zipf-Mandelbrot Scatter plot for A and C parameters for random samples
and sequential windows within The Adventures of Sherlock Holmes . . . .
57
An example workflow illustrating the basic philosophy and order of KNIME
text processing nodes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
65
A data table with a column containing document cells. The documents are
reviews of Italian restaurants in San Francisco. . . . . . . . . . . . . . . . .
67
A column of a data table containing term cells. The terms have been assigned
POS tags (tag values and tag types). . . . . . . . . . . . . . . . . . . . . .
68
Dialog of the OpenNLP NE Tagger node. The first checkbox allows for
specification as to whether or not the named entities should be flagged
unmodifiable. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
69
Dialog of the OpenNLP NE Tagger node. The number of parallel threads
to use for tagging can be specified here. . . . . . . . . . . . . . . . . . . . .
69
Typical chain of preprocessing nodes to remove punctuation marks, numbers, very small words, stop words, conversion to lowercase, and stemming.
70
The Preprocessing tab of the Stop word Filter node. Deep preprocessing is
applied, original documents are appended, and unmodifiable terms are not
filtered. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
Bag-of-words data table with one term column and two documents columns.
The column, “Orig Document” contains original documents. The “Document” column contains preprocessed documents. . . . . . . . . . . . . . . .
72
Bag-of-words data table with an additional column with absolute term
frequencies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
73
List of Figures
3.10
xxxv
Document vectors of 10 documents. The documents are stored in the leftmost column. The other columns represent the terms of the whole set of
documents, one for each unique term. . . . . . . . . . . . . . . . . . . . . .
75
3.11
Chain of preprocessing nodes of the Preprocessing meta node. . . . . . . .
76
3.12
Chain of preprocessing nodes inside the Preprocessing meta node. . . . . .
77
3.13
Confusion matrix and accuracy scores of the sentiment decision tree model.
78
3.14
ROC curve of the sentiment decision tree model. . . . . . . . . . . . . . . .
78
4.1
The text mining workflow used to compute the sentiment score for each user. 84
4.2
Distribution of the level of attitude λ by user, with −20 as minimum attitude
and 50 as maximum attitude. . . . . . . . . . . . . . . . . . . . . . . . . .
84
Scatter plot of frequency of negative words vs. frequency of positive words
for all users. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
85
4.4
Tag cloud of user “dada21”. . . . . . . . . . . . . . . . . . . . . . . . . . .
86
4.5
Tag cloud of user “pNutz”. . . . . . . . . . . . . . . . . . . . . . . . . . . .
86
4.6
Example of a network extracted from Slashdot where vertices represent
users, and edges comments. . . . . . . . . . . . . . . . . . . . . . . . . . .
87
4.7
Scatter plot of leader vs. follower score for all users. . . . . . . . . . . . . .
88
4.8
KNIME workflow that combines text and network mining. . . . . . . . . .
90
4.9
Leader vs. follower score colored by attitude for all users. Users with a
positive attitude are marked green, users with a negative attitude red. . .
91
5.1
Pillreports.net standard report. . . . . . . . . . . . . . . . . . . . . . . . .
96
5.2
Code: Check Python setup. . . . . . . . . . . . . . . . . . . . . . . . . . . .
99
5.3
Code: Creating a sparse matrix in Python. . . . . . . . . . . . . . . . . . .
102
5.4
Code: Confusing output with text in Python. . . . . . . . . . . . . . . . .
103
5.5
Original byte encoding — 256 characters. . . . . . . . . . . . . . . . . . . .
104
5.6
Unicode encoding paradigm. . . . . . . . . . . . . . . . . . . . . . . . . . .
104
5.7
Character bytes and code points. . . . . . . . . . . . . . . . . . . . . . . .
104
5.8
Code: Python encoding for text correct use. . . . . . . . . . . . . . . . . .
105
5.9
Code: Scraping webpages. . . . . . . . . . . . . . . . . . . . . . . . . . . .
107
5.10
Code: Connecting to a database in MongoDB. . . . . . . . . . . . . . . . .
107
4.3
xxxvi
List of Figures
5.11
Code: Regular expressions. . . . . . . . . . . . . . . . . . . . . . . . . . . .
109
5.12
Code: Setting scraping parameters. . . . . . . . . . . . . . . . . . . . . . .
110
5.13
Code: Lambda method. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
112
5.14
Code: DateTime library, Apply & Lambda methods. . . . . . . . . . .
112
5.15
Simple matplotlib subplot example. . . . . . . . . . . . . . . . . . . . . . .
115
5.16
Code: matplotlib subplots. . . . . . . . . . . . . . . . . . . . . . . . . . . .
116
5.17
Weekly count of reports. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
117
5.18
Warning: Column cross-tabulated with Country: column. . . . . . . . . . .
117
5.19
Code: Weekly count of reports submitted. . . . . . . . . . . . . . . . . . .
118
5.20
String length of Description: column. . . . . . . . . . . . . . . . . . . . . .
120
5.21
Code: Setting up vectorizer and models for classification. . . . . . . . . . .
121
5.22
Code: Splitting into train and test sets. . . . . . . . . . . . . . . . . . . . .
121
5.23
Code: sklearn pipeline. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
122
5.24
Code: Model metrics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
122
5.25
Code: Feature selection.
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
123
5.26
Scatter plot of top predictive features. . . . . . . . . . . . . . . . . . . . .
125
5.27
Code: Clustering and PCA models. . . . . . . . . . . . . . . . . . . . . . .
126
5.28
Principal components scatter plot. . . . . . . . . . . . . . . . . . . . . . . .
127
5.29
Code: Tagging words using nltk library. . . . . . . . . . . . . . . . . . . .
128
5.30
Code: Counting word frequency with the collections module. . . . . . . . .
129
5.31
User report: Word cloud. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
129
6.1
Sentiment classification and visualization process. . . . . . . . . . . . . . .
135
6.2
Word cloud for the positive and negative reviews of the mobilephone
category. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
148
6.3
Jigsaw’s welcome screen. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
148
6.4
Import screen in Jigsaw. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
149
6.5
Entity identification screen in Jigsaw. . . . . . . . . . . . . . . . . . . . . .
149
6.6
Word tree view for “screen”. . . . . . . . . . . . . . . . . . . . . . . . . . .
150
List of Figures
xxxvii
6.7
Word tree view for “screen is”. . . . . . . . . . . . . . . . . . . . . . . . . .
151
7.1
A sample of records from the AOL log. . . . . . . . . . . . . . . . . . . . .
155
7.2
A sample from the AOL log divided into sessions. . . . . . . . . . . . . . .
156
7.3
A set of feature vectors from the AOL log. . . . . . . . . . . . . . . . . . .
158
7.4
The Weka GUI chooser.
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
159
7.5
Loading the data into Weka. . . . . . . . . . . . . . . . . . . . . . . . . . .
160
7.6
Configuring the EM algorithm.
. . . . . . . . . . . . . . . . . . . . . . . .
160
7.7
Configuring the visualization. . . . . . . . . . . . . . . . . . . . . . . . . .
161
7.8
100,000 AOL sessions, plotted as queries vs. clicks. . . . . . . . . . . . . .
162
7.9
Four clusters based on six features. . . . . . . . . . . . . . . . . . . . . . .
162
7.10
Three clusters based on seven features. . . . . . . . . . . . . . . . . . . . .
163
7.11
Applying EM using Wolfram et al.’s 6 features to 10,000 sessions from AOL. 165
7.12
Applying EM using Wolfram et al.’s 6 features to 100,000 sessions from AOL. 166
7.13
Applying XMeans (k <= 10) and Wolfram et al.’s 6 features to 100,000
sessions from AOL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
166
Applying EM and Wolfram et al.’s 6 features to 100,000 filtered sessions
from AOL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
167
7.15
Sum of squared errors by k for 100,000 filtered sessions from AOL. . . . .
168
7.16
Applying kMeans (k = 4) and Wolfram et al.’s 6 features to 100,000 sessions
from AOL. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
169
8.1
Windows command prompt. . . . . . . . . . . . . . . . . . . . . . . . . . .
176
8.2
Contents of the SigmaJS folder. . . . . . . . . . . . . . . . . . . . . . . . .
178
8.3
XAMPP control panel. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
193
8.4
News stories. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
195
8.5
Closer view of news stories.
. . . . . . . . . . . . . . . . . . . . . . . . . .
195
9.1
Verifying your Python environment on a Windows machine. . . . . . . . .
201
9.2
Using the NLTK built-in NLTK Downloader tool to retrieve the movie
review corpus. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
202
7.14
xxxviii
List of Figures
9.3
The structure of the movie review corpus. . . . . . . . . . . . . . . . . . .
204
9.4
A sample positive review from the movie corpus. . . . . . . . . . . . . . . .
205
9.5
Performance values of the first NLTK model developed. . . . . . . . . . . .
207
9.6
Performance of NLTK model using bigrams. . . . . . . . . . . . . . . . . .
209
9.7
Performance of NLTK model using trigrams. . . . . . . . . . . . . . . . . .
210
9.8
Performance of NLTK model using the prepare review function. . . . . . .
214
9.9
Performance of the first Naı̈ve Bayes scikit-learn model. . . . . . . . . . . .
217
9.10
Naı̈ve Bayes scikit-learn most informative features. . . . . . . . . . . . . .
218
9.11
Performance of the SVM scikit-learn model. . . . . . . . . . . . . . . . . .
218
9.12
SVM scikit-learn model most informative features. . . . . . . . . . . . . . .
219
11.1
News extract topics. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
242
11.2
Word-cloud of the DTM. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
250
11.3
Cross-validation — optimum number of topics. . . . . . . . . . . . . . . . .
254
11.4
Cross-validation — optimum number of topics (2 to 5). . . . . . . . . . . .
255
11.5
Term distribution. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
258
12.1
Tag frequency distribution for the first 100 tags. . . . . . . . . . . . . . . .
269
12.2
Communities revealed by Infomap, with corresponding sizes. . . . . . . . .
280
12.3
Visualization of our communities graph. . . . . . . . . . . . . . . . . . . .
286
12.4
Visualization of the “R” community. . . . . . . . . . . . . . . . . . . . . .
287
12.5
The “R” community, with the r node itself removed. . . . . . . . . . . . .
288
12.6
The “Big Data” and “Machine Learning” communities (excluding these
terms themselves). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
289
List of Tables
1.1
Creating your repository: Step E – report snapshot I. . . . . . . . . . . . .
10
1.2
Creating your repository: Step E – report snapshot II. . . . . . . . . . . .
10
1.3
Creating your repository: Step E – report snapshot III. . . . . . . . . . . .
11
1.4
Speechstop.txt content. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
2.1
Variation of z-score for the most common words in sequential 5,000-word
windows for the novel Moby Dick. . . . . . . . . . . . . . . . . . . . . . . .
41
2.2
RapidMiner processes and sections where they are described. . . . . . . . .
43
2.3
Process sections for RapidMiner process to calculate rank–frequency
distributions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
46
2.4
Details of texts used in this chapter.
. . . . . . . . . . . . . . . . . . . . .
52
2.5
Details of parameter ranges used in this chapter. . . . . . . . . . . . . . . .
52
5.1
Column descriptions and statistics. . . . . . . . . . . . . . . . . . . . . . .
98
5.2
Example dense matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
101
5.3
Example sparse matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
101
5.4
Geo-coding State/Province: column. . . . . . . . . . . . . . . . . . . . . . .
113
5.5
Summary of Country: and Language: columns. . . . . . . . . . . . . . . . .
114
5.6
Summary of language prediction confidence. . . . . . . . . . . . . . . . . .
114
5.7
Suspected contents (SC Category:) and Warning: label. . . . . . . . . . . .
118
5.8
User Report: string length grouped by Country: and Warning:. . . . . . .
119
5.9
Top 5 models: binary classification on Warning: column. . . . . . . . . . .
123
5.10
Classification accuracy using feature selection. . . . . . . . . . . . . . . . .
124
6.1
Table of used libraries. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
136
xxxix
xl
List of Tables
6.2
Number of reviews per product category. . . . . . . . . . . . . . . . . . . .
136
6.3
Tokenization example. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
140
6.4
Example results for information degree per feature. . . . . . . . . . . . . .
143
6.5
Structure of confusion matrices in case of binary classification. . . . . . . .
144
6.6
Sentiment classification accuracies per product category. . . . . . . . . . .
145
6.7
Resulting average confusion matrix for the mobile phone category. . . . . .
145
8.1
Time windows.
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
180
9.1
Performance comparison of unigrams, bigrams, and trigrams when stop
words have been removed. . . . . . . . . . . . . . . . . . . . . . . . . . . .
211
9.2
Confusion matrix for a Naı̈ve Bayes scikit-learn classifier. . . . . . . . . . .
218
10.1
Examples of sentiment analysis according to each model (pos = positive,
neg = negative, neu = neutral). . . . . . . . . . . . . . . . . . . . . . . . .
234
10.2
Sentiment analysis results (in %, best values in bold). . . . . . . . . . . .
236
10.3
Sentiment analysis results excluding messages without lexicon items (in %,
best values in bold). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
236
Sentiment analysis results with emoticons and hashtags features (in %, best
values in bold). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
238
11.1
Topic document distributions. . . . . . . . . . . . . . . . . . . . . . . . . .
259
12.1
Statistics of raw data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
267
10.4
Part I
RapidMiner
1
This page intentionally left blank
Chapter 1
RapidMiner for Text Analytic Fundamentals
John Ryan
Blanchardstown Institute of Technology, Dublin, Ireland
1.1
1.2
1.3
1.4
1.5
1.6
1.7
1.8
1.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.1
Education Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2
Text Analysis Task Objective . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tools Used . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
First Procedure: Building the Corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.1
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.2
Data Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.3
Creating Your Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4.3.1
Download Information from the Internet . . . . . . . . . . . . . . . .
Second Procedure: Build a Token Repository . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.1
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.2
Retrieving and Extracting Text Information . . . . . . . . . . . . . . . . . . . . . . . . . .
1.5.3
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Third Procedure: Analyzing the Corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.1
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.6.2
Mining Your Repository — Frequency of Words . . . . . . . . . . . . . . . . . . . . . .
1.6.3
Mining Your Repository — Frequency of N-Grams . . . . . . . . . . . . . . . . . . .
1.6.4
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fourth Procedure: Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.1
Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.2
Generating word clouds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.2.1
Visualize Your Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.7.3
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3
3
4
4
4
5
5
5
5
5
11
11
11
17
20
20
20
23
25
27
27
27
27
35
35
35
Introduction
The objectives of this chapter are twofold: to introduce you to the fundamentals of
building a text-based repository; and to explain how you can apply text mining techniques
to documents, such as speeches, in order to discover any valuable insights and patterns that
may be contained within their content. This chapter is strictly focused towards a novice or
start-up level; thus, techniques and processes are explained in a more detailed and visual
manner than would otherwise be necessary.
3
4
1.2
Text Mining and Visualization: Case Studies Using Open-Source Tools
Objectives
There are two objectives that should be achieved by the end of this chapter: an education
objective and a text analysis objective. Every text analysis research needs to have an aim
or a goal; this then informs the researcher on the types of data sources they should seek,
the tools (and data cleaning) required, and the types of patterns they should be aspiring to
uncover.
1.2.1
Education Objective
After this chapter has been read, you should have, at the very minimum, a foundationlevel knowledge of how to build a text corpus and to be able to apply some basic, but
fundamental, text analysis to that corpus and subsequently visualise the results, all using
open-source software.
1.2.2
Text Analysis Task Objective
In this chapter you, the student, are applying text analysis to American Presidential
Inaugural Speeches [1] in order to summarise and therefore compare the most frequently
used words in those speeches. Thus, without having to actually read the documents or
listen to them online, you (or your companions and colleagues) should be able to gain an
understanding of the key terms by quickly glancing at the visual output.
1.3
Tools Used
You should be familiar with the database software MySQL (knowing a GUI-based application, such as HeidiSQL, would be beneficial), data analytics suite RapidMiner 5.3, online
visualization tool Wordle (http://www.wordle.net/), and spreadsheets in order to accomplish the tasks within this chapter.
Note: Ensure to have the following plug-ins installed within your copy of RapidMiner:
Information Extraction, Text Processing, and Web Mining. The Information Extraction
extension provides operators to perform preprocessing of the text-based corpus in order
to facilitate specific extractions of text-based data, such as fragments within sentences
[2]. The Text Processing extension provides all necessary text mining operators to fully
extrapolate and statistically review text-type data within a corpus (e.g., reducing words
to their stem, adding stopword filtering). The Web Mining extension operates in tandem
with the Text Processing extension; thus, this needs to be installed with it. The Web
Mining extension provides operators (such as Get Pages) in which you can retrieve online
based information, for example, by means of web scraping. All of these functions are
comprehensibly explained and demonstrated in Sections 1.4.3 through 1.6.3.
Note: Create a MySQL database called inauguralspeeches. Do not create any tables within
it, as they will be generated automatically within RapidMiner processes (detailed in Section 1.5). This database will be used to store the required speech-related attributes.
RapidMiner for Text Analytic Fundamentals
5
Note: Create a new local repository within RapidMiner called Chapter. This will be used
to store the required processes detailed within this chapter; particularly referenced in
Sections 1.4.3 through 1.6.3.
1.4
First Procedure: Building the Corpus
1.4.1
Overview
The overall objective is to build a corpus that can be mined in later sections. So in order
to do this, you need to obtain a list of functioning URL address links that can be eventually
downloaded, their content mined, analyzed, and produced in visual reports. This section
concentrates on that initial step: creating a viable list of URLs that can be downloaded. In
order to accomplish this task, you require each speech’s unique reference tag to be gathered
(labelled as pid in the address URL structure) and assigned to a fixed URL address to
facilitate appropriate online access to each of the presidential speeches.
1.4.2
Data Source
The source to achieve the text mining objectives of this chapter (outlined in Section 1.2)
is The American Presidency Project website [1], in particular the web page that contains indexes to all U.S. presidents inaugural speeches: http://www.presidency.ucsb.edu/
inaugurals.php.
The initial task for any text mining analyst is to understand the corpus. This particular
web page lists speeches you can read to gain an appreciation of the corpus’s context. This
web page also lists the approximate number of words beside each speech; this is valuable
information as it allows you to gauge the quantity of the tokens in your analysis. To understand the structure of the HTML content, you should also select View Source for the web
document. This knowledge is essential for your extraction text mining model as it informs
you of the types of retrieval methods that will be necessary to collect the fragments of data
required for your analysis by viewing the HTML structure of where those data items are
located.
The language in all the presidential speeches is English, though the terminology and
verbal construction change over the different periods. All speeches are stored, from April
30, 1789 (George Washington) up until the present day (the current American president).
1.4.3
1.4.3.1
Creating Your Repository
Download Information from the Internet
The first procedure in building your repository is solely related to gathering a list of URLs
with unique identifiers (represented by the variable pid in the URL address) that distinctly
tag each speech; these IDs allow for the identification and extraction of the speech’s content
in Section 1.5.
In order to achieve this, a number of RapidMiner operators are required, and are referenced in Figure 1.1.
6
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.1: Creating your repository: Overall process.
Note: The alphabetical steps that label each RapidMiner process (A, B, C...) in the
relevant operator based Figures are outlined by the steps in the appropriate sections.
They are organized in alphabetical order.
Note: It is critical to name each of your processes in a manner that allows you to refer back
to them over time and enables you to instantly recognise what they represent. Text analysis
can create numerous processes due to the type of content involved and outputs required.
Throughout each procedure in this chapter, a text process name will be suggested.
Note: Save the following process within the local repository Chapter as Step1ObtainPresidentUrls.
Step A (Figure 1.2): In order to discover what the range of unique identifiers
are for the speeches, you first have to download and extract information from http:
//www.presidency.ucsb.edu/inaugurals.php, which contains these values within
its HTML content code. The Get Page operator is used to retrieve these values, whose
HTML content contains the necessary identifiers that reference all the inaugural US presidential speeches, for example:
• http://www.presidency.ucsb.edu/ws/index.php?pid=46366
• http://www.presidency.ucsb.edu/ws/index.php?pid=44
In summary, Get Page retrieves the web page via HTTP, and produces the information
as an HTML source code document. This document can then be sliced and mined for the
reference numbers that you wish to retrieve; it should be noted from viewing the source
code that these are in the HTML structure format of
<tr>
<td nowrap class="ver12">
<strong>Grover Cleveland - I </strong>
</td>
<td align="right" nowrap bgcolor="#FFFFFF" class="ver12">
<p>
RapidMiner for Text Analytic Fundamentals
7
FIGURE 1.2: Creating your repository: Step A – Get Page operator.
<a href="http://www.presidency.ucsb.edu/ws/index.php?pid=25824">
March 4, 1885
</a>
</p>
</td>
<td width="80" align="center" bgcolor="#FFFFFF"
class="ver12">1686</td>
</tr>
Step B (Figure 1.3): The element where you can extract the reference number is
pid=25824"
To achieve this extraction, you next select the Process Documents operator within RapidMiner, and select a Cut Document process within that operator’s Vector Creation window
(Figure 1.4).
The Cut Document operator slices the HTML source into segments using specific criteria that can be based on a range of query type mechanisms including String Matching,
Regular Region, Regular Expression; for full reference of these terms, select the query type
drop-down menu within the Cut Document. It should be noted that Regular Region and
Regular Expression are explained further within Section 1.5.
Note: The core difference between the Cut Document operator and the Extract Information operator is essentially that the cut document mechanism slices a document, such
as a speech, with a beginning and end tag, usually stipulated by regular expressions,
whereas the Extract Information operator extracts information anywhere within a structured document. So you could often cut a document and then, within the cut operator’s
segment-processing window, extract information from that sliced portion.
8
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.3: Creating your repository: Step B.
FIGURE 1.4: Creating your repository: Step B – vector window.
Note regarding query type within both Cut Document and Extract Information operators:
• String Matching: Specifies a beginning and ending for a string search and extracts
anything that is between those two parameters.
• Regular Region: This query type is essentially a user-friendly way of adding 2 regular
expressions to stipulate a beginning tag, and an end tag in order to retrieve what is
between those two parameters.
• Regular Expression: These are expressions used to match a group and extract appropriate information. They require detailed set-up and knowledge of various command
words and symbol functions. They tend to be more flexible than other query types.
For this example, String Matching is used as it is deemed the most efficient, where
the query expression states to parse the text string from ”http://www.presidency.
ucsb.edu/ws/index.php?pid” to ”>” (Figure 1.5). This means that for http:
//www.presidency.ucsb.edu/ws/index.php?pid=25824”¿, the segmented source
code would be sliced into fragments of reference numericals, for example =25824.
The next process is to extract those numeric values using an Extract Information process
that is nested within the Cut Document operator’s Segment Processing window, previously
outlined in Figure 1.4. This operator extracts the required information from the structured
content for each reference number from all the inaugural speeches. Using String Matching
again, an attribute, labelled URL is created with a query expression stating, extract values
from between = and ” (Figure 1.6). Thus, the value 25824 is retrieved from the source text.
RapidMiner for Text Analytic Fundamentals
9
FIGURE 1.5: Creating your repository: Step B – Cut Document operator.
FIGURE 1.6: Creating your repository: Step B – Extract Information operator.
Step C (Figure 1.7): The next step is to create an attribute that contains the address
link for each speech in which you wish to crawl and analyze. Select the Generate Attributes
operator within RapidMiner and label it as in Figure 1.8. The content of the attribute
should be http://www.presidency.ucsb.edu/ws/index.php?pid. The attribute
should be titled LINK (Figure 1.8).
Step D (Figure 1.9): You now want to join the extracted reference number contained
within URL with the HTTP link LINK to create a full URL address for each speech and
store this list within a spreadsheet. The Generate Concatenation operator achieves this
requirement by joining these two attributes LINK and URL, with an = sign, as can be
referenced from the following output example: http://www.presidency.ucsb.edu/
ws/index.php?pid=25824.
Step E (Figure 1.10): Finally, the Write Excel operator allows you to store the
generated speech address URLs within a spreadsheet; it should be noted that you should
ensure that the file format is XLS.
Ultimately you should have a spreadsheet report, saved as getlist.xls, that resembles
Table 1.1, Table 1.2, and Table 1.3.
FIGURE 1.7: Creating your repository: Step C – create attributes.
10
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.8: Creating your repository: Step C – attribute name.
FIGURE 1.9: Creating your repository: Step D.
FIGURE 1.10: Creating your repository: Step E – Write Excel operator.
TABLE 1.1: Creating your repository: Step E – report snapshot I.
text
URL
=25800”
25800
ResponseCode
200
ResponseMessage
OK
ContentType
text/HTML
TABLE 1.2: Creating your repository: Step E – report snapshot II.
ContentLength
Date
2014-03-31
14:33:25
LastModified
Expires
RapidMiner for Text Analytic Fundamentals
11
TABLE 1.3: Creating your repository: Step E – report snapshot III.
query key
SCT
LINK
http://www.presidency.
ucsb.edu/ws/index.php?
pid
LINK=URL
http://www.presidency.
ucsb.edu/ws/index.php?
pid=25800
FIGURE 1.11: Build a token repository: Process I.
1.5
1.5.1
Second Procedure: Build a Token Repository
Overview
The main requirement of this section is to collect each speech’s core relevant details,
extracting key information for new attributes such as the president’s name, the speech’s
date, and the content of the speech itself. This extracted information is then stored within
a database repository once successfully retrieved.
The required RapidMiner processes are referenced in Figure 1.11.
1.5.2
Retrieving and Extracting Text Information
Note: Save the following process within the local repository Chapter as Step2RetrieveallSpeeches
Step A (Figure 1.12): Now that the spreadsheet getlist.xls is created and stored from Section 1.4.3, you can actively import and mine the HTML structure of the inaugural speeches.
The first process is to access the spreadsheet using the Read Excel operator. Click on the
Import Configuration Wizard button as displayed in Figure 1.12. Select your file getlist.xls
and click the [Next] button. There are four steps within this process; press [Next] until you
get to the fourth step and click [Finish].
Step B (Figure 1.13): So now that you have a connection to the spreadsheet containing
the URLs, you use the Get Pages operator to retrieve the speeches using the list of speech
HTML addresses. This operator should not be confused with the Get Page operator within
Section 1.4.3, which retrieves just the one page. Get Pages works by examining the attribute
Link=URL and using the data value (i.e., the Web URL links for the speeches) within this
12
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.12: Build a token repository: Process I Step A – Read Excel operator.
FIGURE 1.13: Build a token repository: Process I Step B – Get Pages operator.
attribute to determine the example set and thus the extraction of appropriate information.
Thus, the spreadsheet data itself contained within Link=URL determines and therefore
controls the information that is retrieved for future analysis when using the Get Pages
operator. It functions within a loop by going through each row of the attribute Link=URL
and selecting the next value of that attribute until there are no more available to be called.
Ensure that the link attribute is Link=URL as in Figure 1.13.
Step C (Figure 1.14): You now need to extract the key elements of the president’s
name, date of speech, and the speech itself from the raw source document for each inaugural
speech. Select the Process Documents from Data operator and have the following processes
within its Vector Creation window (referenced in Figure 1.15).
• Extract Information operator (Figure 1.16): The year of the inaugural speech is within
<span class=”docdate” >January 20, 1981 <\span >. Use a Regular Expression to
extract the core element by stating
((?<=docdate">).*?(?=<))
Essentially what this regular expression statement means is that you want to extract
RapidMiner for Text Analytic Fundamentals
13
FIGURE 1.14: Build a token repository: Process I Step C – Process Documents operator.
FIGURE 1.15: Build a token repository: Process I Step C – vector window.
anything that is between docdate” > and < (Figure 1.17). Thus, in this case that text
string would be January 20, 1981. Label the attribute as Inaugural Date.
• Extract Information operator: The president’s name is within <title >Ronald Reagan:
Inaugural Address <\title >. So similar to the previous extraction, use a regular
expression (Figure 1.18)
((?<=title>).*?(?=:))
In this case take anything that is between title > and :; retrieving Ronald Reagan for
this example. Label the attribute as President.
FIGURE 1.16: Build a token repository: Process I Step C – Extract Information operator.
14
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.17: Build a token repository: Process I Step C – extracting date.
FIGURE 1.18: Build a token repository: Process I Step C – extracting president’s name.
Note: Regular Expression contains many symbols and formats (some not very intuitive) to
enable you to write an appropriate text retrieval expression. There are lots of variations of
regular expressions to achieve the token extraction. In the case of the president’s name, the
most flexible method you could use is Grouping and Capturing the required information;
in particular using a variation of this category, broadly labelled as Lookaround grouping,
where the matching only keeps the specific token and not the criteria parameters. In this
case, ?<=title >is considered as a look behind ; it informs the system to match anything
that is matching after title>. In this case that includes .* meaning all arbitrary characters following from the <title >tag. This matching is narrowed as a further criterion
for matching is included: a positive lookahead. In this instance ?(?=:), matches anything
that is before :, so Reagan: would match, but if it was Reag:an, then only Reag would be
extracted. You could, of course write the expression as ((?<=>).*?(?=:)), meaning group
anything between >and: however this is opening your results to matching any other tag
that has >and :, so the President’s name attribute would also contain incorrect data. There
are many more ways in which this could be written to further restrict the results outcome,
such as adding \S, which means include only non-white space character data, so ((?<=title>).*?(?=:))\S. Other types of options include \w (a word letter) and [a-z] meaning show
lowercase. RapidMiner includes plenty of documented help regarding regular expressions,
so it is worth reading that reference material. Use the Edit Regular Expression: query
expression window, activated from the magnifying glass icon in Figure 1.18 to trial your
Regular Expression, as referenced in Figure 1.19. Within Grouping and Capturing there
are four main categories, two of which are previously covered. There is also a negative
lookbehind represented as (?! ....) and negative lookahead represented as (?<!....), both of
which state what to avoid or exclude. As a final note, aside from RapidMiner’s help files,
there is numerous reference material available online, which is worth researching, such
as http://www.rexegg.com/, http://www.regular-expressions.info/, and
http://www.regexlib.com/.
RapidMiner for Text Analytic Fundamentals
15
FIGURE 1.19: Build a token repository: Process I Step C – extracting regular expression.
• Cut Document: Finally the speech is extracted from a particular span class and is
extracted using regular regions as referenced in Figure 1.20, with a region delimiter
of <span class=”displaytext” >) and (\span >) (Figure 1.21), meaning that all text
within that span is captured. Label the attribute as text.
Note: When configuring the Cut Document operator, ensure that nodes seg and doc
(within its Segment Processing window) are joined as in Figure 1.22.
Step D (Figure 1.23): The final step in this process is to store the information using
the Repository Access:Store operator, in this case into a RapidMiner Repository called
Speeches, as referenced in Figure 1.23.
Note: Save the following process within the local repository Chapter as Step3-StoreMySQL
The next requirement within this second procedure element is to create a process that
writes to the MySQL database, inaugurualspeeches, that was referred to in Section 1.3 (as
16
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.20: Build a token repository: Process I Step C – regular region.
FIGURE 1.21: Build a token repository: Process I Step C – cutting speech content.
FIGURE 1.22: Build a token repository: Process I Step C – cut document nodes.
FIGURE 1.23: Build a token repository: Process I Step D – store repository.
RapidMiner for Text Analytic Fundamentals
17
FIGURE 1.24: Build a token repository: Process II.
FIGURE 1.25: Build a token repository: Process II Step A – Retrieve operator.
referenced in Figure 1.24). This enables the data to be accessed, manipulated, and queried
using SQL in the following sections of this chapter.
Step A (Figure 1.25): Use the Repository Access: Retrieve operator to access the
stored speeches repository Speeches.
Step B (Figure 1.26): The next operator you require is the Rename operator, as
you wish to change the names of two attributes, as in Figure 1.26 and Figure 1.27. These
renaming attribute changes improve the readability of your output.
Step C (Figure 1.28): Out of all the attributes in the repository, you wish to store
only three: InauguralDate, President, and Speech. In order to do this you use the Select
Attributes operator, and select your subset as in Figure 1.29.
Step D (Figure 1.30): Finally, to store this information into your MySQL database,
you need to connect to it using the Write Database operator. Within this, establish a
connection configuration to the inauguralspeeches database that you set up in Section 1.3.
Finally input a table name tblSpeeches and select the option overwrite first, append then.
Once you run your process, that table will be created and populated with all available
indexed speeches.
1.5.3
Summary
After this section you now have a fully populated table stored within a MySQL database
that contains three attributes: Speech, InauguralDate, and President. You should understand
and be aware of the differences between regular expressions, string matching, and regular
regions. You should also be familiar now with the differences between Information Extraction
and Cut Document operators.
18
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.26: Build a token repository: Process II Step B – Rename operator.
FIGURE 1.27: Build a token repository: Process II Step B – additional attributes.
FIGURE 1.28: Build a token repository: Process II Step C – Select Attributes operator.
RapidMiner for Text Analytic Fundamentals
19
FIGURE 1.29: Build a token repository: Process II Step C – subset attributes.
FIGURE 1.30: Build a token repository: Process II Step D – Write Database operator.
20
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.31: Analyzing the corpus: Process I.
The following section, moves away from this preparing and building corpus stage into
the areas of analyzing the raw data and subsequently visualizing its output.
1.6
1.6.1
Third Procedure: Analyzing the Corpus
Overview
Now that you have gathered all your information, you want to do something with it.
There are a plethora of ways in which you can search and, subsequently, analyze your
content. In the case of this example, you will write a standard query to retrieve President
Barack Obama’s 2009 inaugural speech and then use RapidMiner’s text mining processes
to perform various functions to it, including applying tokens, stopword dictionaries, and
n-grams. These results will then be stored in a spreadsheet and subsequently summarized
into the most frequently used words which are then displayed using a visualization method
called word clouds. Thus, with word clouds you wish to review the most frequent use of
words; to do this in RapidMiner, you need to tokenize the document explained within this
section.
The required RapidMiner processes are referenced in Figure 1.31.
1.6.2
Mining Your Repository — Frequency of Words
Note: Save the following process within the Local Repository Chapter as Step4TextAnalysis.
Step A (Figure 1.32): As can be referenced from Figure 1.32, you initially have to
query and retrieve relevant information from tblSpeeches using the Read Database operator.
RapidMiner for Text Analytic Fundamentals
21
FIGURE 1.32: Analyzing the corpus: Process I Step A – Read Database operator.
FIGURE 1.33: Analyzing the corpus: Process I Step A – SQL.
Click on the [Build SQL Query] button once you have established a connection labelled
InauguralPresidentSpeeches.
As can be referenced from Figure 1.33, you wish to extract the speech by President
Barack Obama in 2009, using the following SQL statement:
SELECT ‘Speech‘
FROM ‘tblspeeches‘
WHERE tblspeeches.InauguralDate like ’%2009%’ AND
President like ’%Obama%’
Note: The attribute Speech that is used within this query was generated through a renaming procedure included within the process Step3-StoreMySQL.
Step B (Figure 1.34): Once this query-specific statement is complete, the processed
speech has to be extracted and transformed into tokens; this is achieved using the Process
Documents from Data operator. As referenced in Figure 1.34, Term Occurrences is selected
to ensure that the number of recurrent tokens (words) is calculated. There are multiple
text token processes required within this operator’s Vector Creation window, which can be
22
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.34: Analyzing the corpus: Process I Step B – term occurrences.
FIGURE 1.35: Analyzing the corpus: Process I Step B – vector creation window.
viewed in Figure 1.35.
Note: Tokenization breaks the raw source document into individual words, these list formatted words are labelled as Tokens [3].
Since this data is HTML sourced, you have to remove the HTML tags such as <p >etc.,
so you include the Extract Content operator within the Process Documents from Data
vector window (Figure 1.36). From this, you have to break the cleaned source document
into individual words (also called tokens) using the Tokenize operator (Figure 1.37).
There are standard English stopwords such as the and is which you do not want cluttering your analysis; to automatically remove these words, you include a RapidMiner dictionary
operator called Filter Stopwords (English) (Figure 1.38).
Note: Stopwords are regular English words that add no benefit to the analysis such as the
and as.
RapidMiner for Text Analytic Fundamentals
23
FIGURE 1.36: Analyzing the corpus: Process I Step B – Extract Content operator.
FIGURE 1.37: Analyzing the corpus: Process I Step B – Tokenize operator.
Finally, the last operator, Filter Stopwords (Dictionary), is a customized list of words
that you create in which to remove tokens that clutter the results and have no benefit in
their inclusion. This custom dictionary is created via Windows Notepad, and is saved as a
.txt file extension. In this case, the custom stopword dictionary is saved as speechstop.txt
(a copy of which is included with the processes for this chapter) and contains the words
represented within Table 1.4 (which are subsequently excluded from the analysis output).
Simply list the words one line after another that you wish to exclude. Figure 1.39 contains
the required configuration information you need to properly establish a custom dictionary.
Step C (Figure 1.40): To output these results use a Write Excel operator to write
to a spreadsheet (Figure 1.40) called outputanalysis.xlsx. The data is represented with the
words in Row 1, and the frequency of their occurrences in Row 2. For analysis with word
clouds, these need to be copied and transposed into a new worksheet so that the words are
in Column A (labelled Token) and the frequency values in Column B (labelled Amount),
as referenced in Figure 1.41.
This outputted report will be transformed into a word cloud within Section 1.7
24
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.38: Analyzing the corpus: Process I Step B – Filter Stopwords (English) operator.
TABLE 1.4: Speechstop.txt content.
Term
President
America
God
Bless
United
States
Nation
American
Americans
FIGURE 1.39: Analyzing the corpus: Process I Step B – custom dictionary.
FIGURE 1.40: Analyzing the corpus: Process I Step C – Write Excel operator.
RapidMiner for Text Analytic Fundamentals
25
FIGURE 1.41: Analyzing the corpus: Process I Step C – transposed report.
1.6.3
Mining Your Repository — Frequency of N-Grams
Note: Save the following process within the Local Repository Chapter as Step4TextAnalysis-NGRAM.
Listing words by their recurrence frequency, as in Section 1.6.2, provides basic summarization of topics by calculating the most repeated individual words (tokens) in a speech.
However, this type of analysis can lack overall clarity of context. A method to try and gain
this level of nuance is obtained by analyzing words that traditionally could be considered
to be combined together, called n-grams [4]. Examples of these would be United States and
Founding Fathers. In fact, combined tokens (multi words) can enhance the output’s interpretation [3]. Research has also indicated that people tend to prefer viewing a combination
of words instead of single terms [5].
Note: n-gram (terms) are used for combining words.
N-grams can be combinations of two, three, four, or more words. For this section, a two
consecutive-word combination is used.
Note: You will see in literature that 2 n-gram term processes are also referred to as bigrams
[5].
As in Section 1.6.2, Steps A and B are the same procedures. Step B contains two additional operators within the Vector Creation window, as referenced in Figure 1.42.
These two operators are specifically related to the bigrams process. These are Generate
N-Grams (Terms) (Figure 1.43) which has a max length value set at 2 indicating that you are
seeking a bigrams (or two combined tokens) and Filter Tokens (By Content) (Figure 1.44)
which has its condition configured as contains and string set at . This operator thus ensures
that all non-n-grams are excluded from the final output, as you only wish for words that
are joined by .
As in Step C from Section 1.6.2, write the results to a spreadsheet called outputanalysisng.xlsx, and transpose those results to another worksheet; the exact same process as in
Section 1.6.2 (Figure 1.45). These results will be transformed into a visualization report
using word clouds in Section 1.7.
26
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.42: Analyzing the corpus: Process II Step B.
FIGURE 1.43: Analyzing the corpus: Process II Step B – Generate n-Grams (Terms)
operator.
FIGURE 1.44: Analyzing the corpus: Process II Step B – Filter Tokens (by Content)
operator.
FIGURE 1.45: Analyzing the corpus: Process II Step C.
RapidMiner for Text Analytic Fundamentals
1.6.4
27
Summary
After this section, you should be entirely familiar with the basic concepts of retrieving
information using SQL, how to remove HTML code, tokenize the speech documents, and
output them to a spreadsheet. You should also be familiar with the difference in output
of using n-grams and not using them. You should also experiment with different types
of n-grams and also alternative customize stopword dictionaries to understand how that
changes your results. In the next section, the spreadsheet reports that you generated from
your analysis will be converted into word cloud visualization representations.
You could also vary your results by applying different operators within the Process
Documents from Data operator including filtering by length (Filter Tokens (by Length)
operator) checking parts of speech such as nouns or verbs (Filter Tokens (by POS Tags) operator), or stemming words (Stem (Porter) operator). However, the key point to take from
this is not to show how sophisticated your text mining model is or how many operators you
include; it is solely what adds benefit to your overall analysis.
Note: You also have to be extremely careful that you do not lose patterns. For example,
filtering tokens out by length is useful to avoid unnecessary polluting of results; however,
you may also lose key terms. For example, you may configure when analyzing a financial
document to exclude three-letter words; if that is the case then you may lose words
such as IMF (International Monetary Fund). Thus, the key is to read your document(s),
understand the content, note the key points, and only then build your model. The same
applies to stemming, where you truncate a word to its core, such as political to polit and
politic to polit. An over-enthusiastic stemmer could potentially lose valuable insights [5].
1.7
Fourth Procedure: Visualization
1.7.1
Overview
In the previous sections, all the analysis was based within the RapidMiner software application environment, and the reports that were produced were text-based. In this section,
you want to visualise those spreadsheet reports. You could use many online environments
to achieve this. For this section, you will produce your word clouds within the Wordle
www.wordle.net online tool.
1.7.2
Generating word clouds
The first word cloud that you will produce sources its dataset from the transposed report
generated within Section 1.6.2.
Note: word clouds are a quick reference visualization representation of a document (such
as speech), displaying the most frequently used words within its content; the actual font
size of the token within the cloud indicates its frequency [5].
1.7.2.1
Visualize Your Data
1. Open your transposed worksheet from within outputanalysis.xlsx and input the following formula into Cell C2:
28
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.46: Visualization: Layout of transposed worksheet.
=REPT((A1 & " " ),B1)
The Token should be stored within Cell A2, and the Number of Times that it appears
throughout the document is within Cell B2 (Figure 1.46). The formula in Cell C2
recreates the number of times that this token is mentioned in the original source
document by repeating the word (Wordle requires this in order to generate a word
cloud visualization of its frequency). Drag this formula down Column C until you
reach the last token row.
2. Once complete, open www.wordle.net (Figure 1.47), select Create, and within the
subsequently displayed grid box (Figure 1.48) area, copy and paste the string data
stored within the C column. If limited by processing, it would be best to sort the rows
by the most frequent to the least, and only select the top one hundred most frequent
tokens for your word cloud.
3. Once this task is complete, press the Go button, to generate the word cloud represented within Figure 1.49.
4. There are many editing and filtering options for layout (Figure 1.50), colouring (Figure 1.51), limiting the number of tokens on display (Figure 1.52) and changing font
types (Figure 1.53). You are free to experiment here, but remember that the benefit
and aim of your analysis is not to display the best, most flashiest, graphic visualization. Your goal is always to create a clear and insightful report, nothing more. For
example, Figure 1.54 represents the dataset with a limit of 20 maximum-value tokens (avoids cluttering), a horizontal layout (easy to read), and wild variance in the
colouring palette (allows tokens to be distinct from one another).
5. The initial word cloud contains words that do not add much to the analysis. You
can right click on a word to remove it (Figure 1.55). For this report, remove people,
common, day, world, calls, things, generation, generations, time, come, and know and
review the difference (Figure 1.56). It is clear now that economy, peace, and crisis are
tokens that have prominence in the speech.
6. At the end of the word cloud (Figure 1.57), you will see that you can change the layout
of the word cloud (Randomize), open it as a separate window (Open In Window ), print
your image (Print), or save it as a file on your desktop (Save as PNG).
Note: REPT((A1 & ” ” ),B1) is an Excel formula that repeats a string fragment (your
tokens) by a defined number of times (the number of occurrences RapidMiner registered
its frequency). You need to concatenate the original string token with a space character
in order for Wordle to recognise it as multiple words as opposed to a single word term,
so hope hope instead of hopehope.
RapidMiner for Text Analytic Fundamentals
29
FIGURE 1.47: Visualization: Wordle menu.
Note: With Wordle word cloud generation you need to have Java-enabled. Please note
that Google Chrome browsers stopped supporting NPAPI which is required to run Java
applets.
You can apply the same steps to the report (outputanalysisng.xlsx ) that you generated
from Section 1.6.3, to understand the differences between counting individual tokens and
calculating multi words that are combined in their context (Figure 1.58). Instead of the
formula used for the single tokens, use the following instead, as Wordle does not recognise
the underscore marker and thus would separate the words as if they were independent tokens
instead of combined tokens:
=SUBSTITUTE(REPT((A2 & " " ),B2), "_", ":")
Figure 1.59 is limited to 40 tokens, and is further filtered in Figure 1.60
with removal of I related tokens such as I say. From both of the visual representations you can understand that there is a focus on health-related terminology.
Note: SUBSTITUTE(REPT((A2 & ” ” ),B2), ” ”, ”:”) is an Excel formula that repeats
a string fragment (your tokens) by a defined number of times (the number of times RapidMiner registered its frequency) and then substitutes (replaces) the underscore with :.
This is required as Wordle does not recognise and thus treats words (tokens) such as
Founding Fathers as Founding and Fathers. With :, Wordle recognises the bigram as
Founding:Fathers.
30
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.48: Visualization: Copy and paste data into create section.
FIGURE 1.49: Visualization: Speech represented as a word cloud.
RapidMiner for Text Analytic Fundamentals
FIGURE 1.50: Visualization: Word cloud layout.
FIGURE 1.51: Visualization: Word cloud colour.
31
32
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.52: Visualization: Word cloud token limit.
FIGURE 1.53: Visualization: Word cloud font types.
RapidMiner for Text Analytic Fundamentals
FIGURE 1.54: Visualization: Word cloud remove tokens (filtered to 20 tokens).
FIGURE 1.55: Visualization: Word cloud filter option (remove token).
33
34
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 1.56: Visualization: Word cloud filtered (specific tokens removed).
FIGURE 1.57: Visualization: Word cloud options (print, save, new window, randomize).
FIGURE 1.58: Visualization: Layout of transposed bigram worksheet.
RapidMiner for Text Analytic Fundamentals
35
FIGURE 1.59: Visualization: Word cloud bigrams.
FIGURE 1.60: Visualization: Word cloud bigrams filtered (removal of tokens).
1.7.3
Summary
Within this section you created visualization reports of the worksheets produced within
Section 1.6. You will be aware of how to upload a data set, generate a word cloud, exclude
tokens and change the layout of your imagery.
1.8
Conclusion
This chapter has covered the basic tools that you require to start gathering text sources
and exploring their content using software such as RapidMiner Version 5.3 and online visualization environments such as Wordle. You should experiment with checking for parts
of speech and different visualizations. You should also do further reading around the whole
topic area; http://scholar.google.com/ is a notable resource for text mining research papers. Mastering the core techniques outlined in this chapter should enable you to
build upon and enhance your expertise, knowledge and skills by adding to this base using
different corpus, experimenting with other text mining methods and RapidMiner operators.
Bibliography
[1] G Peters. The American Presidency Project [online] http://www.presidency.ucsb.edu/
inaugurals.php. Last Accessed: October 2015.
[2] F Jungermann. Information extraction with rapidminer. In Proceedings of the GSCL
Symposium Sprachtechnologie und eHumanities, pages 50–61. Citeseer, 2009.
[3] S.M. Weiss, N. Indurkhya, and T. Zhang. Fundamentals of predictive text mining, volume 41. Springer, 2010.
36
Text Mining and Visualization: Case Studies Using Open-Source Tools
[4] RapidMiner 5.3 Software. Text sourced within the Help Module: Generate n-Grams
(Terms) operator [online] http://rapid-i.com/wiki/index.php?title=text:generate ngrams (terms)&oldid=2554. Last Accessed: October 2015.
[5] G de Hollander and M Marx. Summarization of meetings using word clouds. In Computer Science and Software Engineering (CSSE), 2011 CSI International Symposium
on, pages 54–61. IEEE, 2011.
Chapter 2
Empirical Zipf-Mandelbrot Variation for
Sequential Windows within Documents
Andrew Chisholm
Information Gain Ltd., UK
2.1
2.2
2.3
2.4
2.5
2.6
2.7
2.8
2.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Structure of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Rank–Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.1
Heaps’ Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.2
Zipf’s Law . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.3.3
Zipf-Mandelbrot . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
RapidMiner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1
Creating Rank–Frequency Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.5.1.1
Read Document and Create Sequential Windows . . . . . . .
2.5.1.2
Create Rank–Frequency Distribution for Whole
Document and Most Common Word List . . . . . . . . . . . . . . . .
2.5.1.3
Calculate Rank–Frequency Distributions for Most
Common Words within Sequential Windows . . . . . . . . . . . . .
2.5.1.4
Combine Whole Document with Sequential Windows . .
2.5.2
Fitting Zipf-Mandelbrot to a Distribution: Iterate . . . . . . . . . . . . . . . . . . . .
2.5.3
Fitting Zipf-Mandelbrot to a Distribution: Fitting . . . . . . . . . . . . . . . . . . .
Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.1
Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.2
Starting Values for Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.6.3
Variation of Zipf-Mandelbrot Parameters by Distance through
Documents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
38
38
39
39
40
41
42
43
43
46
46
47
47
48
52
52
52
53
58
58
59
Introduction
Word rank–frequency distributions are often used as input to various authorship attribution investigations. This chapter uses advanced RapidMiner processes to fit the ZipfMandelbrot distribution to sequential fixed-size windows within a single document to create
a summary of observed fluctuations for an author. These are compared with random samples of the same size from the same document. Different works by the same author as well
as different authors are assessed to determine if there is a consistent pattern for individ-
37
38
Text Mining and Visualization: Case Studies Using Open-Source Tools
ual authors. Some initial evidence of a consistent variation for different works by the same
author is observed.
The study of word rank–frequency distributions (hereafter referred to as rank–frequency
distributions) within documents (see [1] for an extensive and detailed overview) has found
applications in understanding the differences between texts to determine authorship (see [2]
and [3]). The assumption behind this is that authors have unique styles that can be revealed
by the analysis of these distributions. The distributions are observed to follow power law
like behaviour and various empirical distributions, such as Zipf’s law, can be fitted to the
data. The act of fitting to a distribution generates numerical measures that summarise the
fit and these are assumed to relate to stylistic features.
It is observed that as more and more text is written, new unique words continue to
be created. The rate of creation tends to zero as the text increases, but it never reaches
it. A sample of an author’s entire work, such as a book, will contain words with a certain
frequency. The same words from a larger sample will appear less frequently by virtue of the
new unique words in the larger sample. This is counter to the usual rule of statistics that
a larger sample gives more accuracy, and it means that comparisons between samples of
different sizes have to be treated with care (see [1] for extensive detail).
It is also observed that by taking samples of a specific size within a fixed corpus there is
considerable variation in the sampled word frequencies [1]. This variation is more than can
be explained by random chance alone and points to a style variation imposed by the author
as part of the creative process. It is an interesting question to ask whether the variation
itself is different for different authors but is preserved between different books written by
the same author.
As a first step to answering this question, this chapter illustrates this variation for
sequential fixed-size windows within a document and fits the Zipf-Mandelbrot distribution
to each as a way of summarizing the differences. Comparison is also made to random samples
of the same size to illustrate that the sequential windows vary from a random distribution.
By observing how the differences vary for different authors, some evidence is seen that
authors have a variation of style that is different from other authors but which is preserved
for different works by the same author.
The chapter makes extensive use of RapidMiner version 6.1 with sufficient detail to allow
the processes to be reused by others.
2.2
Structure of This Chapter
The rest of this chapter is structured as follows. Related works and various aspects of
rank–frequency distributions are considered in Section 2.3 and Section 2.4. RapidMiner processes are described in Section 2.5. These processes implement various aspects of the overall
process to generate rank–frequency distributions and fit Zipf-Mandelbrot distributions to
the data. Section 2.6 describes the results, Section 2.7 provides discussion, and Section 2.8
summarises.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
2.3
39
Rank–Frequency Distributions
A rank–frequency distribution is constructed by identifying each word in a document
and counting the number of times it occurs. The resulting list of unique words is sorted
with the most common word being given the rank of one, the second the rank of two and so
on. By normalizing the frequencies so they sum to 1, each frequency becomes a probability
that the word appears in the document.
There are various empirical laws and curves that can be fitted to typical rank–frequency
distributions and to the total number of words and unique words. Among the most important
are Heaps’ Law, Zipf’s law and the Zipf-Mandelbrot distribution. These are all described in
the following sections.
It is important to state the definition of what defines a word to ensure repeatability so
that the same words are used for fairer comparisons between different documents. Throughout this chapter, words are defined as sequences of one or more letters separated by non
letters. All words are converted to lowercase, and no filtering to remove any tokens is performed.
2.3.1
Heaps’ Law
Heaps’ law [4] is an empirical law that relates the number of unique words in a document
to the count of all the words with the following formula:
M = KN α
(2.1)
where M is the number of unique words, K is a constant generally between 10 and 100, N
is the total number of words in the document, and α is a constant generally between 0.4
and 0.6.
The formula for this empirical law matches the observation that, as a document increases
in size, the number of unique words increases. This means the ever increasing number of
unique words must reduce the probability of preexisting words which is what is observed in
real texts.
2.3.2
Zipf ’s Law
The empirical law described in 1932 by G.K. Zipf ([5], [6]) applies to rank–frequency
distributions and states that the rank of a word multiplied by its frequency is a constant.
For example, in a document containing 1,000 words, of which 50 are the most common word
(usually “the” in English), the law predicts that the next most common word should appear
25 times and the third most common 17 times. The rule is that the frequency multiplied
by the rank is a constant, so in this case 50 × 1 = 25 × 2 = 17 × 3. Indeed, Heaps’ law and
Zipf’s law express the same conclusion, and it is possible to relate them [7], [8]. The law is
empirical and in practice, all texts exhibit divergences from it.
Mathematically, the law can be expressed as follows:
fr r = C
where the frequency of the word of rank r is fr , and C is a constant.
(2.2)
40
Text Mining and Visualization: Case Studies Using Open-Source Tools
This can be recast into the following formula (for details see [9]):
pr = [rH(N, α)]−1
(2.3)
where pr is the probability of the word of rank r, r is the rank of the word, H(N, α) is
the harmonic number, N is the total number of unique words, and α is the order of the
exponent. For Zipf’s law this exponent is set to 1. The harmonic number of order N , α is
shown as follows:
N
X
H(N, α) =
r−α
(2.4)
i=1
In the limit for large N the harmonic number in the Zipf case simplifies to
H(N ) ≈ ln(N ) + γ
(2.5)
where γ is the Euler-Mascheroni constant with approximate value 0.57721. The value of
expγ ≈ 1.78, allows Equation 2.3 to become Equation 2.6 for large N :
pr ≈ [rln(1.78N )]−1
(2.6)
This shows an important point that the probability of a given ranked word depends
on the total number of words in a document and the larger the document the lower the
probability of finding that word. This has the consequence that comparing documents of
different sizes must lead to different values for the probabilities of words.
For example, if a 5,000-word sample is taken from a larger book, Heaps’ law with parameter values of α = 0.5 and K = 20 in Equation 2.1 would give an estimate for the number
of unique words as 1,414. Using this number with Zipf’s law from Equation 2.6 would give
an estimate that the most common word would appear with a probability of 0.128. Increasing the document to be 100,000 words reduces the probability for the most common word
to 0.107. Taking this further, if the author has written 100 books each 100,000 words in
length the total corpus for that author would consist of 10,000,000 words and the estimated
probability for the most common word would reduce to 0.0860.
Aside from the variation that arises from different sample sizes, the law itself never
fits precisely to real data. These variations can be exploited to reveal differences between
authors, and there exist many authorship attribution techniques that partially use rank or
frequency data (see [10], [11], [3], [12], and [13]).
The law is extremely simple, and various attempts have been made to create a better
fit.
2.3.3
Zipf-Mandelbrot
Mandelbrot proposed a modification to Zipf’s law [14] to enable it to fit more closely
rank–frequency distributions, particularly at both ends of the distribution. The mathematical equation is given in Equation 2.7.
pr = C(r + A)−B
(2.7)
where pr is the probability of a word of rank r, and A, B, C are constants that must be fitted
to the data. The equation is more complex than Zipf’s law, and this affords the possibility
that fitting it to different distributions will yield different values for the constants that could
reveal something about the style being adopted by the author.
Additional discussion about this will be given towards the end of Section 2.4 after
consideration of sampling from a fixed corpus in general.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
2.4
41
Sampling
As stated in the introduction, a sliding window is to be moved through a document.
When sampling randomly from a population, it is possible to place statistical limits on the
characteristics of the sample and in this way, the assumption of randomness can be tested.
As a worked example, in a document containing 100,000 words with the most common
word appearing with a probability of 0.1 then a truly random sample of 5,000 words should
still contain the same proportion of this word. With an observed sample probability and
sample size it is possible to assess how well the assumption holds up against the population
probability if this is already known.
Equation 2.8 is the approximate standard error of the sample proportion
p
standard error ≈ (P (1 − P )/n)
(2.8)
where P is the known population probability, and n is the sample size. For the worked
example, if 5,000 words are sampled and P is known to be 0.1, Equation 2.8 gives 0.00424.
If the observed sample probability is 0.095 this means the observed probability is (0.095 −
0.1)/0.00424 = −1.18 standard deviations from the expected value. Statistically, given the
assumption of randomness, this z-score means the observation is likely to be observed. If
the observed probability was 0.05 the observation would have a z-score of −11.8 and would
therefore be extremely unlikely to be observed.
An assessment of the sample probability of words in a sliding window through a document gives the possibility to assess whether the words written by an author are random or
have been arranged to meet some unconcious style. As an illustration, Table 2.1 contains
data from the novel Moby Dick by Herman Melville [15].
TABLE 2.1: Variation of z-score
windows for the novel Moby Dick.
Window
0-4,999
5,000-9,999
10,000-14,999
15,000-19,999
20,000-24,999
for the most common words in sequential 5,000-word
the
of
−0.707 3.028
−1.617 −0.192
−5.489 −2.833
−0.422 1.129
−2.642 −2.338
and
−0.102
0.651
2.157
−0.270
2.408
a
3.159
7.431
6.072
3.936
−0.336
This table was produced by moving a 5,000-word window through the document and calculating the observed probability for words. The known probabilities for the whole document
were also calculated and this allowed the standard error to be calculated using Equation
2.8. Subtracting the population mean and dividing by the sample error gives a z-score which
is shown in the table. The table shows that there are some large z-scores which it would
be extremely unlikely to observe from a purely random sample. Plotting all the z-scores
as a histogram shows that these sequential windows exhibit significant variations for all
words. For example Figure 2.1 shows this z-score variation for the word “the” within the
novel Moby Dick [15]. This novel contains more than 210,000 words and the figure shows
the observed probability for the 42 sequential windows each of 5,000 words. The results
show that there is considerable variation that cannot be explained using the assumption
that the words are randomly arranged within the sequential windows. Other words are also
observed to have similar large variations, and it would be possible to create a summary
42
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.1: Observed variation for the word “the” for consecutive 5,000-word windows
within the novel Moby Dick.
statistic that captured these. This summary could then be used as a feature vector for an
author to determine if the author exhibited a variation that was unique.
Returning to the Zipf-Mandelbrot distribution from Section 2.3.3, one possibility is to
use this as a way to fit the distribution to the observed probabilities and use the values
of the parameters A, B, C of Equation 2.7 as a proxy for the deviation from randomness.
This makes sense because the rank–frequency distribution is known from the entire book
and the sample provides different probabilities. By keeping the window sizes fixed and by
using probabilities from the entire book, variations relating to different sample sizes can be
minimised.
One final step is to restrict to the most common words. These are typically function
words that by themselves do not convey meaning and which all authors use. This affords the
possibility that the approach may allow different authors to be compared. For this chapter,
the most frequent 100 words are used. To be clear, this means that when a document is
analysed, the most frequent 100 words are selected. When sequential or random windows
are chosen, the presence of the 100 words from the main document is used to determine the
rank–frequency distribution. This means that when comparing books there is a possibility
that different 100 words will be chosen.
2.5
RapidMiner
This section shows the use of RapidMiner [16] to create processes for the key steps to
calculate rank–frequency distributions for documents and sequential windows within the
documents as well as fitting the Zipf-Mandelbrot distribution to these windows using an
evolutionary or genetic algorithm approach. Three processes are provided and are described
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
43
in Table 2.2; the processes are advanced and the key features are described in the referred
sections. All the processes are available from the website for this book.
TABLE 2.2: RapidMiner processes and sections where they are described.
Process name
Details in Section
generateWindowedWordFrequencies
2.5.1
iterateAndExecuteZipfMandelbrotFit 2.5.2
fitZipfMandelbrot
2.5.3
2.5.1
Creating Rank–Frequency Distributions
The high-level process is shown in Figure 2.2. There are four main process sections in
this, labelled A to D, and more detail for each of these is given in Figures 2.3, 2.4, 2.5, and
2.6, respectively. Table 2.3 summarises these and refers to the sections where the detail of
each is given.
FIGURE 2.2: RapidMiner process to calculate word frequencies.
2.5.1.1
Read Document and Create Sequential Windows
This process section (A) reads the document to be processed. This is passed from the
process context to the operator Read Document. The process context reads documents from
a repository. To place documents in a repository it is necessary to use the operator Open
44
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.3: Process Section A within Figure 2.2.
FIGURE 2.4: Process Section B within Figure 2.2.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
FIGURE 2.5: Process Section C within Figure 2.2.
FIGURE 2.6: Process Section D within Figure 2.2.
45
46
Text Mining and Visualization: Case Studies Using Open-Source Tools
TABLE 2.3: Process sections for RapidMiner process to calculate
tions.
Process
Description
Section
A
Read document and create sequential
windows
B
Calculate rank–frequency probability
distribution for entire document and
create word list for most common words
C
Calculate rank–frequency probability
distributions for windows within document
D
Combine whole document and window rank–frequency probability distributions
rank–frequency distribuSection
Section 2.5.1.1
Section 2.5.1.2
Section 2.5.1.3
Section 2.5.1.4
File followed by the Store operator. This is not shown in the process as it is straightforward
to implement. The process section multiplies the document into two copies. One copy is
sent to process section B and the other is sent to a Sub-process operator where it is split
into sequential windows. The size and offset between windows is defined by the macros
windowStepSize and windowLength defined in the process context (in this chapter these are
both 5,000). The operator Window Document uses these macros to create the windows. It
is important to get the input to this operator correct. To do this requires tokenizing the
input document. As mentioned in Section 2.3 the document is tokenized using the Tokenize
operator with parameter set to non letters and all tokens are converted to lower case. There
is one work around that is needed to cater for the addition of the window meta data attribute
added by the Window Document operator. This involves replacing any token with the value
window with another value that does not occur in the text. Later on, within process section
C, this will be reversed. It should be noted that the latest versions of RapidMiner (version
6.1 as at the time of writing: early 2015) cope better with this, but the workaround presented
here will work with older versions of RapidMiner.
2.5.1.2
Create Rank–Frequency Distribution for Whole Document and Most
Common Word List
This process section (B) finds all the words in the original document and makes a rank–
frequency distribution from it. This is done by using the operator Process Documents which
contains the same tokenizing rules as described in Section 2.5.1.1. The Process Documents
operator creates a vector based on term occurrences; the count of the number of times each
word appears in the document. The word list output from this operator is converted to an
example set and various manipulations are performed to sort the result so the most common
word is given the rank of 1. By using the Normalize operator set to proportion transformation, the counts are transformed into probabilities. The process section also outputs a word
list of selected words. Section 2.4 describes using the most common words and this is controlled by the macros firstExample and lastExample defined in the process context. For this
chapter these macros are set to 1 and 100, respectively. To create a word list, the process
section takes the example set corresponding to these words and uses Process Documents
from Data to create the list.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
2.5.1.3
47
Calculate Rank–Frequency Distributions for Most Common Words
within Sequential Windows
Process section C takes the windowed input document and word list of the words of
interest and creates an example set containing an example for each word and an observed
probability for the occurrence of the word in each of the windows within the entire document. The example set that is created contains as many examples as there are words of
interest with one attribute called word describing the word and attributes called window n
where n is a number denoting the starting point for the number of characters through the
document where the window starts. The process section uses some advanced techniques.
Firstly, the workaround for the window meta-data attribute introduced by the Window
Document operator is reversed. This is done by renaming the meta data attribute called
window to be 0 window (this is important later on when transposing) and if there is an
attribute corresponding to the occurrence of the word window in the original text, this is
also renamed. This is done inside a Branch operator because it is entirely possible that a
text and the windows within it do not contain the word window. Secondly, the Transpose
operator is used. This swaps rows and columns with the important point that an id attribute
is created containing the original names of the attributes. By sorting this and knowing that
the window attribute is called 0 window it is possible by sorting to ensure that this is the
first example of the transposed example set. This allows the operator Rename By Example
Values to be used so the name of the attributes of the transposed example set are meaningful. Certain other renaming and manipulation operations are performed to get to the
desired end result.
2.5.1.4
Combine Whole Document with Sequential Windows
The final part of this process (D) is to combine the rank–frequency distribution of the
words of interest of the whole document with the sampled probabilities from each window.
This is done using the Join operator. This process section also uses the Remove Unused
Values operator to remove unused nominal values in the example set and in addition uses
the Materialize Data operator. This last operator should not be needed, but there are times
when it is invaluable as it seems to clear transient memory issues brought about by complex
preprocessing.
2.5.2
Fitting Zipf-Mandelbrot to a Distribution: Iterate
Two RapidMiner processes are provided to calculate the optimum fit of the ZipfMandelbrot distribution to the observed rank–frequency measurements. The first contains
various looping operations to iterate over all the sequential windows and call the second
using the Execute Process operator. The first process is described in this section, and the
second is described in Section 2.5.3.
The process to iterate over all the windows requires an input example set containing
an attribute named rank, which must be a nonregular attribute and at least one regular
attribute named window n where n is a number. The location of the example set is defined in
the process context. Figure 2.7 shows the top-level process. This takes the input example set,
selects only those with the required names and then loops through all the attributes called
window n using the Loop Attributes operator. The inner operators of the loop perform more
processing and execute the RapidMiner process to fit the Zipf-Mandelbrot distribution. The
output of the loop is a collection of example sets. Each example set contains the window
name and the optimum values of the parameters A, B, C for it. The collection is combined
into a single example set using the Append operator, and the Guess Types operator is used
to convert numbers to numeric types to make graphical plotting possible. Returning to
48
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.7: RapidMiner process to execute process for all attributes to fit Zipf-Mandelbrot
distribution.
the inner operators of the Loop Attribute operator, Figure 2.8 shows these. These work
FIGURE 2.8: Detail of RapidMiner process to execute Zipf-Mandelbrot distribution fit.
as follows. The Loop Attributes operator provides a macro corresponding to the current
attribute, which is used to select only this attribute (note the important point that the
original example set must be passed to the first output of the Loop Attributes operator to
make it work properly). The rank attribute will also be selected by virtue of it not being a
regular attribute. The selected attribute is then renamed to probability since the process to
fit to the Zipf-Mandelbrot distribution requires this and then Execute Process is called. The
output from this is the optimum values fitted to the observed data, and an extra attribute
is created corresponding to the current window.
2.5.3
Fitting Zipf-Mandelbrot to a Distribution: Fitting
The top-level process that fits a rank–frequency distribution is shown in Figure 2.9.
Referring to this figure, the operators labelled 1, 2, and 3 ensure there is an example set
passed to the rest of the process. The operator labelled 1 is a Branch operator that detects if
no example set has been passed and will read from the repository in this case. This allows the
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
49
whole process to be tested and used in a stand alone fashion. This is an important technique
when developing and debugging processes. The operator labelled 4 contains the heart of the
FIGURE 2.9: RapidMiner process to fit Zipf-Mandelbrot distribution.
process and is a Loop Parameters (Evolutionary) operator. This will be considered in more
detail shortly. The output from this operator is a performance vector corresponding to the
optimum parameters to fit the distribution and log entries containing the parameters and
total error calculated as the looping operation proceeds. The log entries are converted to an
example set using the Log to Data operator (the operator labelled 6 in Figure 2.9) and an id
attribute is added to ensure the order in which the entries were produced can be retained.
The parameter set output is used to set the parameter values of certain operators (operator
5), and various macros are then set based on these so they can be used to recreate how the
parameters recreate the optimum fit (operator 9).
Returning to the Optimize Parameters operator, Figure 2.10 shows the configuration
of the parameter settings for this operator. This shows that three parameters are defined
and these correspond to the Zipf-Mandelbrot parameters from Equation 2.7. The Optimize
Parameters (Evolutionary) operator uses a genetic algorithm approach to find the smallest
error for a performance vector as these parameters are used in a function to fit to the known
result. The operator repeats the fitting process with different parameter values and the
genetic algorithm approach ensures an efficient convergence to a good solution. Figure 2.11
shows the inner operators within this operator. The operators numbered 1, 2, and 3 in
this operator are there to receive the parameters defined by the optimizing operator itself.
Closer inspection of these reveals that these are in fact Generate Data operators. The only
reason these are used is to work around a restriction exhibited by the Optimize Parameters
50
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.10: Configuration for Optimize Parameters (Evolutionary) operator.
FIGURE 2.11: Details for Optimize Parameters (Evolutionary) operator.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
51
(Evolutionary) operator that prevents it from setting numerical parameters directly into
macros. By using an intermediate operator that is able to take a numeric parameter, numeric
values can be set up and then read by a Generate Macros operator (labelled 4 in Figure 2.11).
The Generate Macros operator uses the built in function param() to do this and sets the
values of macros A, B, and C. Figure 2.12 shows the detail for this operator.
FIGURE 2.12: Details for macro-generation workaround to pass numerical parameters to
Optimize Parameters operator.
Once the macros A, B, and C have been populated, they can be used with the formula in
Equation 2.7 and the example set containing rank and probability values. This is operator
5 from Figure 2.11, and the parameters to this are shown in Figure 2.13. This shows the
calculation of an estimated probability as well as a difference between the estimate and
the known probability from the sample. The error is the absolute value of the difference,
but other error calculations could be used here such as the square. Having calculated the
FIGURE 2.13: Calculation of Zipf-Mandelbrot probability and error from known probability.
error for each example, these are aggregrated using the Aggregate operator (labelled 6),
which simply sums all the individual errors. The next operator allows more calculations to
be performed on the aggregated result if desired. For the results in this chapter, it simply
52
Text Mining and Visualization: Case Studies Using Open-Source Tools
raises the aggregate error to the first power (i.e., no change). This could be changed to find
the square root of the aggregate if desired. The operator is left in as a placeholder only.
The final part of the operator chain calculates a custom performance (operator 9). It
does by using the data value of the single aggregated error calculated in the preceding
operators. By this method, the sum of the errors for each example is reduced to a single
error value and this is the performance for the values of the macros A, B, and C when
calculating the Zipf-Mandelbrot fit. The last operator (labelled 10) writes log entries for the
macros and errors. The Optimize Parameters (Evolutionary) operator has a configuration
setting to minimize or maximize errors; in this case, it is set to minimize.
2.6
Results
2.6.1
Data
Various novels are used in this chapter. These are freely available public domain texts
available from the Project Gutenberg site [15]. Table 2.4 itemizes the ones used in this
chapter.
TABLE 2.4: Details of texts used in this chapter.
Author
Title
Number of words
Jane Austen
Pride and Prejudice
122,815
Jane Austen
Sense and Sensibility
120,718
Jane Austen
Mansfield Park
161,429
Arthur Conan Doyle The Return of Sherlock Holmes
114,407
Arthur Conan Doyle The Adventures of Sherlock Holmes
105,900
Herman Melville
Moby Dick
218,378
Herman Melville
The Piazza Tales
81,089
In all cases, light editing was performed to remove the Project Gutenberg specific preamble and license agreements at the start and end of each document.
2.6.2
Starting Values for Parameters
Consideration of Equation 2.7 leads to some limits on the possible values for the parameters A, B, C. It is important to apply limits for practical reasons since it reduces the
amount of time taken to find optimum solutions. More importantly, however, by limiting
the search to meaningful regions of the search space, solutions can be found that make
sense. With this in mind, the following ranges were used for the parameters.
TABLE 2.5: Details of parameter ranges used in this chapter.
Parameter Range
A
0 to 1
B
−2 to 0
C
0 to 0.2
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
2.6.3
53
Variation of Zipf-Mandelbrot Parameters by Distance through
Documents
As an illustration of what the processes do, Figure 2.14 shows a comparison between
the observed rank–frequency data for an entire document and for a 5,000-word window
within the document. In addition the optimum Zipf-Mandelbrot fit to the observed values
is shown on the same graph. The optimum parameter values for A, B, C are included in the
caption. The processes were run against the books described in Section 2.6.1 to determine
the optimum Zipf-Mandelbrot parameters for all the windows within the documents. In
addition to the 5,000-word sequential windows, 200 random selections of 5,000 words were
also chosen from each of the books and the Zipf-Mandelbrot distribution was fitted to these
(The creation of random samples is straightforward to implement within RapidMiner but
the process is not described in this chapter for reasons of space.)
FIGURE 2.14 (See color insert.): Log of probability and estimated probability as a function
of log rank for the 100 most common words within all of Pride and Prejudice and for the
sequential window from 40,000 to 45,000 words. The values of the parameters from Equation
2.7 are for the whole document: A = 0.958, B = −0.846, C = 0.094 and for the window:
A=0.888, B=-0.761, C=0.076. This graph was produced using R with the basic plot, points
and lines functions used to overlay additional information on a base plot.
Starting with the random samples, examination of the results showed that even for these
there was variation in the fitted parameters. However, it was observed that the variations
seemed to form clusters when 2- or 3-dimensional plots were examined. This showed up most
clearly with the two parameters A and C. There was some similarity between books by the
same author and differences between authors. When the sequential windows were included,
some were observed to coincide with the random clusters. As an illustration, Figure 2.15
shows the results for the book Moby Dick where the Zipf-Mandelbrot parameters A and C
are plotted for each of the 200 random samples and 15 sequential windows.
The random points are observed to form clusters, and some of the sequential points
54
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.15: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Moby Dick. Each window contains 5,000 words; results for
50 random and 42 sequential windows are shown. This graphic and the others in this style
were all produced using R. The basic plot function was used to create the initial graph with
the points function being used to add additional points in a different style on top of the
initial plot.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
55
coincide with them. This gives some confidence that the process to fit the Zipf-Mandelbrot
distribution is indeed finding correct distributions. The sequential windows are often signficantly far from the clusters formed from random samples. The same graphical representation
for the collection of short stories The Piazza Tales also by Herman Melville is shown in Figure 2.16. It is noticeable that the main clustering corresponding to the random samples
FIGURE 2.16: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within The Piazza Tales. Each window contains 5,000 words; results
for 50 random and 15 sequential windows are shown.
is similar to that of Figure 2.15. It is also noticeable that the sequential windows are also
outliers from the main cluster with some similarities in their distribution. As was shown in
Figure 2.1 sequential samples of words exhibit more variation than can be explained from
the assumption that the words are randomly arranged within the windows and so it is not
surprising that the sequential window points diverge from the random points. It is, however,
noteworthy to see that it is the variation in Zipf-Mandelbrot parameters that is showing
this.
However, the differences are not so clear-cut for the other books. Firstly, the two Jane
Austen novels Sense and Sensibility and Mansfield Park are shown in Figures 2.17 and 2.18,
respectively. The general shape of the clustering is different from the novels by Melville (the
limits for the axes have been arranged to be the same to aid comparison) but there does
seem to be some empirical similarity between the books. The sequential samples seem to
show some evidence that they are outliers from the random samples.
Secondly, the two Sherlock Holmes books are compared in Figures 2.19 and 2.20, respectively. The two books show similarities and there is some slight evidence that the points for
the sequential windows are outliers. A more significant number of the random and sequential points have a value for the A parameter near 1. This suggests that the range for this
parameter during the fitting process could be extended.
It is interesting to note that novels written by the same author seem to show similarites
56
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 2.17: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Sense and Sensibility. Each window contains 5,000 words;
results for 50 random and 23 sequential windows are shown.
FIGURE 2.18: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within Mansfield Park. Each window contains 5,000 words; results
for 50 random and 31 sequential windows are shown.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
57
FIGURE 2.19: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within The Return of Sherlock Holmes. Each window contains 5,000
words; results for 50 random and 21 sequential windows are shown.
FIGURE 2.20: Zipf-Mandelbrot scatter plot for A and C parameters for random samples
and sequential windows within The Adventures of Sherlock Holmes. Each window contains
5,000 words; results for 50 random and 20 sequential windows are shown.
58
Text Mining and Visualization: Case Studies Using Open-Source Tools
when the random sample are considered, although the sample size is too small to draw firm
conclusions. It is also interesting to note that the sequential windows show some divergences
from random, but more work would be needed to determine if there is a consistent pattern
to this. This is discussed more in Section 2.7.
2.7
Discussion
The results show evidence that fitting the Zipf-Mandelbrot distribution to randomly
selected 5,000-word samples from an author’s work leads to similarities between books by
the same author. Given the small number of books considered, this result would need a
larger sample to allow firmer conclusions to be drawn.
There is some evidence that sequential windows fitted to the Zipf-Mandelbrot distribution show variations that differ from what would be expected from a purely random sample.
This shows up particularly for the two novels by Herman Melville. There are some questions
to be resolved however.
Firstly, the genetic algorithm approach can sometimes fail to converge on a global solution. The results from the random sample fitting and the sequential window coincidence
with these show evidence that reasonable solutions are being found but it is by no means
certain. This can often be resolved using more iterations although this inevitably would require more processing time. As mentioned in Section 2.6.2, increasing the parameter ranges
would widen the search space and may allow better global solutions to be found or it might
worsen the situation. The only way to determine this is by further investigation. It may be
wise to use an alternative method to fit the distribution to the observed data to check the
algorithm is finding the best solution.
Secondly, the genetic algorithm itself may also repay more analysis. The RapidMiner
environment treats operators as black boxes and detailed knowledge of how they work is not
needed, but there could be some hidden feature of the evolutionary algorithm that causes it
to behave in a way that causes global solutions to be missed. The evolutionary operator has
many parameters, and it would certainly be valuable to explore how these vary the solution
finding ability of the algorithm. The RapidMiner product does have an open-source version
(5.3), and consequently it is possible to examine its source code. The version of the code
used in this chapter is 6.1 so there is a slight possibility that the code has been modified
since the open-source version was released.
Thirdly, the sequential variations are observed empirically by visual inspection to be
different from the random ones, but more statistical rigour is needed to confirm this to
eliminate any wishful thinking.
Lastly, as discussed in Section 2.3.2, probabilities for words are dependent on the size of
the text from which they are taken. This makes comparing works even by the same author
difficult and calls into question the value of the absolute parameters derived from the fitting
process. This means that any summary statistic that captures variation should endeavour
to be independent of sample size to allow comparisons between works to be made more
easily. The Zipf-Mandelbrot parameter set may be a good candidate given the empirical
law that they are derived from, and it is on this basis that they may be a good starting
point for more detailed investigation.
Empirical Zipf-Mandelbrot Variation for Sequential Windows within Documents
2.8
59
Summary
This chapter has shown detailed and advanced RapidMiner processes that determine
the optimum fit of the Zipf-Mandelbrot distribution for sequential windows and random
samples within documents. The distribution of the fitted parameters for all the windows
within a document were compared by the author. Some evidence, with caveats, that the
variation of the fitted parameters is an author-specific trait was observed for the random
samples, and divergences exhibited by the sequential windows was also observed. Future
investigation may be valuable to determine if the caveats can be addressed and whether
there is a genuine style variation that could be exploited to identify authorship.
Bibliography
[1] R.H. Baayen. Word frequency distributions, volume 18. Springer, 2001.
[2] F. Mosteller and D.L. Wallace. Inference in an authorship problem. Journal of the
American Statistical Association, 7:275–309, 1963.
[3] J. Burrows. delta: A measure of stylistic difference and a guide to likely authorship.
Literary and Linguistic Computing, 17(3):267–287, 2002.
[4] H.S. Heaps. Information retrieval: Computational and theoretical aspects. Academic
Press, Inc., 1978.
[5] G.K. Zipf. Selected studies of the principle of relative frequency in language. Harvard
Univ. Press, 1932.
[6] G. Zipf. Human behaviour and the principle of least-effort. Addison-Wesley, Cambridge,
MA, 1949.
[7] D.C. van Leijenhorst and T.P. Van der Weide. A formal derivation of Heaps’ Law.
Information Sciences, 170(2):263–272, 2005.
[8] L. Lü, Z.K. Zhang, and T. Zhou. Zipf’s law leads to Heaps’ law: analyzing their relation
in finite-size systems. PloS one, 5(12):e14139, 2010.
[9] A.W. Chisholm. An investigation into Zipf’s law and the extent of its use in author
attribution. Master’s thesis, Institute of Technology Blanchardstown, IE, 2012.
[10] P. Juola. Authorship attribution. Foundations and Trends in Information Retrieval,
1(3):233–334, 2006.
[11] M. Popescu and L.P. Dinu. Comparing statistical similarity measures for stylistic
multivariate analysis. Proceedings RANLP 2009, 2009.
[12] J. Rybicki and M. Eder. Deeper delta across genres and languages: do we really need
the most frequent words? Literary and Linguistic Computing, 26(3):315–321, 2011.
[13] S. Argamon and S. Levitan. Measuring the usefulness of function words for authorship
attribution. In Proceedings of ACH/ALLC, volume 5. Citeseer, 2005.
60
Text Mining and Visualization: Case Studies Using Open-Source Tools
[14] B. Mandelbrot. An informational theory of the statistical structure of language. Communication Theory, pages 486–502, 1953.
[15] Project Gutenberg. [Last Accessed: July 2015, Available at: http://www.gutenberg.
org].
[16] I. Mierswa, M. Wurst, R. Klinkenberg, M. Scholz, and T. Euler. Yale: Rapid prototyping for complex data mining tasks. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 935–940. ACM,
2006.
Part II
KNIME
61
This page intentionally left blank
Chapter 3
Introduction to the KNIME Text Processing
Extension
Kilian Thiel
KNIME.com, Berlin, Germany
3.1
3.2
3.3
3.4
3.5
3.6
3.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1
Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Philosophy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.1
Reading Textual Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2
Enrichment and Tagging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2.1
Unmodifiabililty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2.2
Concurrency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2.3
Tagger Conflicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.3
Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.3.1
Preprocessing Dialog . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.4
Frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2.5
Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1
Document Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2
Term Cell . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Table Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Example Application: Sentiment Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.1
Reading Textual Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.2
Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.3
Transformation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.5.4
Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
63
64
64
66
66
68
68
68
68
70
72
72
73
73
74
74
75
76
76
77
77
78
79
Introduction
The Konstanz Information Miner (KNIME ) [1] is a modular and open data analytics
platform that enables easy visual assembly and interactive execution of data workflows. Its
visual workbench combines data access, data transformation, initial investigation, predictive
analytics, and visualization. The KNIME Analytics Platform is open-source and available
under GPL license.
KNIME workflows are assembled of modules or nodes. Each node is a single processing
unit that processes the input data and/or model(s) and produces results on its outputs.
An example workflow can be seen in Figure 3.1. Individual settings that affect the node’s
execution can be configured in the dialog of each node. Data are passed from node to node
63
64
Text Mining and Visualization: Case Studies Using Open-Source Tools
via data tables that consist of data cells, encapsulating primitive as well as complex data
types, such as textual documents, networks, images and more.
The KNIME Text Processing extension is an extension of the open-source KNIME Analytics Platform. It is designed to read, enrich, manipulate, and extract textual data, and to
transform it into numerical representations, such as document or term vectors. Once numerical vectors are created, regular KNIME nodes can be applied, for example, for predictive
modeling, clustering analysis, or visualization.
The extension allows for reading and parsing of texts in various formats. It provides
nodes and models for part-of-speech (POS) tagging as well as named entity recognition.
Texts can be filtered, stemmed, and preprocessed in many ways. Different frequencies and
scores for terms in documents can be computed, keywords and topics can be extracted, and
documents can be visualized by tag clouds.
This chapter is organized as follows: installation of the Text Processing extension is
explained first. In section 3.2, philosophy and usage is described, meaning which nodes
can be applied in which order, what needs to be done, and what can be done and during
which step. Section 3.3 describes two new data types storing documents and terms. In
section 3.4 different structures of KNIME data tables are explained, which are required as
inputs by certain nodes. Section 3.5 demonstrates most of the theory about the usage of
Text Processing nodes, the data types and tables by an example application. This section
describes how a classification model can be built on textual data to predict sentiment labels
of movies reviews. The last section concludes this chapter.
3.1.1
Installation
Installing KNIME Analytics Platform and the Text Processing feature is easy. KNIME
is open-source and the latest version can be downloaded from www.knime.org. Versions
for Windows, Linux and MacOS are available as zip archives or installer executables. For
an easy and straight forward installation it is recommended to download and execute these
installers.
Additional extensions, such as the Text Processing extension can be installed conveniently via the KNIME update mechanism. An installation wizard can be started in KNIME
under “File → Install KNIME Extensions”. All the available extension categories are shown,
e.g., KNIME Extensions, KNIME Labs Extensions, or various Community Extensions. The
Text Processing extension can be found in the KNIME Labs Extensions category. Many
more extensions are available for installation, such as extensions for network analysis, XML
processing, image processing, or computational chemistry. For more details about installing
KNIME extensions see http://www.knime.org/downloads/update.
3.2
Philosophy
The order in which Text Processing nodes are organized in a workflow is important. On
the one hand, the nodes require specific structural specification of the input data tables, on
the other hand, the order of certain nodes affects the result of the analysis. For example,
before a filter node—filtering nouns verbs or adjectives—can be used on a set of documents,
a POS Tagger Node has to be applied in advance. This assigns POS tags to all terms of the
documents. Creating a bag-of-words with the BoW Creator node from a set of documents
requires a data table with one column containing documents (document cells) as input. This
Introduction to the KNIME Text Processing Extension
65
node can only be configured and executed properly with an input data table, which consists
of a column of documents. The output data table of this node is a bag-of-words. The BoW
Creator node, and all other nodes of the Text Processing extension require input data tables
of specific structure and create output data tables of a specific structure. Consequently,
certain types of nodes can be used only in a certain order. This order or configuration is
the basic philosophy behind the KNIME Text Processing extension.
To process and analyze textual data with the KNIME Text Processing feature, usually
five important steps have to be accomplished, although not all of them are mandatory:
1. Reading textual data
2. Semantic enrichment by named entity recognition and POS tagging
3. Preprocessing
4. Frequency computation
5. Transformation
The Text Processing node repository contains a category folder for each of these steps. An
example workflow including the nodes for all of the steps is shown in Figure 3.1.
FIGURE 3.1: An example workflow illustrating the basic philosophy and order of KNIME
text processing nodes.
In this example workflow textual data are loaded from a CSV file and converted into
KNIME document cells. Then POS tags are assigned, followed by a cascade of preprocessing
66
Text Mining and Visualization: Case Studies Using Open-Source Tools
nodes, such as filters, case converters and stemmers. Then a bag-of-words is created, term
frequencies are computed, and document vectors are created. These vectors are then used
for predictive model building and clustering analysis, both with regular KNIME nodes. In
addition, topics and keywords are also extracted from the preprocessed documents. In the
following sections each of the five steps will be described in detail.
3.2.1
Reading Textual Data
In the first step, textual data have to be loaded into KNIME and converted to document
cells. Document cells are a specific KNIME data type that can be handled by all nodes of
the Text Processing extension. These cells contain documents represented in a complex data
structure, which allows storage of and access to text at a word and sentence level and also
to meta information such as authors, category, or source. More details about these data
types can be found in section 3.3.
There are two possible approaches to creating document cells. The first option is to
read text from document format files, for example, PDF. The Text Processing extension
provides a set of nodes that parse document format files, extract the textual data, and
create a document cell for each file. These parser nodes are contained in the folder /KNIME
labs/Text Processing/IO of the KNIME node repository. Nodes to parse Microsoft Word,
PDF, flat text files, and more are available.
However, texts are not always available as PDF or Word files. The second approach
therefore involves converting textual data, represented as KNIME string cells, into document
cells. KNIME Analytics Platform provides many nodes to do so. Data can be imported from
files of various formats e.g., CSV and CSV-like formats, XLS, XML, and many more. Data
can be selected from databases or loaded from the Web. To convert textual data from string
cells into document cells the Text Processing extension provides the Strings to Document
node. The node creates a document cell for each row of the input data table. In the node
dialog it can be specified which columns contain the data for which field of the document,
such as title, text, authors, or category.
All parser nodes, as well as the Strings to Document node, apply word tokenization and
sentence detection to the texts to create the document cells. The output of these nodes is a
data table consisting of a column with document cells. An example data table is illustrated
in Figure 3.2. The icon, shown in the header of the second column, indicates that the data
type of this column is document cells. This kind of document list can be used as input by
all nodes of the enrichment and preprocessing category.
3.2.2
Enrichment and Tagging
The second step consists of enriching and tagging terms in the documents. This step is
not mandatory for all text mining applications. Semantic information is added by named
entity recognition and POS tagging. The folder /KNIME labs/Text Processing/Enrichment
in the node repository contains so called Tagger nodes. These nodes scan documents of the
input data table and recognize named entities, for example, names of persons, organizations,
or locations, biomedical-named entities, such as names of genes or proteins and chemical
compounds. Terms that have been recognized as entities are labeled by tags. All Tagger
nodes require an input data table containing exactly one column of document cells. The
output data table consists again of exactly one column of documents with tagged recognized
named entities.
A tag value is assigned to each term recognized as a named entity, e.g., “person” and
a tag type e.g., “NE” for named entity. The tag type represents the domain or type of a
Introduction to the KNIME Text Processing Extension
67
FIGURE 3.2: A data table with a column containing document cells. The documents are
reviews of Italian restaurants in San Francisco.
tagger (e.g., biomedical named entities or chemical named entities). The value represents a
particular characteristic in that domain, e.g., “gene” in the biomedical field.
The POS Tagger node, for example, assigns English POS tags. The assigned tag type
is “POS”, and the values are those of the Penn Treebank tag set [2]. The node applies the
POS model of the Apache openNLP project[3].
The Stanford Tagger node uses the models of the Stanford Natural Language Processing
Group [4] to assign POS tags. The node can be applied to English, German, and French
texts. The corresponding models can be selected in the node dialog. The STTS Stuttgart
Tübingen tag set [5] is used for German texts, and the French Treeblank tag set [6] for
French texts.
Each tagger assigns tags of its own domain and thus uses its own tag type and set of
values. Based on these tag types and values, filtering can be applied subsequently and in
such a way as to extract and visualize the named entities.
Figure 3.3 shows a column that contains terms (term cells). In addition to the word
of the term, the assigned tags are also listed. For example the POS tag “NNS” has been
assigned to the term “People”, indicating a noun plural. The term “Italian” has the tag
“JJ” assigned, indicating an adjective.
To identify and tag standard named entities (e.g., names of persons) the OpenNLP NE
Tagger node can be applied, which uses models from the Apache openNLP project [3]. For
biomedical entities (e.g., names of genes) the Abner Tagger node is applied, which uses the
ABNER [7] model. Chemical compounds can be recognized and tagged by the Oscar Tagger
node based on the model of the OSCAR chemical named entity recognizer framework [8].
The KNIME Text Processing extension also features the Dictionary Tagger node and
the Wildcard Tagger node. These nodes allow for the specification of a dictionary of named
entities to identify. The dictionary of the Wildcard Tagger may contain wild cards and
regular expressions that are used to search for specific terms. These two nodes require a
second input data table. This table must consist of at least one column that contains string
cells, which represent the dictionary entries. The dictionary data table can be created using
existing KNIME nodes, such as the File Reader, or the Database Reader.
68
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 3.3: A column of a data table containing term cells. The terms have been assigned
POS tags (tag values and tag types).
Due to the open nature of KNIME, additional tagger nodes as well as custom tag sets
can also be integrated.
3.2.2.1
Unmodifiabililty
Recognized and tagged named entities can be set as unmodifiable in order to prevent
them from being separated, manipulated, or filtered by subsequent nodes of the preprocessing category. Usually recognized named entities should not be filtered or manipulated
by subsequent nodes in the workflow. It does not make much sense to stem e.g., a first
or last name of a person. To avoid their manipulation of preprocessing nodes, these terms
are flagged unmodifiable by the corresponding tagger node by default, except for POS tagger nodes. In the dialog of each tagger node it can be specified whether recognized named
entities are flagged unmodifiable or not.
Figure 3.4 shows the dialog of the OpenNLP NE Tagger node. The first checkbox in the
Tagger options tab of the dialog specifies whether named entities are flagged unmodifiable
or not.
3.2.2.2
Concurrency
Tagging can be computationally expensive, depending on the used method and model.
To make use of all CPU cores all tagger nodes are parallelized. The data of the input data
table are split up into chunks. The data chunks are then processed in parallel.
In the dialogs of all tagger nodes it is possible to specify the number of parallel threads,
as can be seen in Figure 3.5. Each thread loads a separate tagging model into main memory
to process its data chunk. Note that some models (e.g., from the Stanford Natural Language
Processing Group) do require some hundred megabytes of memory.
3.2.2.3
Tagger Conflicts
If two tagger nodes are applied one after the other, the latter will overwrite the tagging of
the former in the event of conflicts. For example, if, first of all, the OpenNLP NE Tagger node
recognizes the term “Megan Fox” as the name of a person, and then the Dictionary Tagger
nodes recognizes “Megan” as the name of a product (automobile), based on a product list as
a dictionary, the previously recognized term “Megan Fox” is split up and “Megan” (without
“Fox”) is tagged as a product name.
Introduction to the KNIME Text Processing Extension
69
FIGURE 3.4: Dialog of the OpenNLP NE Tagger node. The first checkbox allows for specification as to whether or not the named entities should be flagged unmodifiable.
FIGURE 3.5: Dialog of the OpenNLP NE Tagger node. The number of parallel threads to
use for tagging can be specified here.
70
3.2.3
Text Mining and Visualization: Case Studies Using Open-Source Tools
Preprocessing
The third step is the preprocessing step in which terms are filtered and manipulated.
Filtering is necessary in order to get rid of terms that do not contain very much content
or information considered undesirable as features of document vectors. Manipulation (e.g.,
stemming or case conversion) is performed to normalize terms. All preprocessing nodes are
contained in the /KNIME labs/Text Processing/Preprocessing folder of the node repository.
Besides the regular preprocessing nodes such as stemming or stop word filtering, etc.
there are various other preprocessing nodes provided by the Text Processing extension to
manipulate and filter terms. For each tag type there is a corresponding filter node, which
filters terms with certain tag values. The tag values to filter can be specified in the dialog
of the nodes. The Stanford Tagger node, for example, assigns POS tags from the “STTS”
tag set to German texts. Correspondingly, the STTS filter node filters terms have certain
“STTS” tags assigned. This combination of tagging and filtering allows for a very powerful
identification and extraction of named entities of different types.
Other very powerful preprocessing nodes are the RegEx Filter node and the Replacer
node, which both work based on regular expressions and can be specified in the node
dialog. The Snowball Stemmer node allows for the stemming of texts in many languages,
for example, English, German, Spanish, and Italian. Internally, the node uses the Snowball
stemming library [9]. The Dict Replacer node replaces certain terms specified in a dictionary
with other terms, which are also specified in the dictionary. For example, this node allows
terms to be replaced by their synonyms.
Figure 3.6 shows a typical chain of preprocessing nodes used in many text mining applications. First punctuation marks are removed. Note that information indicating the beginning and ending of sentences is not lost when punctuation marks are removed. It is only
the punctuation characters that are erased from the terms. Next, terms consisting of digits
only are filtered, as well as those terms composed of fewer than three characters. Finally
terms are converted to lowercase, and stemming is applied.
FIGURE 3.6: Typical chain of preprocessing nodes to remove punctuation marks, numbers,
very small words, stop words, conversion to lowercase, and stemming.
Since the introduction of KNIME Text Processing Version 2.9, it has been possible to
apply preprocessing nodes directly to lists of documents as tagger nodes from the enrichment
category. Furthermore, preprocessing nodes can be also applied on bag-of-words data tables.
However, this is not recommended due to the longer processing time.
3.2.3.1
Preprocessing Dialog
The dialogs of all preprocessing nodes contain a “Preprocessing” tab, shown in Figure 3.6. Three settings can be specified in this tab:
Introduction to the KNIME Text Processing Extension
71
1. Deep preprocessing: this option is selected by default. It specifies whether the preprocessing step (e.g., stop word filtering) is applied to the terms contained in each
document. This option needs to be checked if the node is applied to a list of documents,
otherwise the terms in the documents are not filtered or manipulated. It only makes
sense to uncheck this option if the input data table of the nodes is a bag-of-words. In
this case only the terms in the term column of the bag-of-words are processed but not
in the documents themselves.
2. Appending unchanged documents: if deep preprocessing is applied, terms inside
the documents are preprocessed. Sometimes it can be useful to keep the original
documents. This option is checked by default. The original, unchanged documents are
appended in an additional column in either a bag-of-words or a document list data
table.
3. Ignore unmodifiable flag: by default all preprocessing nodes do not apply preprocessing on terms that have been flagged unmodifiable beforehand by any tagger node of
the enrichment category. This unmodifiable flag can be ignored and filtering or manipulation be applied to unmodifiable terms as well. By default this option is unchecked.
FIGURE 3.7: The Preprocessing tab of the Stop word Filter node. Deep preprocessing is
applied, original documents are appended, and unmodifiable terms are not filtered.
72
3.2.4
Text Mining and Visualization: Case Studies Using Open-Source Tools
Frequencies
After preprocessing is finished, frequencies of terms in documents and the whole corpus can be computed. All nodes that compute frequencies are contained in the /KNIME
labs/Text Processing/Frequencies folder of the node repository.
Nodes of the frequency category require a bag-of-words data table as input table. Thus
the data table containing the list of documents has to be transformed into a bag-of-words,
which can be done using the BoW Creator node.
Figure 3.8 illustrates a bag-of-words data table with one term column and two document columns. The original documents are stored in the column, “Orig Documents”. The
“Document” column contains the preprocessed documents. It can be seen that the words
“italian” and “pizzazz” are converted to lowercase. Furthermore, the stop word “with” has
been filtered. The data table structure is explained in more detail in Section 3.4.
FIGURE 3.8: Bag-of-words data table with one term column and two documents columns.
The column, “Orig Document” contains original documents. The “Document” column contains preprocessed documents.
The Text Processing extension provides nodes to compute well-known frequency measures (i.e. the TF node for term frequency, tf , and the IDF node for inverse document
frequency , idf ). In addition to inverse document frequency, there is also an the ICF node
which computes inverse category frequency , icf . This frequency is analog to idf , however
the number of documents in which a term occurs is not used as the divisor; rather the number of documents of a certain category is used. These nodes append an additional numerical
column to the input bag-of-words, which contains the frequency values.
Figure 3.9 shows a bag-of-words with one additional numerical column. The values
are absolute term frequencies computed by the TF node. Each term in each document is
assigned a frequency.
Based on the computed frequencies, terms can be filtered by the Frequency Filter node.
There are more nodes available in this category to create n-grams and count their frequencies
as well as to extract co-occurring terms and their number of co-occurrences.
3.2.5
Transformation
At a certain point the textual data have to be transformed into numerical vectors. A
vector has to be created for each document. This is usually performed by the Document
Vector node. The node creates a binary or numerical representation for each document based
on the filtered bag-of-words input data table. The vector representation can then be used
Introduction to the KNIME Text Processing Extension
73
FIGURE 3.9: Bag-of-words data table with an additional column with absolute term frequencies.
by standard data mining methods, such as clustering or classification nodes. Usually the
transformation into numerical vectors is the final step to be handled by the Text Processing
nodes.
The /KNIME labs/Text Processing/Transformation folder of the node repository contains all nodes that transform structures of data tables or that convert types of columns into
different types. The BoW Creator node can also be found in the transformation category.
The node requires a list of documents as the input table and creates a bag-of-words for
these documents. Furthermore, there are other transformation nodes, which allow for the
conversion of tags to strings, terms to strings, vice versa, and more.
3.3
Data Types
As already mentioned in previous sections, the KNIME Text Processing extension provides two KNIME data types: the document cell, which stores complex documents, and the
term cell, which stores terms. Both types can be handled and are required by nodes of the
Text Processing extension.
3.3.1
Document Cell
Document cells store documents in a complex data structure. The structure allows storage of and access to the text at a word and sentence level and also to meta information such
as authors, category, publication date, and more. A data table containing document cells
can be seen in Figure 3.2. The icon shown in the column header of the “Document” column
indicates that this column contains cells of the document type. Each cell in the data table
contains one complex textual document.
Table views that display document cells only show the title, as can be seen in Figure 3.2. The full text is usually too long to be visualized reasonably in a table cell. However,
documents can of course consist of more text than just a title. The Document Viewer node
can be used to view the full document text, including all meta information.
When a new document cell is created (e.g., by parser nodes or by using the Strings to
74
Text Mining and Visualization: Case Studies Using Open-Source Tools
Document node), word tokenization and sentence detection is applied automatically. Here,
word and sentence tokenizers are used from Apache’s openNLP project. Words are the
finest level of granularity in a document. Each word token, created by the word tokenizer, is
stored separately. Each word consists exactly of one string. Words can be transformed and
combined to terms by tagger nodes. A named entity tagger, for example, can combine the
two words “Megan” and “Fox” to the term “Megan Fox”. Thus, a single term can contain
one or more words. In addition, terms can have tags assigned. Terms, again, are combined
to sentences. Each sentence consists at least of one term.
Meta information, such as category, authors, publication date, and source are stored in
additional fields. Furthermore, generic meta information can be stored in documents as key
value pairs.
3.3.2
Term Cell
For the representation of terms, the term cell type is provided by the Text Processing
extension. A term can contain one or more words and none, one or multiple tags. After
creation of a document, each single word represents a term with no tags assigned. Tagger
nodes can change the granularity of terms and combine multiple words to a term.
Figure 3.3 illustrates a column containing term cells. The term cell icon can be seen in
the column heading. The string representation of a term, shown in table views, is assembled
by its words first, followed by the attached tags in brackets. The name of the tag is depicted
first (e.g., “NNS”) followed by the tag type (e.g., “POS”).
3.4
Data Table Structures
Certain nodes require a certain structure of the input data table in order to configure
and execute properly. There are three different kinds of structures:
1. Document list: one or more document columns, as shown in Figure 3.2.
2. Bag-of-words: at least one document and one term column, as shown in Figure 3.9.
3. Vectors (document or term): one document column and one or more numerical
columns representing terms or documents, illustrated in Figure 3.10.
Document lists are created by parser nodes or the Strings to Document node. Tagger
nodes as well as preprocessing nodes require a list of documents as the input data table.
These nodes can append an additional column containing the original unchanged document
right after the column with the processed documents.
A bag-of-words data table consists of tuples of terms and documents. Each row of a
bag-of-words represents the occurrence of a particular term in a document. This means
that for each unique term contained in a document, one row is created in the corresponding
bag-of-words data table. Depending on the number of documents and the number of unique
words in documents, bags of words can become huge.
The basic bag-of-words data table can be extended by numerical columns that contain
frequencies or the original unchanged documents, which is useful when deep preprocessing
is activated, as shown in Figure 3.9. It can be seen that the document titles of the column in
the middle differ from those of the right column. The original documents on the right still
have the original titles, whereas the titles of the documents in the middle column, affected
Introduction to the KNIME Text Processing Extension
75
by deep preprocessing, have been changed by preprocessing nodes (i.e. stop word filtering
and conversion to lowercase).
Document vectors are numerical representations of textual documents. The features of a
vector are all unique terms that occur in the whole set of documents. A value in the vector
greater than 0 indicates that the corresponding term occurs in a document, 0 indicates that
the term does not occur in the document. Alternatively to the Document Vector node the
Pivoting node can be used on the bag-of-words to create the vectors. As parameters of the
Pivoting node, the document column needs to be specified as group column and the term
column as pivot column.
Figure 3.10 illustrates a data table with document vectors of 10 documents. The
documents are reviews about dim sum restaurants in San Francisco. It can be seen that
all documents contain the terms “dim” and “sum”. Furthermore, in some documents the
terms “wife”, “city”, or “restaurant” occur.
FIGURE 3.10: Document vectors of 10 documents. The documents are stored in the leftmost column. The other columns represent the terms of the whole set of documents, one
for each unique term.
Analogous to the document vectors, term vectors can be created as well. A term vector
is a numerical representation of a term. The features of the vector are all unique documents.
The term vector matrix is the transposed document vector matrix. Term vectors can be
created with the Term Vector node, or alternatively again with the Pivoting node.
3.5
Example Application: Sentiment Classification
Sentiment analysis of free-text documents is a common task in the field of text mining. In
sentiment analysis predefined sentiment labels, such as “positive” or “negative” are assigned
to documents. This section demonstrates how a predictive model can be built in order
to assign sentiment labels to documents with the KNIME Text Processing extension in
combination with traditional KNIME learner and predictor nodes.
The data set used in this example is a set of 2,000 documents sampled from the training
set of the Large Movie Review Dataset v1.0 [10]. The Large Movie Review Dataset v1.0
contains 50,000 English movie reviews along with their associated sentiment labels “positive” and “negative”. For details about the data set see [10]. 1,000 documents of the positive
76
Text Mining and Visualization: Case Studies Using Open-Source Tools
group and 1,000 documents of the negative group have been sampled. The goal is to assign
the right sentiment label to each document.
Figure 3.11 illustrates the workflow used in this example. The textual data are imported
from a csv file and converted to documents. The documents are preprocessed and document
vectors are created. The class labels are extracted and colors assigned based on these labels.
Finally the data set is split up into a training and a test set, and a decision tree model
is built on the training set and scored on the test set. The nodes used for the creation of
documents and the preprocessing nodes collapse in two meta nodes.
FIGURE 3.11: Chain of preprocessing nodes of the Preprocessing meta node.
3.5.1
Reading Textual Data
The workflow starts with a File Reader. The node reads a csv file, which contains the
review texts, the associated sentiment labels, the IMDb URL of the corresponding movie,
and its index in the Large Movie Review Dataset v1.0. In addition to the sentiment column,
the text column is also important here. In the first meta node Document Creation document
cells are created from string cells, using the Strings to Document node. The sentiment labels
are stored in the category field of each document in order to extract the category afterwards.
All columns are filtered with the exception of the column that contains the document cells.
The output of the first meta node is a document list data table.
3.5.2
Preprocessing
The tagging step is skipped here since it is not necessary to recognize named entities
or assign POS tags. All preprocessing nodes are contained in the Preprocessing meta node,
shown in Figure 3.11.
First, punctuation marks are removed, numbers and stop words are filtered, and all
terms are converted to lowercase. Then the word stem is extracted for each term.
In the lower part of the meta node all those terms are filtered from the bag-of-words that
Introduction to the KNIME Text Processing Extension
77
FIGURE 3.12: Chain of preprocessing nodes inside the Preprocessing meta node.
occur in fewer than 20 documents. This is done by grouping by the terms, counting all the
unique documents that contain these terms, filtering this list of terms, and finally filtering
the bag-of-words with the Reference Row Filter node. Thereby we reduce the feature space
from 22,105 distinct terms to 1,500.
The minimum number of documents is set to 20 since we assume that a term has to
occur in at least 1% of all documents in order to represent a useful feature for classification.
This is a rule of thumb and of course can be changed individually.
3.5.3
Transformation
Based on these extracted terms document vectors are then created and used in the
following for classification by a decision tree classifier. In this example bit vectors were
created by the Document vector node. The values of bit vectors are 1 or 0 depending on
the presence of a term in a document. However, the previously computed tf values or any
other scores or frequencies computed beforehand could be used as vector values as well.
3.5.4
Classification
Any of the traditional mining algorithms available in KNIME can be used for classification (e.g., decision trees, ensembles, support vector machines). As in all supervised learning
algorithms a target variable is required (see, e.g., [11]). In this example the target is the
sentiment label, which is stored as category in the documents. The target or class column is
extracted from the documents and appended, as string column, using the Category to class
node. This category can then be used as the target class for the classification procedure.
Based on the category a color is assigned to each document by the Color Manager node.
Documents labeled “positive” are colored green; documents labeled “negative” are colored
red.
78
Text Mining and Visualization: Case Studies Using Open-Source Tools
The data set is then split up into a training (70%) and test set (30%). A decision tree
is trained on the training set and scored on the test set. The accuracy of the decision
tree model is 93.667%. Figure 3.13 shows the confusion matrix of the scorer node. The
FIGURE 3.13: Confusion matrix and accuracy scores of the sentiment decision tree model.
corresponding receiver operating characteristics curve can be seen in Figure 3.14.
FIGURE 3.14: ROC curve of the sentiment decision tree model.
The aim of this example application is to clarify and demonstrate the usage of the
KNIME Text Processing extension rather than achieve the best accuracy. Other learner
nodes, such as support vector machines, or decision tree ensembles can be applied easily
in KNIME in order to build better models. Furthermore, scores, such as tf − idf could
be used instead of bit vectors. In addition to features that represent single words, n-gram
features representing multiple consecutive words can be used as well. Cross validation could
be applied to achieve a more precise estimation of accuracy.
Introduction to the KNIME Text Processing Extension
3.6
79
Summary
In this chapter the KNIME Text Processing extension has been introduced. The functionality as well as the philosophy and usage of the extension have been described. It has
been shown which nodes need to be applied and in what order to read textual data, add
semantic information by named entity recognition, preprocess the documents, and finally
transform them into numerical vectors which can be used by regular KNIME nodes for clustering or classification. The KNIME data types provided by the Text Processing feature,
as well as the document cell and the term cell have also been explained. These techniques
were demonstrated as an application example, building a sentiment classification of movie
reviews.
More information and example workflows for clustering and classification of documents,
usage of tag clouds, and other visualizations can be found at http://tech.knime.org/
examples.
All results were generated with the open-source KNIME Analytics Platform. KNIME
workflows and data are provided with this chapter. The required KNIME extension is the
Text Processing extension.
Bibliography
[1] M. R. Berthold, N. Cebron, F. Dill, T. R. Gabriel, T. Kötter, T. Meinl, P. Ohl, C. Sieb,
K. Thiel, and B. Wiswedel. KNIME: The Konstanz Information Miner. In Studies in
Classification, Data Analysis, and Knowledge Organization (GfKL 2007). Springer,
2007.
[2] Mitchell M., G. Kim, M. A. Marcinkiewicz, R. MacIntyre, A. Bies, M. Ferguson,
K. Katz, and B. Schasberger. The penn treebank: Annotating predicate argument
structure. In Proceedings of the Workshop on Human Language Technology, HLT ’94,
pages 114–119, Stroudsburg, PA, 1994. Association for Computational Linguistics.
[3] Apache opennlp, 2014. http://opennlp.apache.org.
[4] C. D. Manning, M. Surdeanu, J. Bauer, J. Finkel, S. J. Bethard, and D. McClosky.
The Stanford CoreNLP natural language processing toolkit. In Proceedings of 52nd
Annual Meeting of the Association for Computational Linguistics: System Demonstrations, pages 55–60, 2014.
[5] Stts stuttgart tübingen tag set, 2014. http://www.isocat.org/rest/dcs/376.
[6] A. Abeillé, L. Clément, and booktitle=Treebanks pages=165–187 year=2003 publisher=Springer Toussenel, F. Building a treebank for french.
[7] B. Settles. ABNER: An open source tool for automatically tagging genes, proteins and
other entity names in text. Bioinformatics, 21(14):3191–3192, 2005.
[8] L. Hawizy, D. M. Jessop, N. Adams, and P. Murray-Rust. Chemicaltagger: A tool for
semantic text-mining in chemistry. Journal of cheminformatics, 3(1):17, 2011.
[9] Snowball stemming library, 2014. http://snowball.tartarus.org.
80
Text Mining and Visualization: Case Studies Using Open-Source Tools
[10] A. L. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, and C. Potts. Learning
word vectors for sentiment analysis. In Proceedings of the 49th Annual Meeting of the
Association for Computational Linguistics: Human Language Technologies, pages 142–
150, Portland, Oregon, USA, June 2011. Association for Computational Linguistics.
[11] M. R. Berthold, C. Borgelt, F. Hoeppner, and F. Klawonn. Guide to Intelligent Data
Analysis: How to Intelligently Make Sense of Real Data, volume 42 of Texts in Computer Science. Springer-Verlag, 2010.
Chapter 4
Social Media Analysis — Text Mining Meets
Network Mining
Kilian Thiel
KNIME.com, Berlin, Germany
Tobias Kötter
KNIME.com, Berlin, Germany
Rosaria Silipo
KNIME.com, Zurich, Switzerland
Phil Winters
KNIME.com, Zurich, Switzerland
4.1
4.2
4.3
4.4
4.5
4.6
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
The Slashdot Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Text Mining the Slashdot Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Network Mining the Slashdot Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Combining Text and Network Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.1
Introduction
81
82
82
85
88
89
92
Social media channels have become more and more important for many organizations in
order to reach targeted groups of individuals as well as understand the needs and behaviors
of their users and customers. Huge amounts of social media data are already available
and growing rapidly from day to day. The challenge comes in accessing that data and
creating usable and actionable insights from it. So far, there are three major approaches
that are typically used to analyze social media data: channel reporting tools, overview
score-carding systems, and predictive analytics with focus on sentiment analysis. Each has
its useful aspects but also its limitations. In this chapter we will discuss a new approach
that combines text mining and network analysis to overcome some of the limitations of the
standard approaches and create actionable and fact based insights.
Sentiment analysis on the one hand is a common approach in text mining to estimate
and predict the attitude of users with respect to products, services or certain topics in
general. On the other hand centrality measures are used in the field of network analysis to
identify important nodes (users), so-called influencers. In our approach we combine these
two techniques in order to analyze the sentiments of influencers of a social media platform,
as well as the impact of users with certain sentiments. We show that participants who are
81
82
Text Mining and Visualization: Case Studies Using Open-Source Tools
very negative in their sentiment are actually not highly regarded as thought leaders by the
rest of the community. This result goes against the popular marketing adage that negative
users have a very high effect on the community at large. To explain and demonstrate the
approach we used the publicly available Slashdot data set. The results have been created
with the open-source KNIME Analytics Platform, the Text Processing extensions, and the
Network Mining extension.
This chapter is organized as follows. First, an introduction to the Slashdot data set is
given. Section 4.3 explains the sentiment analysis approach used to determine the attitude
of each user. In section 4.4 we show how the user network was created and influencers are
identified. Section 4.5 describes how the two techniques are combined and new insights have
been generated. Section 4.6 concludes this chapter.
4.2
The Slashdot Data Set
Slashdot1 is a popular website, which was created in 1997. It publishes frequent short
news posts mostly about technological questions and allows its readers to comment on them.
The user community is quite active with more than 200 responses to a thread tending to
be the rule rather than the exception. Most of the users are registered and leave comments
by their nickname, although some participate anonymously. The data we used are a subset
of the Slashdot homepage provided by Fundación Barcelona Media4.2 The subset contains
about 140,000 comments to 495 articles about politics from a total of about 24,000 users.
4.3
Text Mining the Slashdot Data
In a first step our goal is to identify negative and positive users, that is, to determine
whether the known (not anonymous) users express predominantly positive or negative opinions, attitudes, feelings, or sentiments in their comments and articles.
In order to measure the sentiment of a user a level of attitude is determined, which
measures whether a user writes his or her comments and articles mainly negatively or positively. The level of attitude can also be used to categorize the users afterwards. To categorize
sentiments a lexicon containing words (clues) uses their polarity values. The polarity of a
word specifies whether the word seems to evoke something positive or something negative.
Possible polarity values are: positive, negative, both, and neutral. Naturally, the lexicon is
incredibly important as not only the language but the contextual usage of the language for
the given audience is significant. With KNIME, the lexicon that is most appropriate for
your text data can be chosen freely, or alternately, KNIME can be used to build or modify
an available lexicon to suit your tasks.
Due to the lack of labeled training data a model that predicts sentiment labels, as shown
in Chapter 3 cannot be trained and used here. The predictive model described in Chapter 3
was trained on movie reviews. These reviews differ compared with Slashdot articles and
comments in terms of spelling type, vocabulary, topics, and domain. For this reason the
predictive model of the previous chapter cannot be used on the Slashdot data set.
1 http://slashdot.org/
2 http://www.knime.org/files/slashdot.zip
Social Media Analysis — Text Mining Meets Network Mining
83
For the Slashdot data, the sentiment analysis of the user comments and articles is based
on the MPQA subjectivity lexicon [1, 2], a publicly available lexicon. In this lexicon, the
words as well as their polarity have been identified and collected manually and automatically
from annotated and nonannotated data.
Before applying text mining, it is important to understand the structure of the Slashdot
data and how it is read into KNIME and transformed. An article is an initial contribution.
A comment is a note or a reply to an article or to another comment. Each article with all
its following comments and notes represents a document. Users write their comments and
notes on many documents. To quantify the user attitude, we then need to navigate through
all the documents and measure the amount of negativity and positivity the person has been
expressing.
As we have seen, a word can be seen as positive or negative just by itself or can vary
depending on the context. The frequency of negative and positive words throughout a
document defines the attitude of the document. Similarly the frequency of negative and
positive words, among all words used by specific users across all documents, defines the
attitude of the user. The more negative words used by a user, the more negatively the user
attitude is perceived. In contrast, the more positive words a user uses, the more positively
the user attitude is perceived.
We excluded the “anonymous” user from the analysis, since this represents a collection of many different users rather than a single user and therefore carries no interesting
information.
For each nonanonymous user, the frequencies of positive and negative words, respectively,
fpos (u) and fneg (u), are calculated over his or her contributions across all documents. The
difference between such frequencies defines the user attitude as λ(u) = fpos (u) − fneg (u).
Positive λ define positive users, and negative λ define negative users.
In the first part of the workflow built to text mine the Slashdot data in Section 4.2, a
Table Reader node and a few traditional data manipulation nodes read the data, remove
the anonymous user and isolate the posts. Each post is then converted into a document
cell to allow further text analysis operations. At the same time, another branch of the
same workflow reads data from the MPQA corpus, extracts the polarity associated with
each term, and creates two separate sets of words: the set of positive words and the set
of negative words. Finally, the Dictionary Tagger node associates a sentiment tag to each
word of the Document column.
Now that all words in the posts are tagged as positive or negative, we can proceed with
the calculation of the level of attitude for each post and for each user. Therefore, a bag of
words is created and absolute term frequencies are computed. Negative and positive term
frequencies are then aggregated over user IDs to obtain the total frequency of negative and
positive words for each user. The level of attitude of each user is then measured as the
difference between the two term frequencies. A picture of the text mining workflow used to
compute the sentiment score for each user can be seen in Figure 4.1.
We would like to categorize the users using only three categories “positive”, “neutral”,
and “negative” based on their level of attitude. We assume that the user level of attitude
is Gaussian distributed (Figure 4.2) around a mean value µλ with a variance σλ and that
most users around µλ are neutral. Therefore, we assume that users with a level of attitude λ
inside µλ ±σλ are neutral, while users with λ in the left queue of the Gaussian (λ ≤ µλ −σλ )
are negative users and users with λ in the right queue of the Gaussian (λ ≤ µλ + σλ ) are
positive users.
Based on the calculated values for µλ and σλ , the binning process results in 67 negative
users, 18,685 neutral, and 1,131 positive users. Figure 4.3 shows a scatter plot of all known
users. The x axis represents the frequency of positive words, and the y axis represents the
84
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 4.1: The text mining workflow used to compute the sentiment score for each user.
FIGURE 4.2: Distribution of the level of attitude λ by user, with −20 as minimum attitude
and 50 as maximum attitude.
Social Media Analysis — Text Mining Meets Network Mining
85
frequency of negative words used by a user. Negative users are colored red, positive users
green, and neutral users are gray.
FIGURE 4.3 (See color insert.): Scatter plot of frequency of negative words vs. frequency
of positive words for all users.
The most prolific users — that is, those that have written the most (positive and negative
words) — are positive or neutral users. The user with the highest number of words (positive
and negative) is “99BottlesOfBeerInMyF”, which can be seen in the right top corner of
Figure 4.3. However, he or she is not the user with the highest level of attitude.
The most positive user is “dada21” with over 1,992 positive words and only about 1,196
negative words. He or she is a frequent writer as well. A tag cloud of the 500 most frequent
nouns and adjectives of “dada21” can be seen in Figure 4.4. Stop words have been removed,
positive words are colored green, negative words are colored red, and neutral words are
colored gray. The most frequent word is “government” followed by “money” and “people”.
It is clear that there are more positive than negative words.
In contrast, the user with the lowest level of attitude is “pNutz”, with only 43 positive
words and 109 negative. This user is not a frequent writer. It seems that this user simply
wished to vent his anger once but did not want to participate in a reasonable conversation. A
tag cloud of the 500 most frequent nouns and adjectives of “pNutz” can be seen in Figure 4.5.
Stop-words have been removed, positive words are colored green, negative words are colored
red, and neutral words are colored gray. Here, the most frequent word is “stupid” and the
negative words outnumber the positive words.
The average word frequency (positive and negative) of positive users is at 327, much
more than that of negative users with 192. Thus, negative users do not write frequently.
86
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 4.4 (See color insert.): Tag cloud of user “dada21”.
FIGURE 4.5 (See color insert.): Tag cloud of user “pNutz”.
Social Media Analysis — Text Mining Meets Network Mining
4.4
87
Network Mining the Slashdot Data
Network analysis of social media data focuses on the relationships between individuals
using their communication on particular topics as the connectors between them. These
networks can become incredibly complex, but advanced networking techniques not only
identify the network, but can translate it into either graphical representations or solid
numeric features, which can be used for analytics.
FIGURE 4.6: Example of a network extracted from Slashdot where vertices represent users,
and edges comments.
The main goal of this network analysis is to detect leaders and followers based on their
activity in the Slashdot forum. In order to do so we create a network G(V, E) that reflects
the posts of the Slashdot forum users. In the network each vertex v ∈ V represents a user.
Comments in the forum are represented by directed edges e = (u, v) ∈ E with the source
vertex u representing the user who has commented on an article and the target vertex
v representing the user who has written the initial article. For example the connection
between “User1” and “User4” in Figure 4.6 indicates that “User1” has commented on an
article in the forum written by “User4”. After filtering all articles and comments submitted
by anonymous users, the created network consists of 20,912 vertices representing users and
76,437 edges representing comments.
In order to detect the main leaders and followers, we borrowed the Hyperlink Induced
Topic Search algorithm [3] from web analytics. This centrality index was mainly developed
to improve the results of web searches by discovering the most authoritative web pages for a
broad search topic. It is an iterative algorithm that assigns each vertex v ∈ V two different
nonnegative weights, called authority weight xhvi and hub weight y hvi . Starting with an
initial value of 1 for the authority and hub weight of each vertex the authority weight is
computed as
X
xhvi =
y hui
u:(u,v)∈E
and the hub weight as
y hui =
X
xhvi .
v:(u,v)∈E
After each iteration the weights are normalized to maintain the invariant.
A vertex v is assigned a high hub weight y hvi if it refers to many vertices with a high
authority weight. A vertex v is assigned a high authority weight xhvi if it is referenced by
many vertices with a high hub weight. Therefore, a high hub weight is assigned to users who
frequently react to articles posted by others; in contrast, a high authority weight describes
those users whose articles generate a lot of comments.
Figure 4.7 shows a scatter plot of the leader vs. follower score for all users in the net-
88
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 4.7: Scatter plot of leader vs. follower score for all users.
work. The x axis represents the follower score based on the hub weight, whereas the y axis
represents the leader score based on the authority weight.
The user that attracts immediate attention is “dada21”, who has the highest authority
weight of 1 and a very high hub weight of 0.7. This user thus receives a lot of comments from
other users in his or her posts (high leader score) and at the same time comments often on
other users’ articles and comments. When inspecting the posts one can see that this user is
indeed one of the most active users with regard to the political topics in Slashdot.
Another user that might be of interest is the user with synonym “Carl Bialik from the
WSJ”. This user has a very high authority weight of 0.6 but a very low hub weight of 0,
implying that he is followed by a very high number of people, but never responds to anyone
elses entries.
On the opposite side of the scatter plot, we find the user “Doc Ruby”. “Doc Ruby” has
the highest hub weight of 1 and only a moderate authority weight of 0.3, meaning that he
leaves a lot of comments on other users posts but rarely writes a post of his own, and, if
he/she does, rarely receives a comment. This makes him one of the top followers.
4.5
Combining Text and Network Mining
Text mining and network are now widely used as analytic approaches for revealing new
insights in social media data. However, each technique follows its own very specific goal.
In text mining, the emphasis is on translating the textual data into sentiment in a
Social Media Analysis — Text Mining Meets Network Mining
89
carefully controlled process, which places the emphasis on words and expressions within a
given context. However, the information about the actual creator of the text, the sentiment
expressed in it, and the counts and numbers of readers and responders cannot reveal the
relevance of that person with respect to all others in that community, nor can it reveal
the relative interactions between that person and other people and how they relate to each
other.
Network mining, on the other hand, does a very good job of identifying how individuals
interact with each other. It does not rely on a categorical captured “thumbs up” or “star
rating” of individuals to rate the importance of a person, but rather identifies those people
of influence and those that are followers through physical nodes and connectors. Network
mining is also very good at identifying anomalies, such as “I will vote for you and you will
vote for me”, which is a classic challenge for text mining used in sentiment analysis. And
yet powerful, networking analysis alone cannot provide us with any information about the
context.
Taking advantage of having both network mining and text processing available within
the KNIME environment, we combined the results from the sentiment analysis with the
results from the network analysis in order to better position each user inside his or her
community in terms of influence (leaders vs. followers) and sentiment (positive, neutral,
and negative users).
The workflow that integrates the text mining results with the network analysis results
can be seen in Figure 4.8.
In general the goal of marketing is to identify negative and neutral users and, by means
of dedicated operations, to convince them to become positive users. However, working on
all possible users might be impractical and expensive. Thus, from time to time a new
marketing trend emerges trying to identify only the top influencers and to act on them. An
even better strategy, in terms of saved resources, would be to identify only the top negative
and/or neutral influencers. Indeed, it might not be as effective to try to influence an already
positive user, since he or she is already writing mostly positive comments and an excessive
marketing action might push him/her train of thoughts in an unwanted direction.
The real power of predictive analytics and visualization comes when we combine both
scatter plots, as shown in Figure 4.9. The x axis represents the follower score, and the y axis
represents the leader score. In addition, users are colored by their attitude: red for negative
users; green for positive users; and gray for neutral users.
The top influencers are the users with the highest leader score. These users are most
likely to have a high impact on other users, since their articles and comments are widely
read and used as reference by other users. What first becomes clear is that very few negative
users have a high leader score. One such user who clearly requires further investigation is
“Vicissidude”, the top negative attitude leader. Virtually all other negative attitude users
are only occasionally followers and have almost no leader influence. This goes against the
popular marketing adage that says that all negative attitude users are relevant.
Positive users such as “dada21” or “Doc Ruby” would also not be relevant for marketing
actions since they are positive anyway, even though they have a high leader score.
On the other hand, we can clearly identify users like “Carl Bialik from the WSJ” who
would be top targets for a marketing campaign. They are neutral or negative and are
followed by many other users.
90
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 4.8: KNIME workflow that combines text and network mining.
Social Media Analysis — Text Mining Meets Network Mining
91
FIGURE 4.9 (See color insert.): Leader vs. follower score colored by attitude for all users.
Users with a positive attitude are marked green, users with a negative attitude red.
4.6
Summary
In this chapter we explained and demonstrated a new approach to analyze social media
data. The approach combines text mining and network analysis techniques in order to
create fact-based, heterogeneous, and actionable insights into social media user behavior.
This would not have been possible by using just one analytic technique.
Combining sentiment analysis from online forum posts together with reference structures
from the quotation network has allowed us to position negative and positive users in context
with their relative weight as influencers or followers in the underlying discussion forum. The
approach was demonstrated on the publicly available Slashdot data set. It can be applied
to other data sets that contain both text and networking relationship data.
The approach can be improved by applying a predictive model for sentiment classification that has been trained on labeled data of the used social media platform. Further
enhancements are the integration of additional relevant data sources, such as company
and product names, political parties known users of products, and so on. The additional
data would enhance the method capability of identifying, segmenting, and drilling down to
interesting user groups.
All results have been created with the open-source KNIME Analytics Platform. KNIME
92
Text Mining and Visualization: Case Studies Using Open-Source Tools
workflows and data are provided with this chapter. The required KNIME extensions are
Text Processing, Network Mining, and R Integration.
Bibliography
[1] T. Wilson, J. Wiebe, and P. Hoffmann. Recognizing contextual polarity in phrase-level
sentiment analysis. In Proceedings of the conference on human language technology
and empirical methods in natural language processing, pages 347–354. Association for
Computational Linguistics, 2005.
[2] MPQA subjectivity lexicon, 2014. http://mpqa.cs.pitt.edu/lexicons/subj lexicon.
[3] J.M. Kleinberg. Authoritative sources in a hyperlinked environment. Journal of the
ACM, 46(5):604–632, 1999.
Part III
Python
93
This page intentionally left blank
Chapter 5
Mining Unstructured User Reviews with
Python
Brian Carter
IBM Analytics, Dublin, Ireland
5.1
5.2
5.3
5.4
5.5
5.6
5.7
5.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.1
Workflow of the Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.2
Scope of Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.3
Technical Stack Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.3.1
Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.3.2
MongoDB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4
Python for Text Mining . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4.1
Sparse Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4.2
Text Representation in Python . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1.4.3
Character Encoding in Python . . . . . . . . . . . . . . . . . . . . . . . . . .
Web Scraping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.1
Regular Expression in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2.2
PillReports.net Web Scrape . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Cleansing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Visualization and Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.1
Python:Matplotlib . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.4.2
Pillreports.net Data Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.1
Classification with sklearn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.5.2
Classification Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Clustering & PCA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.6.1
Word Cloud . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
95
97
97
99
99
100
100
100
103
105
106
106
108
110
114
115
115
119
120
122
124
128
130
131
Introduction
Pillreports.net is an online database of reviews of Ecstasy pills. The site states its aim as
providing qualified information to consumers of Ecstasy about the quality of the pill in terms
of its effects, contents, and the possible presence of adulterants that may induce unwanted
or unexpected effects for the consumer.
In consumer theory, illicit drugs are experience goods, in that the contents are not
known until the time of consumption. As such, sites like Pillreports.net may be viewed as
an attempt to bridge that gap and to potentially flag instances where a particular identifiable
pill is producing undesirable effects.
95
96
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 5.1: Pillreports.net standard report.
Mining Unstructured User Reviews with Python
97
Each entry/record contains a set of nominal (Color, Logo, Rating . . . ), binary (Consumed,Tested, Warning), numeric (Height, Width), temporal, and geographic attributes.
There are two free text entries (Description, User Report). In addition, a series of Comments can be associated with each report. Figure 5.1 displays a screen shot of a typical
entry.
Table 5.1 presents descriptions and some summary statistics for each column, indicating
the original name as found on the Pillreports.net site and the name as stored in the database.
The mixture of data types in the presence of the free text entries presents an ideal
opportunity to examine if there are underlying patterns in the free text fields that can be
linked back to the data in the binary/nominal columns. This chapter will exploit this and
work through the following examples:
• A classification problem using the free text field Description: as the input variables
and the Warning: = Yes/No field as the class label.
• A cluster analysis of the User Report: field to determine if there are underlying identifiable patterns.
5.1.1
Workflow of the Chapter
The chapter presents a sequence of processes required to set the stage in order that the text
mining objectives can proceed:
• Scraping the data from Pillreports.net - 2.Web Scraping
• Data shaping and cleansing - 3.Data Cleansing
• Data visualization and exploration - 4.Data Visualization
• Classification - 5.Classification
• Clustering - 6.Clustering
The emphasized text refers to an iPython Notebook associated with each chunk of
the overall process. Each notebook presents a series of steps for each process, the Python
libraries used, and the actual code. The coding was completed using Python (Version 2.7.8).
In addition to Python, a MongoDB database was used to store the data as they were
being scraped from Pillreports.net. Links are provided to instructions for installing and
setting up MongoDB on your machine. A CSV file of the cleansed data is also provided. It
is possible to jump straight to the visualization and exploration processes using the CSV
file if desired.
iPython Notebook is a browser-based computation environment. A notebook is a
JSON document that can be converted to Python script using the iPython terminal. As the
code is stored in a JSON document, there are methods provided to convert it to other formats including HTML, Latex, and PDFs, providing very flexible sharing and documentation
possibilities.
5.1.2
Scope of Instructions
The information and discussion provided in this chapter will focus on the technical aspects
of implementing the previously outlined project workflow solely using Python and iPython
98
TABLE 5.1: Column descriptions and statistics.
DB
Name
ID:
mongo id:
Description:
Data
Type
ID
ID
text
User Report:
User Report:
UR Length:
Desc Length:
Language:
Language Percentage:
Texture:
Color:
Edges:
Logo:
Name:
Shape:
Height:
Width:
State/Province:
Country:
Date Submitted:
Last Updated:
Member Since:
text
continuous
continuous
nominal
continuous
text
text
text
text
text
text
continuous
continuous
text
nominal
date
date
date
Member Name:
User Admin:
text
binary
Total Comments:
continuous
Report Stars:
continuous
Report Votes:
continuous
Tested:
Mandelin Reagent:
Marquis Reagent:
Mecke Reagent:
Robadope Reagent:
Simons Reagent:
Rating:
Suspected Contents:
SC Category:
Warning:
Consumed:
binary
nominal
nominal
nominal
nominal
nominal
nominal
nominal
nominal
binary
binary
Texture:
Color:
Edges:
Logo:
Name:
Shape:
Height:
Width:
State/Province:
Date Submitted:
Last Updated:
Submitted By:
Report Quality Rating:
Tested:
Mandelin Reagent:
Marquis Reagent:
Mecke Reagent:
Robadope Reagent:
Simons Reagent:
Rating:
Suspected Contents:
Warning:
Consumed:
Description
Incremental ID for referencing webpages
System generated key
Free text field, focuses on describing what an ecstasy pill looks like, but
may vary.
Free text field, documents user experiences of consuming ecstasy pill.
Generated: string length of User Report:
Generated: string length of Description:
Generated: Language of User Report: free text field.
Generated: Confidence level of the prediction of Language: label
Series of columns detailing various attributes of the ecstasy pill under
review.
Geographic location of where report was submitted from, user entered.
Generated: based on State\Province: column
Date of report submission.
Date report was last updated or revised.
Generated: Date owner of report is member of the site based on Submitted by: column
Name and joining date of a member
Generated: Member name taken from Submitted By: column
Generated: Boolean value indicating if member is a site Administrator\Moderator
Generated: Total comments per report
Review of report with stars (1-5) and number of voters
Generated: Number of stars given to a report based on Report Quality
Rating:
Generated: Number of votes given to a report based on Report Quality
Rating:
Binary value Yes\No indicating if ecstasy pill has been tested.
Series of different types of testing kits and input of the value (typically
a color)
Rating of the pill in terms of believed active ingredients
User opinion on the active ingredients in the pill
Generated: Refined grouping of Suspected Contents:
Binary value Yes\No indicating if pill has unwanted effects.
Binary value Yes\No indicating if user consumed the pill.
Unique
Entries
5001
5001
4975
Blank
Cells
0
0
0
min
value
26940
max
value
34120
average
3225
1809
1131
8
3021
1743
1259
1092
2556
3496
989
67
64
1629
56
4988
3374
1386
1762
0
0
0
1981
1238
6
2080
20
1
7
3184
3169
1
0
0
1600
205
0
1
9336
5406
702
369
0.571
0.999
0.998
0.2
0.4
99.9
99.9
5.19
7.74
USA (2051), Australia (969), Netherlands (364)
2011-08-11
2014-11-14
2011-08-11
2014-11-14
2005-09-05
2014-11-14
2546
2
0
0
TRUE ( 205), FALSE (4796)
51
84
1
55
10.7
44
1252
0
4.8
3.97
22
1252
1
23
3.04
2
7
11
7
4
4
7
31
5
2
2
0
4014
2525
3667
4794
3972
0
0
0
0
0
yes(2746), no (2255)
Black (1375), Other (272), Purple (269)
Tested Only (1445), MDxx High (875), MDxx Medium (948), Adulterated (821)
MDMA (1821), MDxx (897), Unknown (828)
mdxx (2859), unknown (1038), amphet (504)
no (3184), yes (1817)
yes (3244), no (1757)
Text Mining and Visualization: Case Studies Using Open-Source Tools
Pillreports.net
Name
ID:
id
Description:
Mining Unstructured User Reviews with Python
99
Notebook. The rationale behind choosing Python is discussed, and the pros/cons of using
Python in the context of various steps in the data mining project are highlighted.
At various points throughout, an overview of some of the more thorny issues that a user
may experience with various Python libraries and coding conventions is presented.
By focusing on the technical aspects, there is light treatment given to the theoretical aspects of text mining and/or data mining algorithms. [1] provides a very good overview of the
data mining process and many of the most often used algorithms. For a more mathematical
and statistical focus to the data mining methods see [2].
The remainder of this section outlines the following:
• Setting up the technical stack
• A discussion on the support for sparse matrices support in Python
• A discussion on character encoding with ASCII and UTF-8 in Python
5.1.3
Technical Stack Setup
5.1.3.1
Python
Download Python and Python Notebook from the Anaconda distribution available at
https://store.continuum.io/cshop/anaconda/.
Anaconda will download most of the popular Python packages with its distribution. To
start with Python Notebook simply open a Command Prompt/Terminal window and
type ipython notebook. This will open up a browser window. Create a new notebook
and you can enter the commands shown in Figure 5.2. The second command will output
the Python version that you are running.
In [1]: print ("Hello World")
Hello World
In [2]: import sys
print sys.version
2.7.8 |Anaconda 1.9.2 (x86 64)| (default, Aug 21 2014, 15:21:46)
[GCC 4.2.1 (Apple Inc. build 5577)]
FIGURE 5.2: Code: Check Python setup.
As stated many of the libraries used in the chapter install with the Anaconda distribution. Anaconda provides a package manager "conda". It installs libraries from its own
package repository and as such not all libraries are available, especially more experimental libraries. The standard Python package manager is pip. This is used to download the
additional libraries.
There are five additional libraries used in the project that are not part of the Anaconda distribution. The installation instructions are listed below. Open a new Command
Prompt/Terminal and enter each command separately. Your machine must be connected
to the internet and it will grab each library and install it.
• pip install pymongo
100
Text Mining and Visualization: Case Studies Using Open-Source Tools
• pip install beautifulsoup4
• pip install langdetect
• pip install seaborn
• pip install git+git://github.com/amueller/word_cloud.git
5.1.3.2
MongoDB
MongoDB is an open-source database. It is a NoSQL database, sometimes referred to as a
document store. NoSQL stands for Not only SQL. In the context of this chapter NoSQL
is a nonrelational database. This has beneficial properties, in that it is very flexible with
regard to specifying the column names of the data being scraped.
This is important when doing a large Web scrape, where each webpage may contain
different/new data. The structure of the webpage may be the same but the content differs.
This is the case for Pillreports.net.
In this chapter, 5,001 records were scraped, resulting in 26 columns, of which 15 were
present for each page. Some of the columns only appeared in the middle of the scrape and
could not be determined in advance. Using a NoSQL DB as the initial repository for the
data meant this was not an issue. To understand it, think of when you fill out a form on a
website: only some of the entries are mandatory, these are the 15 columns.
Installing MongDB is relatively straightforward. It can be downloaded from http:
//mongodb.org/downloads. Setting up the initial instance can be a bit more tricky.
There are very good instructions for all the major operating systems at docs.mongodb.
org/manual/. The main thing to remember when initially setting up is that you need to
create a data directory called data/db where all the databases will be stored.
After the setup, apart from making sure the MongoDB server is running, all interaction
with the MongoDB will be done via the pymongo library.
5.1.4
Python for Text Mining
Before setting off into the coding for the chapter, it is worth taking a few minutes reading
over the next two sections. The first explains what sparse matrices are. In any text mining
project the ability to handle sparse table entries (lots of zeros) is crucial. The second contains a discussion on character encoding. This is within the context of Python but may be
applicable in other languages. Understanding how Python handles text will save you a lot
of frustration in the long run.
5.1.4.1
Sparse Matrices
Most text mining tasks require that the data are represented in a matrix or vector form. This
is referred to as a document-term-matrix (DTM) or the transpose term-document matrix
(TDM).
A collection of documents is referred to as a corpus. For example, take the Roald Dahl
poem, Hey Diddle Diddle, shown at the start of Figure 5.3.
Taking each line of the poem as its own individual document, the poem represents a
corpus of six documents. The corpus can be represented by a matrix as presented in Table
5.2.
A 6 × 25 matrix is created representing 6 documents and 25 unique words. This is not
very efficient as many of the entries are zero. In fact only 30 cells contain nonzero entries.
(Note the information about the order of the words is lost in the matrix representation.
Mining Unstructured User Reviews with Python
101
work
...
...
...
...
...
...
h
0
0
0
1
0
0
whic
0
0
0
0
1
0
we
cash
0
0
0
0
0
1
up
caref
ul ly
0
1
0
0
0
0
until
by
0
0
1
0
0
1
the
are
0
1
0
0
0
0
take
and
Line/Word
Line 1
Line 2
Line 3
Line 4
Line 5
Line 6
al l
TABLE 5.2: Example dense matrix.
0
0
0
1
0
0
0
1
0
0
0
2
0
0
1
0
0
0
0
0
1
0
0
0
0
1
0
1
1
1
0
0
0
0
1
0
0
0
0
0
0
1
Using ngrams of words can preserve some of the word order.)
A sparse matrix provides a more efficient way to store the information. Sparse matrix
store references to the nonzero elements only. There are a number of different methods of
implementing sparse matrices; dictionary of keys (DOK), list of list (LL) and co-ordinate
list (COO). The sklearn machine library in Python uses the latter method. COO stores the
information in {row, column, value} form where; row is a vector of row indexes indicating
the row location of the nonzero values, column is a vector of the column indexes indicating
the column location of the nonzero values and value are the nonzero values in the matrix.
The entries are sorted by row index and column index to enable faster search and
retrieval. A representation of the poem in COO format is shown in Table 5.3. There are 30
entries to represent the poem, compared with 150 when the zeros are included.
TABLE 5.3: Example sparse matrix.
Line
Line 2
Line 3
Line 6
Line 2
...
Line 5
Line 6
Line 5
Line 6
Word
all
and
and
are
...
we
we
which
work
Count
1
1
1
1
...
1
1
1
1
row,col,value
2,1,1
3,2,1
6,2,1
2,3,1
...
5,23,1
6,23,1
5,24,1
6,25,1
The Python sklearn.feature extraction.text method can be used to convert a corpus
of documents to a sparse matrix representation (COO) of word counts.
The Python code for creating the representation of the Hey Diddle Diddle poem is shown
in Figure 5.3. From this small toy example, the memory used is reduced by 1,448 - 464 = 784
bytes or 2/3. This memory saving scales exponentially as the size of the corpus grows.
A key advantage of using Python is that a sparse matrix can be passed to many of the
machine learning methods in the sklearn library and they will work, (a notable exception
being tree based algorithms including Random Forests). The ease and flexibility of this
102
Text Mining and Visualization: Case Studies Using Open-Source Tools
In [1]: import pandas as pd
poem="""Hey diddle diddle,
We are all on the fiddle,
And never get up until noon.
We only take cash,
Which we carefully stash,
And we work by the light of the moon."""
#Split poem by "," or ".". and save in dataFrame df
#Use ’re’ package - regular expressions
#Include newline \n splitting as will remove
import re
lines=re.split(’,\n|.\n’,poem)
df=pd.DataFrame({"Lines":lines})
In [3]: #Create Sparse and Dense Matrix
from sklearn.feature_extraction.text import CountVectorizer
countVec = CountVectorizer()
poemSparse=countVec.fit_transform(df[’Lines’])
poemDense=pd.DataFrame(poemSparse.todense(),
columns=countVec.get_feature_names())
In [4]: print "Dense matrix non-zero values:",((poemDense!=0).sum()).sum()
print "Sparse matrix entries:",len(poemSparse.data)
Dense matrix non-zero values: 30
Sparse matrix entries: 30
In [9]: #Size of the data frame in memory is
#values + index + column names
print "Dense matrix size:",poemDense.values.nbytes + \
poemDense.index.nbytes + poemDense.columns.nbytes, "bytes."
print "Sparse matrix size:",poemSparse.data.nbytes + \
poemSparse.indptr.nbytes + poemSparse.indices.nbytes, "bytes."
Dense matrix size: 1448 bytes.
Sparse matrix size: 464 bytes.
FIGURE 5.3: Code: Creating a sparse matrix in Python.
Mining Unstructured User Reviews with Python
103
representation can significantly speed a project’s development time and reduce computation
overheads.
5.1.4.2
Text Representation in Python
The data used in this project is scraped off the Web. As such it is impossible to control
what types of text will be returned. The text may contain accented characters { être — to
be in French, köszönöm — thank you in Hungarian } or indeed characters outside that of
the Latin alphabet; Cyrillic, Chinese Mandarin, and Arabic. Python 2.7 can produce some
confusing output when working with text.
In [8]: eng="thank you"
hun="köszönöm"
print ("String Length:",len(eng),"Type:",type(eng))
print(hun, "String Length:",len(hun),"Type:",type(hun))
#k,s,z,n,m print correctly
(’String Length:’, 9, ’Type:’, <type ’str’>)
(’k\xc3\xb6sz\xc3\xb6n\xc3\xb6m’, ’String Length:’, 11, ’Type:’, <type ’str’>)
In [9]: eng="thank you"
hun=u"köszönöm"
print(eng, "String Length:",len(eng),"Type:",type(eng))
print(hun.encode(’ascii’,’replace’), "String Length:",len(hun),
"Type:",type(hun))
#Type has changed. String Length of hun reduced 11 to 8.
(’thank you’, ’String Length:’, 9, ’Type:’, <type ’str’>)
(’k?sz?n?m’, ’String Length:’, 8, ’Type:’, <type ’unicode’>)
FIGURE 5.4: Code: Confusing output with text in Python.
These types of output may also occur when reading files in. The cause of this is character encoding. When computers were in their infancy, characters were encoded using ASCII
(American Code Standard Information Interchange). ASCII, created in 1963, original contained 128 bytes. The original ASCII coverage is shown in Figure 5.5.
Computers converse in bits and bytes. An ASCII encode is quite simply a lookup table.
Given the byte \x4D a program can look up the table and know to print out a capital M
and so on. The encoding standard was extended further with Latin-1 and Windows 1252
giving a range of 256 bytes on a lookup table. These include the original subset of ASCII
byte lookups.
Clearly 256 bytes is not enough to represent the full set of characters of all the languages
in existence. A proliferation of encodings ensued to address these shortcomings. Unicode was
and is the solution. Unicode is a catalog of over 1 million integers. Every possible character
in the world has its own unique integer. (Currently about 10% of the possible spaces are
used).
The Unicode-integer or code point is encoded according to a selected encoding standard
that translates it to bytes. These bytes can then be sent from computer to computer. Once
received they can be decoded to get the relevant code point and its character. The misuse
and interchange of the terms (encode, encoding, Unicode) can cause significant confusion.
Figure 5.6 illustrates the paradigm.
104
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 5.5: Original byte encoding — 256 characters.
FIGURE 5.6: Unicode encoding paradigm.
FIGURE 5.7: Character bytes and code points.
Mining Unstructured User Reviews with Python
105
UTF-8 is one such encoding standard. In fact it is the de facto standard with over 90%
of all webpages now encoded as such. Figure 5.7 shows the word pillreports and a mocked
up version of the same word using strange but recognizable characters from other alphabets.
The mapping between the characters and the original ASCII byte is shown. In the mocked
up version, the layer of hex code points between the character and the subsequent byte(s)
is displayed. The original ASCII characters have a code point that maps to the same byte
as back in 1963.
5.1.4.3
Character Encoding in Python
The previous section detailed the accepted Unicode Standard. This section outlines its
implementation in Python. There are some data type conventions that require explanation
and the reader should be aware that these differ significantly between Python 2.7.x and
Python 3.x.
Python 2.7 has two <basestring> types; <str> and <unicode>. <str> are bytes
whereas <unicode> is composed of unicode code points. The differences between the
<str> and <unicode> are illustrated in Figure 5.8. The code is entered in an Python
Notebook.
The unicode symbol for a smiley face is entered and printed. The corresponding code
point and byte length (of the code point) are printed. Note that the encoded has defaulted
to UTF encoding.
The <str> byte representation of the smiley face symbol is then printed out, 3 bytes
long. Finally an attempt is made to encode the symbol using ascii which throws an error
as ascii does not have a representation for a smiley face.
In [11]: #Create unicode symbol for a smiley face
#u\u263a - unicode code point
smiley=u"\u263A"
print(smiley)
In []: ord(smiley) # = 9786 decimal value of code point = 263A in Hex.
In []: len(smiley) # = 1 This is a <unicode>. Length in code-points
In []: # prints ’\xe2\x98\xba’ the bytes - it is <str>
# its length = 3 - 3 bytes
smiley.encode(’utf8’)
In []: #Encode unicode point with ascii
#Error no coverage in ascii for this code-point
print u"\u263A".encode(’ascii’)
FIGURE 5.8: Code: Python encoding for text correct use.
In the above example a u is prefixed before the string entered. This tells the Python 2.7
interpreter that this is a <unicode> type. If this was not included the interpreter would
assume it was <str> type. In Python 3.0 this is reversed, the default without a prefix is
<unicode> and a prefix b must be specified to get a <str> type.
Relating this back to text mining, a firm understanding of the difference between getting
an error message with \xe2\x98\xba relating to a <str> of bytes and similarly \u263A
relating to <unicode> will help debugging if they should arise.
106
Text Mining and Visualization: Case Studies Using Open-Source Tools
In order to avoid these errors, explicitly encode files with the correct encoding, (’UTF8’,ASCII), when reading a file at the start of a project and similar encoding a file when
writing it out after processing has completed.
The encoding must be known when reading in a file. However, it is reasonable to assume
that it is UTF+8 unless otherwise stated. You will see this convention used in all the code
accompanying this chapter. (Note if you open up the accompanying CSV files in Excel
and examine the free text columns, especially for non–English-language entries, you will
see strange symbols. Excel is using its default ANSI encoding for reading the file in, not
UTF+8 that the file was explicitly encoded with.)
For a very informative video on everything discussed in this section [3] provides an
excellent overview.
5.2
Web Scraping
There are three parts involved in Web scraping part of the chapter. Firstly we need to be
able to navigate to the correct location, the webpage. Secondly we need to be able to traverse
that webpage and select the parts that are of interest and finally we need somewhere to
store that data. The Python library urllib2 is used to navigate to the correct landing
page. The BeautifulSoup library traverses the webpage and the pymongo library is used
to save the data to a collection in a database located on the local MongoDB server.
urllib2 is a Python module that can be used for fetching URLs. The most basic way
to use the library is to send a request to a server (pillreports.net) and providing a successful
connection has been made, read in the data as a string.
BeautifulSoup takes advantage of the nested tree structure of HTML documents. It
automatically converts a document to unicode and parses the document. It is then possible
to traverse the document tree to find any relevant parts of the document. A little bit of
knowledge of HTML tags is required to supply the correct arguments. Figure 5.9 displays
an example of using both libraries.
Note the use of the .text method to extract the content from within a HTML tag. In
the example, the text is extracted from a string bounded by HTML tags in the form <td
>This is HTML in side tags <td >. In fact BeautifulSoup is using regular expressions
to achieve this. Note in the code proper, a regular expression is used instead of .text
method, in order to remove white space and extra HTML tags that may be within the text.
See the discussion in Section 5.2.1.
Once the information has been extracted from the webpage, it must be stored. A MongoDB is used as the database. The reasons for choosing MongoDB were outlined in Section 5.1.3.2.
The pymongo package is used to connect to a running MongoDB server. A new database
is created with two collections, one for the reports and the other for the associated comments.
For each webpage, a Python dictionary is created with the relevant data extracted. The
dictionary is then saved to the relevant collection, recursively (similar to inserting a row
into a table). Figure 5.10 illustrates these conventions.
Mining Unstructured User Reviews with Python
In [39]: from bs4 import BeautifulSoup
import urllib2
#Open webpage and scrape it.
web_path = ("http://www.pillreports.net/index.php?"
"page=display_pill&id=34087")
web_page = urllib2.urlopen(web_path).read().decode(’utf-8’)
#Convert to Beautiful soup object.
soup=BeautifulSoup(web_page)
In [40]: #Find all the tables in soup - findChildren
#Returns all <table> tags
no_tables = len(soup.findChildren(’table’))
print "Total Tables: "+str(no_tables)
#Find all the rows - find_all
#Returns all <tr> tags in <table>
no_rows = len(soup.find_all(’table’)[0](’tr’))
print "Rows in first Table: "+str(no_rows)
#Find the cells - find_next
#Returns the next HTML tag. Should be <td> for table cell
#but may not be.
first_cell=(soup.find_all(’table’)[0](’tr’)[3]).find_next()
second_cell=first_cell.find_next()
print "Text of First Cell in Row 4: "+first_cell.text
print "Text of Second Cell in Row 4: "+second_cell.text
Total Tables: 3
Rows in first Table: 21
Text of First Cell in Row 4: Date Submitted:
Text of Second Cell in Row 4: November 10, 2014, 4:28 pm GMT
FIGURE 5.9: Code: Scraping webpages.
In [ ]: #Python Dictionary
post={"Data Submitted:":"December 24, 2014",
"Rating:" : "Adulterated",
"Warning:" : "yes"}
import pymongo
#Connect to MongoDB server with default settings.
conn=pymongo.MongoClient()
db = conn[myfirstdatabase] #Create the database
reports = db.reports #Create collection in database
reports.insert(post) #Insert dictionary into collection
conn.close() #Close the connection
FIGURE 5.10: Code: Connecting to a database in MongoDB.
107
108
5.2.1
Text Mining and Visualization: Case Studies Using Open-Source Tools
Regular Expression in Python
Regular expressions (prefixed with r in Python) are a text pattern that can be used to
match content in a string. This can be for replacement, verification, splitting, etc. They are
very powerful in simplifying many text processing tasks. However they need to be used with
skill. Err on the side of simplicity when creating them.
Regular expressions use a wide array of special characters (and combinations) that cause
different behaviors. In Python these characters are ( . ˆ$ * + ? { } [ ] \ — ( ) - ). In addition
the letters (A,b,B,d,D,s,S,w,W,z) when prefixed with \ also have special behaviors.
It is beyond the scope of this chapter to go into a huge amount of detail. In the accompanying notebooks regular expressions are used at a number of points. Figure 5.11 provides
an overview of the main ways that they are employed.
The sub() function of the Python re module performs a search and replace across a
whole string. All matches are replaced with the specified string. It is possible to limit the
number of replacements by setting the count parameter to a nonnegative integer.
The findall() function returns all matches to a whole string. It is possible to set a
starting and ending position to begin/finish searching the string. The returned result will
be in a list.
Both methods make a call to compile() method; therefore, it is possible to create an
instance of this method with the parameters pre-set.
Placing parentheses () around a regular expression string, causes a regular expression to
find groups or pairs of matching strings returned in tuples. The tuples can then be indexed.
You may notice many similarities between the functions for the re module and the
BeautifulSoup library. BeautifulSoup is in part a very neat wrapper around regular expression targeted towards HTML code. For a comprehensive overview of regular expressions
[4] covers every possible aspect that you will ever need in many programming languages
including Python.
5.2.2
PillReports.net Web Scrape
This section describes the Web scraping of the pillreports.net site itself. A quick look at any
of the reports on the site, indicates that the data is very neatly structured inside sets of
HTML tables. Using a trial and error process, similar to the methods shown in Figure 5.9,
it is relatively easy to work out that there are three tables. The first contains the report
and third contains the associated comments.
The address of the website is an incremental ID for each report. It is straightforward to setup a loop to run through a number of iterations increasing the ID to move
through reports. In this chapter the most recent report was selected; at the time of the
scrape, 17/11/2014, this was www.pillreports.net/index.php?page=display_
pill&id=34120. The id was decremented (34120, 34119,. . . ) until 5,001 pages had been
captured. Note this required 7,180 iterations, as over 2,000 webpages did not return a result.
The pseudo code for the process is outlined in Algorithm 5.2.2. The full code for the
scraping of pillreports.net is provided in iPython Notebook: 2.Web Scraping.
In addition to the logical checks in place to ensure the structure for each webpage is as
expected, there are also checks to ensure that the encoding is correct and has worked. The
incremental reference system may have reports that are deleted or that are not published.
This will not affect the webscrape as these are simply discarded and an error message is
printed. The code can be used straight away provided there is a MongoDB server running
on your machine.
The two parameters that need to be changed are a) the number of reports to download and b) the starting place. These parameters are shown in Figure 5.12. The data will
Mining Unstructured User Reviews with Python
In [118]: import re
text="""Hey diddle diddle, We are all on the fiddle,
And never get up until noon. We only take cash,"""
In [119]: #1. Sub
print (re.sub(’We’, ’They’, text))
Hey diddle diddle, They are all on the fiddle,
And never get up until noon. They only take cash,
In [120]: #2. Findall
#{4,6} 4 to 6 in a row.
print (re.findall(r"[a-zA-Z]{4,6}", text))
#+ = 1 or more * = 0 or more repetitions.
print re.findall(r’n[o|e][a-z]*’,text)
#()Indicates grouping \w and letter
print re.findall(r’([\w]+)dd([\w]+)’,text)
[’diddle’, ’diddle’, ’fiddle’, ’never’, ’until’, ’noon’, ’only’, ’take’, ’cash’]
[’never’, ’noon’]
[(’di’, ’le’), (’di’, ’le’), (’fi’, ’le’)]
In [121]: #3. Search
#Returns first match only in groups.
a=re.search(r’([\w]+)dd([\w]+)’,text)
print a.group(0),a.group(1),a.group(2)
diddle di le
In [122]: #4. Using Compile
re_1=re.compile(r’We’)
re_2=re.compile(r’n[o|e][a-z]*’)
print re_1.sub(’They’,text)
print re_2.findall(text)
Hey diddle diddle, They are all on the fiddle,
And never get up until noon. They only take cash,
[’never’, ’noon’]
diddle
FIGURE 5.11: Code: Regular expressions.
109
110
Text Mining and Visualization: Case Studies Using Open-Source Tools
Algorithm Looping through pillreports.net webpages.
Require: target = positive integer (number of pages to collect)
Require: top index = positive integer (starting index of website
Require: connection to MongoDB database; collection created (reports)
Ensure: MongoDB Server Running
for count < target do
Pass the index to web scrape() for downloading the page
if number of table not 3 then
return false
else
return soup
end if
if soup not false then
Pass (soup,index, db connection, db collection) to parse tables()
Require:
pill reports dict - empty python dictionary
for row in table do
col1=row.findnext()
col2 = col1.findnext()
python report dict[col1] = col2
end for
insert python report dict into MongoDB collection reports
end if
end for
be saved into a new MongoDB database named with the current date in the form pillreports DDMonYY.
In []: #Scrape x number of pages starting at the most recent.
target=50
top_index=34120
In []: #Introduce time delay to stop overloading sites.
from time import sleep
while download_count <= target:
sleep (0.25)
soup=web_scrape(top_index)
FIGURE 5.12: Code: Setting scraping parameters.
A final word of caution, if you hit a server with too many requests at once you may
be blocked from accessing it. If that is the case it may be worthwhile encapsulating your
requests within a time delay as shown in Figure 5.12.
Mining Unstructured User Reviews with Python
5.3
111
Data Cleansing
The previous section outlined the scraping of the data from the Web. As part of that
process, regular expression were applied to the pillreports.net landing page, to extract the
relevant features. In effect this was the start of the data cleansing, stripping away the HTML
tags.
This section briefly outlines more aspects of data cleansing namely data tidying and data
enrichment. Cleansing or shaping of the textual fields (Warning: and User Report:) continue
throughout the project and are discussed further in the sections detailing the classification
(5.5) and clustering (5.6) objectives.
Data tidying is relatively straightforward. In this chapter the tasks are defined as
(a) Renaming the columns (5.2)
(b) Cleaning up the date format (5.3)
(c) Remove trailing mm to get integer millimeter (5.7)
(d) Splitting column Report Quality Rating: (5.8)
(e) Categorize Suspected Contents: column (5.9)
(f) Splitting the Submitted By: column (5.10)
(g) Reorder the columns (5.11)
(h) Stripping out white space on binary labels (5.12)
These steps are well outlined in the accompanying iPython Notebook: 3.Data
Cleansing. The numbers in brackets refer to the actual code chunk in the notebook.
Many of the steps use regular expression methods as described in Section 5.2.1.
The data is read from the MongoDB into the ubiquitous pandas data frame. pandas
is a library for data analysis in Python. A data frame is a tabular structure of rows and
columns similar to a database table or a spreadsheet.
In addition to regular expression, a few Pythonic constructs (lambda, apply) and
the datetime library are utilized in the data cleansing.
The lambda construct allows for the creation of functions without specifying a name
for the function or even a return value. They are sometimes referred to as anonymous,
throwaway or mini functions. Figure 5.13 concisely illustrates the concept.
The input to the function in the above case (x) is specified on the left of the colon and
the operation on (x) on the right.
Apply is a method of the pandas library. It applies an operation to a column of a
pandas data frame using the row index. (In the example presented in Figure 5.14 the
operator is an unnamed lambda function, applied to every row.)
The datetime module is utilized to recode columns containing date information. The
datetime module supplies a number of classes for manipulating dates and times. In the
example presented in Figure 5.14 the strptime() method creates a datetime object from
a string.
The string is represented with a series of defined characters ( for example %B is the full
length English word for a month). Once a string has been converted to a datetime object
it is then possible to apply a number of defined operations to work out time differences
(called time deltas), days of the week, etc. It is good practice to set all dates to one unified
112
Text Mining and Visualization: Case Studies Using Open-Source Tools
In [8]: #Defined Function
def square_func (x) :
return x*x
#Using Lambda
square_lamb = lambda x: x*x
print(square_func(5))
print(square_lamb(5))
25
25
FIGURE 5.13: Code: Lambda method.
In [51]: import pandas as pd
data=[u’January 3, 2014, 3:19 pm GMT’,
u’March 17, 2014, 1:02 am GMT’,
u’November 14, 2014, 5:19 pm GMT’]
df = pd.DataFrame(data,columns=["Date"])
In [52]: from datetime import datetime
#use strptime to represent dates
as Timestamp(’2014-11-09 14:05:00’)
df[’Date2’] = df[’Date’].apply(lambda d:
datetime.strptime(d, "%B %d, %Y, %I:%M %p GMT"))
#Standard format
print df[’Date2’][2]
print "The first day is "+(df[’Date2’][0]).strftime("%A")
#Create a time delta
days=(df[’Date2’][2]-df[’Date2’][0]).days
print str(days)+" days between dates."
2014-11-14 17:19:00
The first day is Friday
315 days between dates.
FIGURE 5.14: Code: DateTime library, Apply & Lambda methods.
Mining Unstructured User Reviews with Python
113
format, as it allows for much quicker development later on.
The data enrichment consists of three steps
(a) Recode the State/Province: column to Country (5.4)
(b) Calculate the number of comments per report (5.5)
(c) Determine the language of entries in User Report: using the langdetect library (5.6)
Recoding of the State/Province column was the most laborious part in the cleaning process.
The column is free text and could be filled with a multitude of values. The original column
had 1,629 different entries. The Google Refine application was used to cluster these
based on similarity, resulting in 1,158 unique entries. These were then hand coded by sight
in Excel to their respective countries. This is far from ideal and not recommended for large
datasets, but in this case was acceptable. Table 5.4 illustrates some example cases. Note
the value So-Cal an abbreviated version of South California coded to the United States. A
lot of granularity is lost with this methodology.
TABLE 5.4: Geo-coding State/Province: column.
Geo Coding
Refine Cluster Row Count
MELBOURNE
107
Chicago
93
CALIFORNIA
91
Country
Australia
USA
USA
NSW-Sydney
So-Cal
Australia
USA
60
57
The comments associated with each report were also saved to the MongoDB during
the webscrape. A reference to the corresponding reports was included, the report ID: The
comments collection is extracted from the MongoDB, the number of ID:s summed and this
is merged with the data frame. See step 5.5 in the accompanying notebook.
The final part of data enrichment involves the Google developed langdetect library.
langdetect is a Python port of a Google created Java library for language detection. The
library was developed using machine learning. It was supplied with a volume of Wikipedia
articles from different languages and trained using Naı̈ve Bayes. The library is the resultant
Naı̈ve Bayes ruleset or filter. It can detect 49 languages using the created ruleset with 99%
accuracy, requiring a minimum 10–12 words depending on the language.
The langdetect methods were applied to the User Report: column. This was chosen
over Description: column as it had more empty entries. The machine learning in this chapter
is limited to English only and the labels determined by langdetect are used to filter the
rows. If Description: column is used, it is possible to get a label indicating English but a
blank entry in the User Report: field. The choice of User Report: therefore has an effect on
the number of rows used in the Classification and Clustering processes. This is discussed
further in Section 5.5.
A synopsis of the results from applying langdetect to the data is shown in Table 5.5.
The determined language is shown with the generated Country: column. The Unknown are
predominately empty or very short entries for User Report: field.
114
Text Mining and Visualization: Case Studies Using Open-Source Tools
TABLE 5.5: Summary of Country: and Language: columns.
Country of
Report
USA
Australia
Netherlands
England
Unknown
Ireland
49 other countries
Total
Rows
2051
969
364
324
282
217
794
5001
en
1065
591
208
223
160
122
593
2962
nl
Language
tr pl ru
af
no
29
1
3
32
19
20
1
2
2
2
2
1
1
1
Lang. not
determined
986
378
127
101
119
95
175
1981
(en=English, nl=Dutch, tr=Turkish, pl=Polish, ru=Russian, af=Afrikaans, no=Norwegian)
Table 5.6 displays a summary of the prediction confidence as generated by the
langdetect ruleset.
TABLE 5.6: Summary of language prediction confidence.
Language
en
nl
tr
pl
ru
af
no
Total
Rows
2962
32
20
2
2
1
1
Prediction
min
max
0.571 1
0.714 1
0.571 1
0.857 1
1
1
0.571 0.571
1
1
Confidence
avg.
std.
0.999 0.014
0.92
0.108
0.943 0.126
0.929 0.101
1
0.571
1
(1,981 rows have no language determined, 1,762 rows have no text, remainder typically too short)
5.4
Data Visualization and Exploration
The success of data mining and by extension text mining projects lie in part with the
visualization and exploration of the data.
In Section 5.1.4.1 the power of Python’s implementation of sparse matrices and the tight
integration with the sklearn library was extolled as a driver in choosing Python as the
data mining tool for text mining. A downside to choosing Python is that there currently does
not exist an easy to use and mature library for data visualization, certainly not comparable
to the ease of use of the ggplot2 package [5] in the R statistical programming language.
Up until very recently the Matplotlib library was really the only option for visualization in Python. In the last few years, a number of new libraries have emerged, some built
upon the foundations of Matplotlib.
Some of these (Bokeh, Chaco, Plotly, Vincent) are orientated towards interactive visu-
Mining Unstructured User Reviews with Python
115
alizations either locally or hosted. Others (Seaborn, ggplot) aim to emulate the ease of use
of ggplot2 for R. For a good treatment of their various strengths and weaknesses see [6].
All of the above libraries are in nascent stages of development. Therefore currently the
best option is to use Matplotlib as the base tool and the other tools where appropriate.
For this chapter the Seaborn library was used at different points. It is built on top of
Matplotlib and aims to remove much of the time consuming coding burden of creating nice
plots in Matplotlib, more of which in the next section (5.4.1).
5.4.1
Python:Matplotlib
A quick search of questions on stackoverflow.com tagged with Matplotlib reveals ≈ 10,250
tags. The next nearest of the libraries listed in the previous section is Bokeh ≈ 130 and
Seaborn ≈ 100 (as of 1 Feb, 2015), reflecting Matplotlib’s popularity, perpetuity, and pain!
Matplotlib has a low-level API. The amount of code required to create only moderately
complex statistical visualizations can be onerous, certainly compared to that of ggplot2.
Seaborn is built on-top of Matplotlib and and provides nice wrappers for creating
nicely formatted visualizations without the coding and time overhead. It also works well
with pandas data frames. However, it is relatively new and therefore it does not have the
full suite of functions or controls as are often required.
The best way to learn these libraries is through repeated use. Section 5.4.2 presents the
results of data exploration with a mixture of visualizations and tables. It provides links to
the accompanying notebooks. The focus of the code explanations is on rolling up data to
the point where a visualization with Matplotlib is possible.
Before presenting the results, Figure 5.15 shows a simple Matplotlib subplot containing
a line chart and bar graph. Figure 5.16 presents the associated code with 3 alternate ways of
achieving this plot with Matplotlib. This goes some way to explaining why it can be difficult
to get going with Matplotlib. It is an extensive library with a huge amount of open-source
development and thus has a lot of cross over functionality.
FIGURE 5.15: Simple matplotlib subplot example.
In Matplotlib, there is the concept of the current plot or often referred to as an axes (not
axis). Until a new plot is generated, all commands entered are targeted at the most recently
generated graphic. Alternatively a handle to the individual plot object (ax1,ax2) can be
used to pass commands to a particular graphic. For a good introduction to Matplotlib refer
to [7].
5.4.2
Pillreports.net Data Exploration
This section focuses on describing the downloaded data. The code for the actual plots can
be found in iPython Notebook: 4:Visualization.ipynb. Where code is presented
it is to illustrate how rows are aggregated before passing to the Matplotlib method.
116
Text Mining and Visualization: Case Studies Using Open-Source Tools
In []: %matplotlib inline
import numpy as np
import matplotlib.pyplot as plt
fib = (1, 8, 5, 3, 2,21,13)
x = np.linspace(1,len(fib),num=len(fib))
In []: plt.figure(1,figsize=(12,4))
#1. Using Subplots
plt.subplot(121)
plt.plot(fib),plt.title("Title A")
plt.subplot(122)
plt.bar(x,fib),plt.ylabel("Y Label: B")
In []: #2. Using figure and axes
fig, ((ax1,ax2)) = plt.subplots(1,2,figsize=(12, 4))
ax1.plot(fib),ax1.set_title("Title A")
ax2.bar(x,fib),ax2.set_ylabel("Y Label: B")
In []: #3. Using ravel()
fig, axs = plt.subplots(1,2, figsize=(12, 4))
axs = axs.ravel()
for i in range(2):
if i==0:
axs[i].plot(fib)
if i==1:
axs[i].bar(x,fib)
axs[i].set_title((’Title ’+str(i)))
FIGURE 5.16: Code: matplotlib subplots.
Mining Unstructured User Reviews with Python
117
A total of 5,001 reports were downloaded. The User: and Description: columns
represent the corpus of documents in the usual sense of text mining. The earliest report
submitted report was on 08/08/2011 and the most recent on 17/11/2014. On average there
were 29 reports being submitted per week for this time period. Figure 5.17 displays graphs
of the weekly count of reports submitted both by year and over the whole time period.
FIGURE 5.17 (See color insert.): Weekly count of reports.
There is an obvious decline in number of reports being submitted over the time period.
In 21 weeks’ worth of data from 2011 there averaged 41 reports per week. This drops to
20 reports by 2014. The causes of this are not known. Interestingly the number of reports
with the value Warning=No increases over this time period from 50% to 70% as shown in
Figure 5.18.
(a) Split by Warning:
(b) Count by Country: column
FIGURE 5.18: Warning: Column cross-tabulated with Country: column.
118
Text Mining and Visualization: Case Studies Using Open-Source Tools
It could be that there is an increased usage of testing kits to determine the presence
of adulterants in ecstasy pills over this period of time. The breakdown of the suspected
contents of a pill is shown in Table 5.7.
TABLE 5.7: Suspected contents (SC Category:) and Warning: label.
Suspected Contents
Category
amphet (Amphetamines)
MDxx
Other
pip (Piperzines)
Unknown
Warning:
yes percent
yes
107
397 79
2809 50
2
34
287 89
3
276 99
231
807 78
no
Table 5.7 is used as an illustrative example for the data aggregation or roll up that is
required to build much of the visualizations presented in this section. Many visualizations
rely on comparison counts of categorical variables and/or the comparison of quantitative
variables split by categorical variables. Three alternative methods to achieve the results in
the Table 5.7 are presented in the code snippet shown in Figure 5.19.
In [83]: import pandas as pd
import numpy as np
df=pd.read_csv("Data/prReports.csv",encoding=’utf-8’)
In []: #Pivot Table
df.pivot_table(index=’SC Category:’, columns=’Warning:’,values="ID:",
aggfunc=len)
#Cross Tabulation
pd.crosstab(df[’SC Category:’],df[’Warning:’],margins=False)
bySCCategory=df.groupby([’SC Category:’])
bySCCategory[’Warning:’].value_counts().unstack()
In []: #len=number of rows, mean, max of UR Length, by w
bySCCategory[’UR Length:’].agg([len,np.mean,np.max])
FIGURE 5.19: Code: Weekly count of reports submitted.
The pivot table is relatively straight forward to understand if you have any familiarity with the concept from spreadsheets. The categorical variables for rows and columns
are specified, then the value that is under observation and the summation measures are
provided. In this case the ID: column is passed and its len or number of rows counted.
Passing the ID: is bit of a workaround in the above example. When counts of rows are
required pandas provides a much simpler crosstab function to achieved this.
The final method shown is groupby which is arguably the most powerful and flexible of
the above methods. Using groupby allows for multiple variables and multiple aggregation
functions to be computed simultaneously as required.
Much of the rolling up processes presented here are hard coded into the new visual-
Mining Unstructured User Reviews with Python
119
ization libraries discussed. It is beneficial to expose these calculations, as it can aid better
understanding if you have an appreciation of what is going on under the hood. [8] provides
a comprehensive overview of aggregating data using the pandas library. The focus of the
data mining objectives set out, is to build supervised and unsupervised models based on
the textual data using the Description: and User Report: columns. An analysis of these
columns is required. String lengths is a useful way of examining textual data to understand
if key differences exist for certain groupings.
Table 5.8 displays the string lengths of User Report: grouped by Country: and Warning:
columns. Only the top 7 countries are displayed.
TABLE 5.8: User Report: string length grouped by Country: and Warning:.
Country
USA
Australia
Netherlands
England
Unknown
Canada
Ireland
Scotland
User Report: String Length
All
Warning=Yes
Warning=No
Rows Rows Mean std. Rows Mean
Len.
Len.
1131
505
1297
1261 626
1481
631
186
929
745
445
973
258
21
719
642
237
849
229
47
872
579
182
1079
184
66
836
892
118
1062
163
148
810
1016 15
1357
126
34
885
894
92
977
91
18
1178
1183 73
1158
std.
1228
738
944
875
960
805
717
804
There are differences in length based on the country of origin, with the USA, Scotland
and Australia having longer responses. When looking at the mean of the string length split
by Warning: value, there is slight increase in the length where Warning:=no. It should be
noted that the string length of Description: entries have a log normal distribution. This has
long been observed in language composition [9].
A visualization of the string lengths transformed by the natural logarithm (loge x) is
shown in 5.4.2. In addition a boxplot and violin plot are presented using the plotting
functions from the Seaborn library (see step 4.9 of the accompanying notebook).
The violin plot shows the distribution for Warning:=Yes has a less dense mean peak.
This may indicate that people give either sharper/shorter responses or longer/detailed responses in this instance.
The next section presents an outline of taking the textual data in Description: column
and manipulating it to a format where its features (the words) can be used as the input
variables of a classification model.
5.5
Classification
The Warning: field is populated from a drop-down list with a binary option Y es/No, indicating whether a pill being considered is believed to be adulterated i.e. contains ingredients
not comparable to 3,4-methylenedioxy-methamphetamine, commonly referred to as MDMA,
the key active ingredient of ecstasy.
120
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 5.20: String length of Description: column.
This provides a good target label to build a prediction on. The Description: field was
chosen as the text field to generate input features for the prediction model. The Description:
free text field was chosen over User Report: as it is mandatory and the latter optional. In
addition it was believed that a label Warning:=Yes may preclude an entry into User Report:
field. (Would a person consume a pill if they believed it to be adulterated?)
The distribution between the labels is No (N=3,184, 64%) and Yes (N=1,817, 36%).
There is no significant skew in the label distributions, therefore the measure of Accuracy is
suitable for reporting model results.
TP + TN
(5.1)
TP + FP + TN + FN
where T P is true positive, T N is true negative,F P is false positive and F N is false negative.
Accuracy =
The next section moves through the process of setting up a classification model with
the sklearn library. The full code is presented in accompanying iPython Notebook:
5.Classification.
5.5.1
Classification with sklearn
The sklearn package allows for the development of classification models in straightforward and intuitive steps. This allows for manageable creation of very complex models.
The first step after reading in the data is to set up the methods for vectorizing the text
data into a sparse matrix as outlined in Section 5.1.4.1. A basic outline of this is shown in
Figure 5.21
The CountVectorizer tokenizes text into individual words, creating the bag of words
Mining Unstructured User Reviews with Python
121
In [1]: import pandas as pd
df=pd.read_csv("Data/prReports.csv",encoding=’utf-8’)
In [2]: #Set up data transformation and model parameters
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
vec = CountVectorizer(min_df=0.001,max_df=0.95)
from sklearn.naive_bayes import MultinomialNB
clf = MultinomialNB()
FIGURE 5.21: Code: Setting up vectorizer and models for classification.
representation. In the presented example, the default parameters are used, that is uni-ngram
and frequency count. It is possible to change the parameters for multi-ngrams and/or binary
occurrence. The min df and the max df parameters are set to (0.001 , 0.95) respectively.
min df determines how infrequent words are handled. In this example any word that
appears in fewer than 0.1% of documents (or five documents with 5,000 documents) should
be discarded. Similarly max df determines how very frequent words should be treated.
Frequent words may, but not be, limited to english stopwords. In this domain it is conceivable
that the word drug may be frequent.
The TfidfTransformer can be used to create a term-frequency inverse document
frequency or tf-idf representation of the corpus. The calculation of the tf-idf for a term t is:
Count of t in document
Total number of terms in document
Count of documents
IDF (t) = loge
Count of documents containing t
T F (t) =
tf-idf(t) = T F (t) ∗ IDF (t)
(5.2)
(5.3)
(5.4)
The effect of tf-idf is to increase the importance of a word according to the number of
times it occurs in a document and offset it by the frequency with which it occurs across all
the documents, in effect filtering for domain specific common words.
In [7]: #Split the data. Recode text label as numeric (TRUE/FALSE)
X=df[’Description:’].values
y=(df[’Warning:’]==’no’).values.astype(int)
from sklearn import cross_validation
X_train, X_test, y_train, y_test = cross_validation.train_test_split(
X,y,test_size=0.4,random_state=79)
FIGURE 5.22: Code: Splitting into train and test sets.
Figure 5.22 illustrates the cross validation.train test split method to split
the data into training and test sets. Note the use of the random state. This controls
how the random number generator of the splitting method is created, so as to allow the
comparison of various models. If this was not held constant a different subset of rows
would be selected each time the method was called. This could potentially obscure any
122
Text Mining and Visualization: Case Studies Using Open-Source Tools
improvements or degradation of the score of accuracy when altering model parameters.
This would be typically removed once the model is finalized.
Figure 5.23 introduces the sklearn Pipeline class. Pipeline is a very powerful feature of the sklearn library, allowing the chaining or serialization of data transformations
before applying the data to the model. In essence it builds a data stream. A lot of the
heavy lifting in machine learning are the transformations and adjusting of parameters for
generating the input feature set. Using pipeline facilitates this especially when passing
user defined function or wrapping it in a loop and altering the parameters (see step 5.4 of
the accompanying notebook).
In []: #Set up a pipeline, transformations and
#model will be serialized.
from sklearn import pipeline
pipe = pipeline.Pipeline([(’vectorizer’, vec),
(’tfidf’,TfidfTransformer()),
(’classifier’, clf)])
pipe.fit(X_train, y_train)
FIGURE 5.23: Code: sklearn pipeline.
The final part in classification is reporting the model results. Figure 5.24 presents the
metrics class with a number of different methods for outputting relevant results.
In []: #Use metrics methods to print get results.
from sklearn import metrics
print (metrics.accuracy_score(pipe.predict(X_test), y_test))
print (metrics.classification_report(pipe.predict(X_test), y_test))
print (metrics.confusion_matrix(pipe.predict(X_test), y_test))
FIGURE 5.24: Code: Model metrics.
5.5.2
Classification Results
In the code for this chapter, a series of vector representations of the text in the Description:
column, with different settings were created. Two classifiers were used: Naı̈ve Bayes and
Stochastic Gradient Descent. A sklearn pipeline was wrapped inside a loop and the
vector representation and model were interchanged. This resulted in a matrix of (2 models
× 6 vector representations × 2 term counting schemes ) 24 combinations. The top 5 results
based on training accuracy are presented in Table 5.9.
From a baseline accuracy of 0.710 in the first model (as outlined in Figure 5.22) the accuracy only increases slightly to 0.72 using Naı̈ve Bayes and relatively straightforward vector
representation with words containing alpha character only. Stochastic Gradient Boosting
with tf-idf weighting, stemming and 1-3 ngrams achieves only a modest improvement.
The number of features is quite large, 2,785 and 16,847, respectively. [10] have cited that
a significantly reduced feature set can improve accuracy in text classification. The next step
therefore is to reduce the feature set.
Taking the best results for both the model types (Naı̈ve Bayes and Stochastic Gradient Descent) feature selection is applied. sklearn provides simple methods for feature
selection, presented in Figure 5.25.
Mining Unstructured User Reviews with Python
123
TABLE 5.9: Top 5 models: binary classification on Warning: column.
Rank
1
2
3
4
5
Model
NV
SGD
NB
SGD
SGD
Number
Features
2,785
1,6,847
2,905
1,1,819
167,136
Train
Acc.
0.722
0.722
0.72
0.715
0.711
Train
std.
0.014
0.024
0.016
0.024
0.021
Test
Score
0.726
0.705
0.733
0.7
0.706
Count
freq
tfidf
freq
tfidf
tfidf
Vectorization
Ngrams Stemmed
1 gram
1-3 grams
1-3 grams
1-2 grams
1-3 grams
X
Minimum and maximum document frequency held at constant percentages (0.001, 0.95), respectively.
In []: from sklearn.feature_selection import SelectKBest, chi2
for sel in [150,300,450,600]:
#Feature selection based on select k number of features
fs = SelectKBest(chi2, k=sel)
X_train_reduced = fs.fit_transform(X_train, y_train)
X_test_reduced = fs.transform(X_test)
pipe.fit(X_train_reduced, y_train)
FIGURE 5.25: Code: Feature selection.
124
Text Mining and Visualization: Case Studies Using Open-Source Tools
Currently the implementation of sklearn.feature selection is limited to the chisquared measure. This allows selection of features using the SelectPercentile where
the top % predictive words are selected based on the chi-square measure. It is possible to
write a user defined function to determine other measures such as information gain, etc.
and pass these into the feature selection method. The results from applying feature
selection are shown in Table 5.10.
TABLE 5.10: Classification accuracy using feature selection.
Model
Naı̈ve Bayes; word frequency
count, alpha characters only, 1
gram (2,785 features)
SGD; tf-idf weighting, alpha
characters only, 1-3 grams
(16,487 features)
Select
Feat.
417
557
450
600
300
450
%
Feat.
0.15
0.2
0.162
0.036
0.018
0.027
Train
Acc.
0.779
0.793
0.782
0.779
0.755
0.765
Train
std.
0.021
0.022
0.022
0.019
0.014
0.015
Test
Acc.
0.722
0.722
0.722
0.723
0.723
0.712
While feature selection improves the training accuracy it appears to overfit the data
as there is only a modest increase, if any, in the test score. Not withstanding this, feature
selection provides an interesting way to examine the predictive power of individual words.
Using many of the code methods outlined above and combining with the process for generating plots outlined in Section 5.4.1, a scatter plot of the top 250 predictive uni-gram words
was created based on the chi-square measure. The code for this is contained in step 5.6 of
iPython Notebook: 5.Classification.
The coding process is very similar to classification, with one key difference. Rather than
counting the words using term frequency (TF) or TF-IDF, the binary occurrence is used.
Binary occurrence gives a term a value of one it occurs in a document, regardless of the
number of times its occurs in that document, zero otherwise. A zoomed in version of the
plot is shown in Figure 5.26.
The effect of the using binary occurrence is the plot shows the count of documents that
a particular term occurs in across the whole corpus. The count is split by the class label
Warning:=yes/no of each document. They size of the bubbles represents the chi-square
value of that term.
It is interesting to note that words like (sour, stay, away, warning, speed) are associated with Warning:=yes. ”Stay away” probably has more predictive power as a bi-gram.
Contrast this with the words for Warning:=no such as amazing, instantly, tolerance and
words associated with testing kits mecke, mandeline, reagent. There is a significantly lower
probability of a Warning=Yes if the pill has been tested, as illustrated in Table 5.7. The
examination of the individual words predictive power, motivates examination of the User
Report: entries detailed in the next section.
5.6
Clustering & PCA
The previous section outlined a supervised learning exercise. This section explores unsupervised learning using two methods Clustering and Principal Component Analysis (PCA).
Mining Unstructured User Reviews with Python
125
FIGURE 5.26: Scatter plot of top predictive features.
126
Text Mining and Visualization: Case Studies Using Open-Source Tools
These are applied to the User Report: column. Performing unsupervised learning on textual
data may be considered as data exploration [11] depending on the context of the project.
It is assumed that the user is familiar with Clustering. PCA may be less familiar. In
structured data, one of the challenges is understanding the interaction between multiple
variables. Many times, pairwise calculations of correlation between two variables are used.
Sometimes there is a requirement to reduce the number of variables, while retaining as
much as the information as possible as possible. PCA is one such data reduction technique
to transform a large number of variables to a smaller set. It is beyond the scope of this
chapter to sufficiently explain PCA. For a good treatment see [12].
The processes for setting up vector representation of the User Report: is similar to that
detailed in Section 5.5.1. Note that binary occurrence is used and Tf-idf weighting is used
to control for documents of varying lengths. The full code for this section is contained in
the accompanying iPython Notebook: 6. Clustering & PCA.
Setting up a cluster model is exactly the same as outlined for classification models in
Figure 5.22. To run PCA, the sparse matrix that is the output of the vectorization of
the words, must be first converted to a dense matrix as the PCA method of sklearn
currently does not handle sparse matrices. Example code is presented in Figure 5.27
In []: from sklearn import cluster
from sklearn.decomposition import PCA
#Clustering
k = 3, kmeans = cluster.KMeans(n_clusters=k)
kmeans.fit(data)
#PCA convert the sparse matrix to dense() representations
pca = PCA(n_components=2)
UR_PCA = pca.fit(data.todense()).transform(data.todense())
FIGURE 5.27: Code: Clustering and PCA models.
After PCA has been applied the first two principal components are added to the data
frame. These are a series of numeric values and thus can be visualized on a scatter plot. A
exploratory question arises, do these principal components represent a particular phenomena
of the data set? An attempt at interpreting the meaning of the principal components is
provided using available categorical variables.
The components are plotted on a scatter plot using matplotlib. In addition the points
are colored based on a particular nominal field, including the membership label as a result
of the clustering (three clusters). Figure 5.28 presents the output. Here a matrix of six plots
is created, each plot’s point being colored by a different categorical variable.
Moving through the graph, plot (1) is colored by the cluster membership created with
k-means. Compare it to plot (3). This represents the documents colored by the value of the
Warning: column.
There is a definite split in plot (3) and this seems to be replicated to some extent in the
cluster membership displayed in plot (1). Therefore, the second principal component (along
the y axis) could be said to be exposing the variation in words used based on the value of
Warning:=yes/no and the k-means clustering algorithm finds this also.
Plot (2) is colored by creating a binning of the User Description: field based on its string
length, with equal numbers of points per bin (low, medium, high - string lengths). The split
here also shows some relation to the clustering of the documents in plot (1). Moreover it
could be said the first principal component (along the x axis) is capturing the length of the
documents.
Mining Unstructured User Reviews with Python
FIGURE 5.28 (See color insert.): Principal components scatter plot.
127
128
Text Mining and Visualization: Case Studies Using Open-Source Tools
Plot (4) represents the points colored by the SC Category: column. Note the points where
SC Category: = mdxx appear to follow the same pattern as the points where Warning: =
no, show in plot (3). This is not surprising as Table 5.7 illustrated significant differences in
distribution of the Warning: values based on the suspected contents.
Plot (5) is colored by the country of origin of the report. There appears to be the outline
of a pattern here, perhaps indicating differing english vocabularies based on the origin of
the report. The final plot (6) is based on the day of the week a report was posted. It was
hypothesized that reports filled in or around the weekend would vary somewhat. This is not
supported here.
5.6.1
Word Cloud
Taking the observations reviewed already, it might be interesting to visualize the differences in language use of the User Experience: field based on the Warning: label. The SC
Category: label appears to mimic the Warning: value and could be used as a proxy. As it
is multinominal it can provide more granularity than a binary variable.
A wordcloud of the User Experience: field split the value of SC Category: provides a very
accessible way of visualizing this. Using all the words can overwhelm such a visualization
type. It is often better to focus on descriptive or rich words. To achieve this the words can
be filtered by their type (e.g., noun, adjective, adverb).
The first step is to tag each word in the column with their relevant tag using the nltk
library. The words can then be filtered based on their type. The code for this is shown in
Figure 5.29.
In []: #Using nltk tag all words with their part of speech.
#Returns tuples (word,pos).
from nltk import pos_tag
clusterDF[’Tagged’]=(clusterDF[’User Report’]).map(
lambda x: pos_tag(word_tokenize(x)))
#Select only those words that are common adjectives and common nouns
clusterDF[’Rich_words’]=(clusterDF[’Tagged’]).map(
lambda x: ’ ’.join([y[0].lower() for y in x
if y[1] in [’JJ’,’NN’]]))
FIGURE 5.29: Code: Tagging words using nltk library.
Wordclouds are based on frequency counts of words. The word with the highest count
receives the largest relevant plotting space (or font size). Sometimes if stopwords are left in
or a particular word has very high frequency it dominates the the wordcloud. In order to
control for this, the top words can be filtered out.
The collections module provides an extremely fast counter for achieving this. In
the example shown in Figure 5.30, each row of the data frame is joined to form one very
large string. This is then tokenized. The defaultdict loops through and adds any unseen
words as a key and for seen words increments a counter or value.
The top words can then be viewed and added to a list of stopwords as necessary. The
final step involves using the wordcloud library developed by [13]. Using a set of controls
within a for loop, allowing for the increment of the number of stopwords, a matrix of plots
is created, shown in Figure 5.31.
This yields some very interesting differences between the language used in the text
column User Report: based on the associated value in Suspected Contents:. Users who take
MDMA substances focus more of adjectives/nouns associated with empathy. Contrast that
Mining Unstructured User Reviews with Python
In []: #Join all words into a long string
words = ’ ’.join(clusterDF[’Rich_words’]
from collections import defaultdict
word_counts = defaultdict(int)
#Tokenize each word loop over and created count
for w in word_tokenize(words):
word_counts[w] += 1
FIGURE 5.30: Code: Counting word frequency with the collections module.
FIGURE 5.31: User report: Word cloud.
129
130
Text Mining and Visualization: Case Studies Using Open-Source Tools
with users who take Amphetamine type substances where the focus is on adjectives/nouns
related to trajectory or energy. The last column are those where the user does not know
what the substance being consumed is. These are dominated by the words associated with
the physical body.
5.7
Summary
This chapter outlined the processes for working through a text mining project using Python.
The binary classification model based on the Description: field achieved a modest 1.9 point
increase from the baseline model (.714) to (.733).
The application of clustering was validated visually using principal components projection and available labels. While the clustering was not entirely successful, the subsequent
scatter plot visualizations led to a new avenue of investigation. This is often the case in
data mining.
Numerical validation of clusters was not completed. There are a wide variety of numerical
measures to validate clusters; see [14] for a comprehensive list. However cluster validation is
currently an area of open research. Methods for selecting the correct number of clusters is
more a matter of usability and interpretation rather than a defined measurement. For more
information about the mathematical underpinnings of clustering see [15].
(Clustering in the context presented in this chapter may not be the best data mining
technique to expose underlying patterns. Topic modeling may be a more suitable approach.)
The word cloud, while relatively elementary in their creation, was powerful and could
certainly be a convincing visualization to aid understanding of Ecstasy drug culture. For
the reader that wishes to do more experimentation using the aforementioned data sets there
a wealth of possible questions that could be posed.
The classification section focused on modeling the Description: free text column. The
Description: column contains significantly less information than the User Report: column.
These could be swapped and or combined and differences in the model accuracy appraised.
The classification model focused on the binary label Warning:. It is possible to use a
different column for multinominal classification (SC Category:) or indeed use a real number
for a regression model with either of the (Report Stars:, Report Votes:, Total Comments:)
columns.
The collection of associated comments for each report is completely unexplored (it is
provided unprocessed in a CSV file). At a simple exploration level, which reports generate
the most comments? Which users post comments regularly? Can comments be clustered?
What are the characteristics of a comment discussion that generate an Admin/Moderator
user intervention? These are just a few possible ideas.
From a programming perspective there are lots of avenues of further investigation. Many
of the processes shown, especially in the classification section could be parallelised to speed
up the computation time. The pipelining method outlined, is only briefly utilized. It
can be extended to build very complex chains of data transformations.
One the best resources for learning about Python and data/text mining is through
iPython notebooks and PyCon conferences that occur regular. Many of these are recorded
and posted on YouTube with links to the accompanying notebooks.
Python for text mining can be very rewarding. There is a very vibrant community with
lots of people who are willing to share their knowledge and experience. All the libraries
detailed here are continually being developed and improved. Finally by using the iPython
Mining Unstructured User Reviews with Python
131
environment it is very easy for you to become part of this community and share your
knowledge and insights in your domain.
Bibliography
[1] P-N Tan, M. Steinbach, V. Kumar, et al. Introduction to data mining, volume 1.
Pearson Addison Wesley, 2006.
[2] K. P. Murphy. Machine learning: a probabilistic perspective. MIT press, 2012.
[3] N. Batchelder. How Do I Stop the Pain.
sgHbC6udIq), 2012.
(https://www.youtube.com/watch?v=
[4] J. Goyvaerts and S. Levithan. Regular expressions cookbook. O’reilly, 2009.
[5] H. Wickham. ggplot2: elegant graphics for data analysis. Springer, 2009.
[6] C. Moffitt.
Overview of Python Visualization Tools.
visualization-tools-1.html), 2015.
(http://pbpython.com/
[7] A. Devert. matplotlib Plotting Cookbook. Packt Publishing Ltd, 2014.
[8] W. McKinney. Python for data analysis: Data wrangling with Pandas, NumPy, and
IPython. O’Reilly Media, Inc., 2012.
[9] W. C. Wake. Sentence-length distributions of Greek authors. Journal of the Royal
Statistical Society. Series A (General), pages 331–346, 1957.
[10] Y. Liu, H. Tong Loh, and A. Sun. Imbalanced text classification: A term weighting
approach. Expert Systems with Applications, 36(1):690–701, 2009.
[11] S. M Weiss, N. Indurkhya, and T. Zhang. Fundamentals of predictive text mining.
Springer Science & Business Media, 2010.
[12] I. Jolliffe. Principal component analysis. Wiley Online Library, 2002.
[13] A. Mueller. Python word cloud. (https://github.com/amueller/word cloud), 12 2014.
[14] M. Charrad, N. Ghazzali, V. Boiteau, and A. Niknafs. Nbclust: An r package for determining the relevant number of clusters in a data set. Journal of Statistical Software,
61(6):1–36, 2014.
[15] R. Xu and D. Wunsch. Clustering, volume 10. John Wiley & Sons, 2008.
This page intentionally left blank
Chapter 6
Sentiment Classification and Visualization of
Product Review Data
Alexander Piazza
Institute of Information Systems, Friedrich-Alexander-University Erlangen-Nuremberg,
Germany
Pavlina Davcheva
Institute of Information Systems, Friedrich-Alexander-University Erlangen-Nuremberg,
Germany
6.1
6.2
6.3
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Used Libraries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Set Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.1
Step 1: Preparation and Loading of Data Set . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.2
Step 2: Preprocessing and Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . .
Substitution of Emoticons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Tokenization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopwords and Punctuation Marks Removal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Text-normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Feature Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.3
Step 3: Sentiment Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.4
Step 4: Evaluation of the Trained Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . .
6.2.5
Step 5: Visualization of Review Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
133
133
135
135
136
137
138
139
139
140
140
141
142
143
145
150
152
Abstract
The chapter presents a process for sentiment classification of product review data. Our
data set consists of reviews on six categories of products from Amazon.com. We use the
star review system to determine whether the review is positive or negative. Emoticons
substitution, tokenization, stopwords removal, and text normalization are applied, features
are generated, which then are used for training the classifier. The resulting classifier is
evaluated based on a k-fold stratified cross-validation strategy using the accuracy and a
confusion matrix as measure for determining the quality of the prediction. As a final step,
we demonstrate two visualization techniques to reveal the context behind the sentiment.
133
134
6.1
Text Mining and Visualization: Case Studies Using Open-Source Tools
Introduction
As human beings, it is only natural for us to express our opinion on different topics,
products, companies, and anything else that is part of our daily lives. The task of sentiment
analysis, also known as opinion mining, is detecting the polarity of these opinions. The
term opinion in itself has many definitions in the literature. However, the focus of sentiment
analysis are mainly the opinions, which express or imply positive or negative statements [1].
Sentiment analysis has gained an increasing importance in the last years, especially due
to the rising of social media. Sharing our opinions on products and services, and companies
mining these opinions has become the rule rather than the exception. This is a reason why
so many advancements in the field of natural language processing are being made.
The content that we as users generate on a daily basis (also known as user-generated
content, or UGC) is of great value for companies because of several reasons: they have an
easy access to what their customers think about their products and services, what they like
or dislike, as well as suggestions for improvement. Companies can also respond in real time
to the demands and complaints of the customers.
In spite of such major advantages, the UGC poses its challenges for the data analyst —
data volume is increasing rapidly, and since it is mostly text and thus, unstructured data, it
is increasingly difficult for a company to process such data. Other issues are related to the
way users tend to express themselves by using mostly colloquial language and abbreviations.
For a machine to understand language in a way human does, it is necessary to have large
amounts of training data. Even in such case, however, the machine cannot detect the fine
line between sarcasm and literal meaning.
There are quite a few tools for analysing text data, including sentiment classification.
The advanced commercial tools (Sysomos, Radian6) that major companies are using, do a
good job in extracting data from the web based on user-defined queries, as well as diverse
visualization techniques for text data (e.g., buzz graphs, word clouds) so that the user
can immediately get a glimpse of the main topics in large amounts of articles. Nevertheless,
even these expensive tools cannot provide a high accuracy of the sentiment. On articles from
newspapers and magazines, the sentiment tends to be of the highest accuracy due to the
precise and correct language. From our practical experience, even in such cases the accuracy
is around 80 to 85%, depending on the language (e.g., English, German). When social media
content like tweets, forum data and other user reviews are considered, the sentiment is often
of a very low accuracy. For this reason, some companies for which analysis of UGC is of high
value develop their own in-house solutions, which are combining the strengths of commercial
and open-source tools for data analysis, such as R and Python.
In this chapter, we present a method for sentiment classification of product reviews from
amazon.com. The goal of this chapter is less to provide an optimized process for predicting
the sentiment of the given data set, but more to provide a first framework which enables
the reader to conduct his or her own experiments and optimizations within the code. The
reviews are not manually labelled as positive or negative. We chose such data intentionally
since in a real-world scenario it would be a very time-consuming task to do so. We present
a different option by using the 5-star user ratings system (also used by [2]). We also present
visualization techniques that assist in revealing the context behind the sentiment. This
is important because a company would like to know not only the sentiment but also the
product features to which the sentiment refers. For example, the information that there is
an 80% negative statement on a query related to a newly released product, is valuable only
to a very low extent. It is very important to detect the exact reason for the negative reaction
of the customers towards the product (e.g., shape or text and image labels are often such
Sentiment Classification and Visualization of Product Review Data
135
reasons, with the latest example of retailer Zara removing a T-shirt due to its resemblance
to an attire worn by Jewish people during the Holocaust).1
In the rest of the chapter we present and explain the process for sentiment classification
and visualization following the steps presented in Figure 6.1.
FIGURE 6.1: Sentiment classification and visualization process.
We start with an overview of the libraries we used for the classification and description of
the data set. In step 1, we explain how to load and parse the data set. In step 2, we explain
the preprocessing and extraction of features required for training the classifier. In step 3,
the classification of the sentiment is demonstrated followed by evaluation of the sentiment
accuracy in step 4. Finally, in step 5 we show some options for visualizing the data with
the open-source tools Wordle and Jigsaw. The chapter concludes with a brief summary.
Used Libraries
Table 6.1 gives an overview of the libraries we used for the sentiment classification part
of this chapter along with the link where the libraries can be downloaded. The algorithms
demonstrated were implemented and tested on a computer running on Windows 8. Nevertheless the implementation should be platform agnostic. Because of stability issues it is
recommended to use only 32-bit versions of Python and the used modules.
Data Set Description
For this project, we chose a data set that consists of reviews for six product categories:
camera, mobile phone, TV, laptop, tablet, and video surveillance system. [3] collected the
data set from www.amazon.com.2 The reviews are organized in separate folders according
to the product category and are in JSON (JavaScript Object Notation) format. JSON is
an open data interchange format that stores data in attribute-value pairs. Table 6.2 shows
the number of reviews in each category and the number of positive and negative reviews,
excluding multiple postings of the same review. The total number of reviews is 20,821, which
is sufficient for meaningful analysis. Moreover, for our purpose, it is useful to have different
types of products reviewed by customers based on diverse characteristics, mainly for the
visualization part of this chapter. Since some of the reviews do not have text content (i.e.,
1 For more information, see: http://www.ibtimes.co.uk/zara-removes-childs-holocaust-pyjamas-salefollowing-complaints-1462767
2 The data set can be downloaded from: http://sifaka.cs.uiuc.edu/ wang296/Data/.
136
Text Mining and Visualization: Case Studies Using Open-Source Tools
Software
Python 3.4.1
for Windows
(32 Bit)
NumPy 1.8.1
Superpack
win32
Description
Python installer for Windows
Link
https://www.python.org/downloads/
NumPy is a general-purpose
array-processing package designed to efficiently manipulate
large multi-dimensional arrays of
arbitrary records (python.org).
NLTK
A Python package for natural
3.0.0b1.win32 language processing
Scipy 0.14.0- A Python package containing alwin32gorithms for numeric and sciensuperpacktific computing
python 3.2
https://pypi.python.org/pypi/numpy/
https://pypi.python.org/pypi/nltk/
https://pypi.python.org/pypi/scipy/
TABLE 6.1: Table of used libraries.
they contain only a star review) we consider just reviews with text content. Sometimes the
same review is posted multiple times. In such case, we kept only one entry of the review.
Product Category
Number
views
Video Surveillance
TVs
Tablets
Mobile Phone
Laptops
Cameras
Total:
2,790
2,365
1,049
4,471
2,473
7,673
20,821
of
re-
Number
of
unique positive
comments (4 &
5 stars)
672
1,507
523
2,707
1,615
4,041
11,065
Number
of
unique negative
comments (1 &
2 stars)
223
180
120
678
269
695
2,165
TABLE 6.2: Number of reviews per product category.
Each product and review have certain attributes as follows:
• Product attributes: product ID (unique), name, product features, listing price, and
image URL.
• Review attributes: review ID (unique), author, title, content, overall rating, and date.
For sentiment analysis, we need only the review attributes content and overall rating.
6.2
Process
In the following, a program code for conducting sentiment analysis with Python and
NLTK is introduced. In each subsection, parts of the code are explained. For reproducing
Sentiment Classification and Visualization of Product Review Data
137
the example conducted in this chapter, the following source fragments should be stored in
one file like sentiment analysis.py.
In the beginning of the source file, the necessary modules shown in Listing 6.1 have to
be imported. The json and the glob modules are built-in modules in Python, the first for
accessing data in the JSON format, the latter for navigating and accessing the file system
with Unix-style pattern (refer to lines 1 – 2). From the NLTK library we load modules
for preprocessing and feature creation (lines 4 – 9), which are used in the preprocessing in
Section 6.2.2 step 2 and modules for the classification (lines 11 – 12) in Section 6.2.3, step 3,
which are described more in detail in the following pages. The last two imports are related
to the classifier evaluation in step 4, where we use the arrays from the numpy module for
the data preparation, and the cross validation function from the scikit-learn module for
conducting the evaluation.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
import j s o n
import g l o b
from
from
from
from
from
from
nltk
nltk
nltk
nltk
nltk
nltk
. t o k e n i z e import w o r d t o k e n i z e
. c o r p u s import s t o p w o r d s
. stem . s n o w b a l l import SnowballStemmer
. c o l l o c a t i o n s import B i g r a m C o l l o c a t i o n F i n d e r
. m e t r i c s import BigramAssocMeasures
. u t i l import bi gr am s
import n l t k . c l a s s i f y . u t i l
from n l t k . c l a s s i f y import N a i v e B a y e s C l a s s i f i e r
import numpy a s np
from s k l e a r n import c r o s s v a l i d a t i o n
Listing 6.1: Imported modules.
6.2.1
Step 1: Preparation and Loading of Data Set
In the first step we describe how we read the data from the JSON files, parse these data
and filter out flawed review entries.
We have bundled all necessary steps for loading and filtering the review data from the
JSON files in the function readJSONdata(), which is shown in Listing 6.2. The argument
opinionData of the function indicates the path to the directory containing the JSON files
(e.g., “c:\projects\AmazonReviews\mobilephone\*”) and returns two lists of strings having
all positive and negative reviews (pos reviews or neg reviews).
Within the function, first all files in a directory are read within the opinionData
folder by using the glob module, which offers file system access possibilities similar to
unix-environments. This means Unix shell wildcards like *.zip and relative pathnames like
“..\TVs\*” can be used. First, we read the names of all JSON files in the directory opinionData (line 2), and then read iteratively all JSON files (line 6) into json data. In lines 9-10
we check if the JSON file has a “content” and “overall” entry. The “content” entry holds the
text of the review and the “overall” entry the star rating. If the review entry has an overall
rating of 4 or 5 stars it is classified as positive, and if it has 1 or 2 stars as negative. The
review is added to the pos reviews or neg reviews list if it contains a review text (line 13 or
18, third condition). Since reviews with 3 stars are considered as being neither positive nor
negative, they are omitted.
For the later sentiment classification, it is important to consider only comments consisting of natural text and to avoid text not belonging to the review like code fragments,
which could appear during an imperfect Web crawling. In our Amazon data set there are
138
Text Mining and Visualization: Case Studies Using Open-Source Tools
139 reviews containing only code or markup fragments. These are filtered out by the fourth
condition in lines 12 and 17 by not considering review texts starting with “<!--”, which
indicates a comment in HTML.
As a last step we want to ensure that each comment is loaded only once for the later
analysis to avoid noise in the classification. Therefore, we filter out multiple occurrences of
the same comments in lines 14 – 15 for positive reviews and lines 19 – 20 for negative ones.
1 def readJSONdata ( o p i n i o n D a t a ) :
2
a l l f i l e s = glob . glob ( opinionData )
3
pos reviews = [ ]
4
neg reviews = [ ]
5
f o r elem in a l l f i l e s :
6
with open ( elem ) a s j s o n f i l e :
7
json data = json . load ( j s o n f i l e )
8
i f j s o n d a t a . g e t ( " Reviews " ) :
9
i f " Content " in j s o n d a t a [ " Reviews " ] [ 0 ] . k e y s ( ) :
10
i f " O v e r a l l " in j s o n d a t a [ " Reviews " ] [ 0 ] . k e y s ( ) :
11
i f ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 5 . 0 " or j s o n d a t a [
" Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 4 . 0 " ) and \
12
j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] and not s t r ( j s o n d a t a [ "
Reviews " ] [ 0 ] [ " Content " ] ) . s t a r t s w i t h ( "<!−−" ) :
13
tmp = s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] )
14
i f not tmp in p o s r e v i e w s :
15
p o s r e v i e w s . append ( tmp )
16
i f ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 1 . 0 " or j s o n d a t a [
" Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 2 . 0 " ) and \
17
j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] and not s t r ( j s o n d a t a [ "
Reviews " ] [ 0 ] [ " Content " ] ) . s t a r t s w i t h ( "<!−−" ) :
18
tmp = s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] )
19
i f not tmp in n e g r e v i e w s :
20
n e g r e v i e w s . append ( tmp )
21
return ( p o s r e v i e w s , n e g r e v i e w s )
Listing 6.2: Load review data.
6.2.2
Step 2: Preprocessing and Feature Extraction
After loading the data, the reviews have to be preprocessed and the features extracted
so they can be provided to the classification algorithm. The demonstrated preprocessing
steps may differ depending on the data set used. For instance, if the sentiment analysis is
performed on Twitter data, URLs have to be removed as they normally do not have an
influence towards the sentiment of the tweet. In sentiment analysis a string of characters,
whole words, and word combinations (so-called n-grams) are considered as features. In this
chapter we use a combination of single words (unigrams) and two-word combinations (bigrams).
We grouped all preprocessing and feature extraction steps in a function called preprocessing(), which is illustrated in Listing 6.3. As an argument the function takes a list of strings
each holding a text review, and the polarity of these comments, which might be positive
or negative. This function is later called two times: once for preprocessing the positive
comments, and once for preprocessing the negative ones. During the next steps, the preprocessing is demonstrated with the following example sentence:
Example: One which is smaller, but large enough to hold the camera but not large enough
to hold the charger, and your lunch. :-)
Sentiment Classification and Visualization of Product Review Data
139
1 def p r e p r o c e s s i n g ( a l l R e v i e w s , p o l a r i t y ) :
2
stemmer = SnowballStemmer ( " e n g l i s h " )
3
s t o p s = set ( s t o p w o r d s . words ( ’ e n g l i s h ’ ) + [ ’ . ’ , ’ , ’ , ’ ! ’ , ’ ? ’ , ’ ; ’ , ’ : ’ , ’ [ ’ , ’
] ’ ] ) − set ( [ ’ not ’ ] )
4
s m i l e y h a p p y = [ ’ : −) ’ , ’ : ) ’ , ’ ( : ’ , ’ ( −: ’ , ’ =) ’ , ’ (= ’ ]
5
s m i l e y s a d = [ ’ : −( ’ , ’ : ( ’ , ’ ) : ’ , ’ ) −: ’ , ’ =( ’ , ’ )= ’ ]
6
s m i l e y w i n k = [ ’ ; −) ’ , ’ ; ) ’ , ’ ( ; ’ , ’ ( −; ’ ]
7
pos reviews tok = [ ]
8
reviewsPreprocessed unigram = [ ]
9
reviewsPreprocessed bigram = [ ]
10
reviewsPreprocessed = [ ]
11
12
f o r r e v i e w in a l l R e v i e w s :
13
tmp = s t r ( r e v i e w ) . l o w e r ( )
14
f o r s m i l e y in s m i l e y h a p p y :
15
tmp . r e p l a c e ( s m i l e y , ’SMILEY HAPPY ’ )
16
f o r s m i l e y in s m i l e y s a d :
17
tmp . r e p l a c e ( s m i l e y , ’SMILEY SAD ’ )
18
f o r s m i l e y in s m i l e y w i n k :
19
tmp . r e p l a c e ( s m i l e y , ’SMILEY WINK ’ )
20
tmp = n l t k . w o r d t o k e n i z e ( tmp )
21
tmp = [ word f o r word in tmp i f word not in s t o p s ]
22
tmp = [ stemmer . stem ( word ) f o r word in tmp ]
23
r e v i e w s P r e p r o c e s s e d u n i g r a m . append ( l i s t ( set ( tmp ) ) )
24
25
b i g r a m F i n d e r = B i g r a m C o l l o c a t i o n F i n d e r . f r o m w o r d s ( tmp )
26
tmpBi = b i g r a m F i n d e r . n b e s t ( BigramAssocMeasures . c h i s q , 1 0 0 0 )
27
r e v i e w s P r e p r o c e s s e d b i g r a m . append ( l i s t ( set ( tmpBi ) ) )
28
29
def l a b e l i n g ( f e a t u r e s ) :
30
return d i c t ( [ ( f e a t u r e , True ) f o r f e a t u r e in f e a t u r e s ] )
31
32
f i n a l F e a t u r e s = [ ( l a b e l i n g ( f e a t u r e s ) , s t r ( p o l a r i t y ) ) f o r f e a t u r e s in
reviewsPreprocessed bigram ]
33
f i n a l F e a t u r e s += [ ( l a b e l i n g ( f e a t u r e s ) , s t r ( p o l a r i t y ) ) f o r f e a t u r e s in
reviewsPreprocessed unigram ]
34
return ( f i n a l F e a t u r e s )
Listing 6.3: Preprocessing and feature extraction.
Substitution of Emoticons
In general emoticons can play a significant role during sentiment classification [4]. From a
sentiment classification point of view, one challenge is that different emoticons exist having a
similar meaning. For instance, “:-)” and “=)” both indicate a feeling of happiness. To bring
such emoticons to a canonical form we substitute them to a unique symbolic representation
in form of a specific string. In lines 4 – 6 smilies are grouped into happy, sad, and wink
ones. Then in lines 14 – 19 all these smilies are replaced by the strings SMILEY HAPPY,
SMILEY SAD or SMILEY WINK. By this, the classifier treats different smilies with a similar meaning in the same way. It is recommended to apply this step before the tokenization
process, as depending on the tokenization approach the symbols relating to one emoticons
can be grouped to one token [:-)] or each symbol to an individual one [:,-,)]. The used list
of emoticons is covering only three classes of common emoticons. By further defining more
classes with more emoticons it is expected to have a positive influence to the classification
performance.
Example: One which is smaller, but large enough to hold the camera but not large enough
to hold the charger, and your lunch. SMILEY HAPPY
140
Text Mining and Visualization: Case Studies Using Open-Source Tools
Tokenization
In simple terms, tokenization means dividing a given text into smaller and meaningful
elements like sentences and words. For example, let us assume we have the following review
on a phone:
The battery of the Nokia phone is crap and the display as well.
After tokenization, the sentence could take the following form as illustrated in Table 6.3.
The
battery
of
the
Nokia
phone
is
crap
and
the
display
as
well
.
TABLE 6.3: Tokenization example.
The final output depends on the tokenizer applied. It could also be the case that “well”
appears together with full stop next to it (e.g., “well”) or the full stop might not be present
at all. Similarly, when we have “don’t” in the sentence, possible tokens could be “don’t” or
“do” and “n’t” . In our case we use the default tokenization function nltk.word tokenize(),
which first splits up a given text into its sentences by using the so-called Punkt-SentenceTokenizer followed by splitting up the sentences in words by applying the Treebank-wordtokenizer (line 20).
Example: [’one’, ’which’, ’is’, ’smaller’, ’,’, ’but’, ’large’, ’enough’, ’to’, ’hold’, ’the’,
’camera’, ’but’, ’not’, ’large’, ’enough’, ’to’, ’hold’, ’the’, ’charger’, ’,’, ’and’, ’your’, ’lunch’,
’.’, ’SMILEY HAPPY’]
Stopwords and Punctuation Marks Removal
Stopwords are words that occur frequently in text documents and possess only low
characteristics from a text classification point of view [5] [6]. Within Python there is a
defined corpora of stopwords for the English language in the module nltk.corpus called
stopwords, and it includes words like a, or, am, I, and as. In addition, we remove punctuation
marks like the full stop, comma or exclamation mark — in Listing 6.3 they are defined in
the variable stops in line 3 and removed in line 21. As illustrated in line 3 along with the
predefined words we add additional symbols like brackets to the stopwords list, as we expect
they do not have a significant impact on the sentiment classification. Moreover, we remove
the word “not” from the predefined stopword list, as it can have a significant impact on the
classification. In the last step of the tokenization process punctuation marks are removed.
Example: [’one’, ’smaller’, ’large’, ’enough’, ’hold’, ’camera’, ’large’, ’enough’, ’hold’,
’charger’, ’lunch’, ’SMILEY HAPPY’]
Text-normalization
In natural languages, words are modified according to their grammatical context. For
text classification in general, and also specifically for sentiment classification it is beneficial
to have a canonical and generalized form of the words. Therefore, as a first step, all characters of the text are set to lowercase (see line 13), so words like “camera” and “Camera”
are treated case insensitively as the same word. Afterwards, stemming is applied to derive
a generic form of a word independent of its inflections.
As it is the case with tokenization, there are different stemmers that could be used for this
purpose, and the right choice will depend on our goal [7]. NLTK in version 3 provides various
implementations of stemming supporting languages like English, German, and Spanish. In
addition, lemmatization is available. Lemmatization is a more advanced form of stemming,
which uses the contextual and grammatical information related to the word to determine
the lemma (i.e., the root of the word) [8] but at the same time is computationally more
Sentiment Classification and Visualization of Product Review Data
141
expensive. In general, there is a trade-off between the benefits of bringing words into their
canonical form by stemming and losing valuable information. An example for information
loss is “defense” and “defensive”, which by stemming will be reduced to “defens”. The former
word has a more positive connotation, whereas the latter a more negative; however, both
are treated after the stemming as the same. From our experience with product information,
stemming has a slightly positive influence on the sentiment classification, but we advise to
conduct experiments with and without stemming based on the own data and context the
reader is intending to apply sentiment classification. We have used the Snowball stemmer
(see line 22) which is initialized for the English language in line 2.
Example: [’one’, ’smaller’, ’larg’, ’enough’, ’hold’, ’camera’, ’larg’, ’enough’, ’hold’,
’charger’, ’lunch’, ’smiley happi’]
Feature Extraction
The features for the classifier consist of a combination of uni- and bigrams. Unigrams are
individual words resulting from the tokenization process. Bigrams are a version of n-grams
consisting of two words. Studies have shown that bigrams are effective features for classifying
text, especially in the context of sentiment analysis. Since bigrams are less common than
individual words, they help the classifier to make better decisions [9].
In general, n-grams in sentiment classification can be regarded in two ways. One is to
consider them as so-called presence features, which means it is only considered whether
a word appears or not in a text, and the other way is to consider them according to the
frequency they appear in a given text. In the latter case the frequency takes the role of a
weight of the feature. In sentiment classification experimental results show that the appearance based approach seems to be more effective, as the frequency of words are not related
to their degree of sentiment indication [10]. Therefore, we check only if a word appears
in the review but not its frequency. This is done in Listing 6.3 in line 23 by adding the
preprocessed words which are stored in the variable tmp by adding it to a list called reviewsPreprocessed unigram. We eliminate multiple occurrences of words by converting the
list of preprocessed words into a set data type. As in the mathematical concept of sets, this
data type holds only one instance of each element. Then this set is casted again into a list
data type for the later processing.
In the context of the classification of sentiment polarization, the individually extracted
features consisting out of unigrams and bigrams have different predictive performance. From
these features only those having a larger predictive performance should be selected. Normally features mainly occurring only in one of the two classes (positive or negative reviews) are considered to have a greater predictive power than features occurring in a similar frequency in both classes. The latter case can be considered as noise, as it does not
provide clear and meaningful information to the classifier. To quantify such relationships
between classes and features, measures like the information gain or the chi-square measure are used [5]. For that NLTK provides so-called n-gram association measures within
the NLTK.metrics.association module. Within this chapter we limit the demonstration at
bigram features and the selection based on the chi-square measure. Therefore in line 25 first
we are creating bigrams from the preprocessed words in variable tmp by creating an instance
of the BigramCollocationFinder class by using the method from words(). In the following
line the 1,000 most significant bigrams based on the chi-square independence test are returned. The value of 1,000 is based on our experience with the given data set, but varying
this parameter is in general a good idea for conducting further experiments to improve the
classifier. To illustrate this, in the following example we show unigrams and bigrams based
on the previous example sentence.
Example bigrams: [(’charger’, ’lunch’), (’enough’, ’hold’), (’larg’, ’enough’), (’lunch’,
142
Text Mining and Visualization: Case Studies Using Open-Source Tools
’smiley happi’), (’one’, ’smaller’), (’camera’, ’larg’), (’hold’, ’camera’), (’hold’, ’charger’),
(’smaller’, ’larg’)]
Example unigram: [[’one’, ’smaller’, ’larg’, ’enough’, ’hold’, ’camera’, ’larg’, ’enough’,
’hold’, ’charger’, ’lunch’, ’smiley happi’]]
As a last step during the preprocessing and feature creation the derived unigrams and
bigrams have to be labeled depending on whether they appear in a positive or negative
review. Therefore, we define a function labeling in line 29, which transforms the processed
data into a dictionary data type with annotated label, which is held into the variable polarity
and contains in our case the string “pos” for positive and “neg” for negative reviews.
Example: results of labeling the data with a “pos” label:
[({’enough’: True, ’smiley happi’: True, ’smaller’: True, ’camera’: True,
’charger’: True, ’not’: True, ’lunch’: True, ’larg’: True, ’hold’: True, ’one’: True},
’pos’), ({’not’: True, ’camera’: True}, ’pos’), ({’charger’: True, ’lunch’: True},
’pos’), ({’enough’: True, ’hold’: True}, ’pos’), ({’enough’: True, ’larg’: True}, ’pos’),
({’lunch’: True, ’smiley happi’: True}, ’pos’), ({’one’: True, ’smaller’: True}, ’pos’),
({’hold’: True, ’camera’: True}, ’pos’), ({’charger’: True, ’hold’: True}, ’pos’),
({’larg’: True, ’not’: True}, ’pos’), ({’larg’: True, ’smaller’: True}, ’pos’)]
6.2.3
Step 3: Sentiment Classification
Following the feature extraction explained in the previous subsection, in a next step
we continue with the sentiment classification process by training our classifier. This will
allow us to automatically classify new reviews into positive or negative. Within the context
of sentiment analysis the most applied classifiers are Naı̈ve Bayes, Maximum Entropy and
Support Vector Machines. The first two are implemented within the NLTK library, the
latter is provided in the the scikit-learn module. In this chapter we are focusing on the Naı̈ve
Bayes approach which provides good accuracy and performance in sentiment classification.
Nevertheless, the readers are encouraged to try the other two classifiers as well.
Simply put, the Naı̈ve Bayes classifier makes observations from the training data and
estimates the probability of a negative or positive sentiment based on the presence of certain
features. It is especially useful for classification tasks with a large number of attributes or
dimensions that need to be taken into consideration to calculate the probability of an event
to occur (or in this case a sentiment to be positive or negative). It treats all features as
independent, which leads to a simple yet effective classifier [11]3 .
We encapsulated all needed functionalities within the function sentiment classification
in Listing 6.4, which takes as argument a list of training features and a list of test features.
After we initialized the variables for the later determination of a confusion matrix in line
5, an instance of the Naı̈ve Bayes classifier is created and trained with the training data
(line 7). Afterwards we iterate through all test samples (line 8-20), applying one review per
iteration at the classifier (line 9), and determine the type of classification error (line 11-20).
We determine the features with the most informative values according to the classifier in
line 22 (this is commented out in the listing). An example result of the most informative
features for the product category mobilephone can be seen in Table 6.4. In the end the
function returns the four values of the determined confusion matrix.
3 For
more information on the Naı̈ve Bayes we refer you to [12] and [13]
Sentiment Classification and Visualization of Product Review Data
Feature
(’return’, ’item’) = True
(’money’, ’back’) = True
(’sent’, ’back’) = True
(’high’, ’recommend’) = True
(’took’, ’phone’) = True
(’not’, ’return’) = True
crisp = True
(’worst’, ’phone’) = True
shot = True
(’not’, ’unlock’) = True
refund = True
143
Information degree regarding sentiment
neg : pos = 22.6 : 1.0
neg : pos = 22.2 : 1.0
neg : pos = 21.6 : 1.0
pos : neg = 19.5 : 1.0
neg : pos = 17.3 : 1.0
neg : pos = 17.3 : 1.0
pos : neg = 17.3 : 1.0
neg : pos = 16.8 : 1.0
pos : neg = 16.6 : 1.0
neg : pos = 16.5 : 1.0
neg : pos = 15.6 : 1.0
TABLE 6.4: Example results for information degree per feature.
1 def s e n t i m e n t c l a s s i f i c a t i o n ( t r a i n f e a t s , t e s t f e a t s ) :
2
#i n i t i a l i z e c o n f u s i o n m a t r i x v a r i a b l e s
3
#TN = t r u e n e g a t i v e ; TN = t r u e p o s i t i v e
4
#FN = f a l s e n e g a t i v e ; FP = f a l s e p o s i t i v e
5
(TN, TP, FN, FP) = ( 0 , 0 , 0 , 0 )
6
7
c l a s s i f i e r = NaiveBayesClassifier . train ( trainfeats )
8
f o r ( t e s t R e v i e w , p o l a r i t y ) in t e s t f e a t s :
9
predictedSentiment = c l a s s i f i e r . c l a s s i f y ( testReview )
10
#c o u n t r e s u l t i n g p r e d i c t i o n c l a s s e s
11
i f p r e d i c t e d S e n t i m e n t == p o l a r i t y :
12
i f p r e d i c t e d S e n t i m e n t == "neg" :
13
TN += 1
14
else :
15
TP += 1
16
else :
17
i f p r e d i c t e d S e n t i m e n t == "neg" :
18
FN += 1
19
else :
20
FP += 1
21
#uncommen t h e f o l l o w i n g l i n e t o show t h e most i n f o r m a t i v e f e a t u r e s
22
#c l a s s i f i e r . s h o w m o s t i n f o r m a t i v e f e a t u r e s ( 3 0 )
23
return ( (TN, TP, FN, FP) )
Listing 6.4: Sentiment classification.
6.2.4
Step 4: Evaluation of the Trained Classifier
In general the accuracy of classifiers is measured by dividing a given data set into a
training and testing sample. This process is known as validation and it is used to assess the
ability of our classifier to accurately classify new unlabeled reviews. One of the most used
strategies for evaluating classifier is the k-fold cross-validation approach. In this approach
the given data set is divided into k mutual-exclusive subsets. During the evaluation phase,
k-1 subsets are used for training the classifier and one subset for testing its accuracy. The
average accuracy is then estimated by iterating k-times through all possible subset combinations until all subsets are used once as test set. In practice, a 10-fold cross-validation has
shown good results for estimating the overall classification accuracy [14]. In this chapter
we apply the so-called stratified version of the 10-fold-cross validation. Stratification means
that the frequency of all class representations in the k-1 subsets are similar to the one of
144
Text Mining and Visualization: Case Studies Using Open-Source Tools
Predicted Sentiment
Positive
Negative
Actual Sentiment
Positive
Negative
T rueP ositive F alseP ositive
F alseN egative T rueN egative
TABLE 6.5: Structure of confusion matrices in case of binary classification.
the original data set. In a sentiment classification task, this will ensure that the ratio of
positive and negative opinions appear in each of the k subsets in equal frequency.
To assess the quality of the sentiment classifications we calculate the general accuracy
measure as well as the contingency matrix. The accuracy is calculated by the following
CorrectP redictions
formula: accuracy = numberOf
N umberOf AllP redictions . To get a more detailed quality assessment of
the classifier’s prediction, besides the overall accuracy, we determine the so-called confusion
matrix. The confusion matrix indicates the frequency of correct and incorrect classifications
for both positive and negative sentiments. Table 6.5 illustrates the structure of such a
matrix. Expecially in cases where the classes are immoderately unbalanced, as it is with the
data set used in this chapter, this is an important aspect to consider. For instance in the
mobile phone category there are approx. four times more positive than negative reviews.
In the extreme case that a classifier just predicts every review to be positive, without
considering any features of the review, we get an accuracy of 79%. By this each positive
review is classified correctly, and each negative one incorrectly. In our case of predicting the
sentiment of reviews for marketing decision, we rather need both classes to be predicted in
a similarly good quality, in order to get an accurate view on the customers’ opinion, and not
having one class beeing very accurate, and one beeing too inaccurate. To evaluate the quality
of the prediction of the classes occuring with minor frequency, an overall confusion matrix is
calculated. The matrix concists out of four entries: True Positive (TP) and True Negative
(TN) entries, indicating the correct classification of positive and negative opinions; and
False Positive (FP) and False Negative (FN) entries, indicating the incorrect classification.
In Listing 6.5 we first initialize three variables holding statistics about the maximal,
minimal and average accuracy determined during the k-folds. In line 11 the number of folds
are defined. As mentioned already, in practice often 10 folds are used. Then we get the
positive and negative comments by loading the data specified in variable opinionDataset
with the function readJSONdata(), which is explained in Section 6.2.2 step 2. Afterwards
we preprocess these data in lines 9 and 10 by the function defined in Section 6.2.3 step 3.
For the evaluation we use the cross validation module from the scikit-learn library (line 25).
Therefore, we have to prepare an array storing the labels for review data, which we store
in the variable data in line 19. The cross-validation function is expecting a numpy array.
One special feature of these arrays is the possibility to index the field with a list of indices,
which is used by the cross validation function. We define the array in line 22 by labeling
the positive reviews with a 1 and the negative reviews with a −1. Then we instantiate a
cross validation object by giving the labeled list and the review data as an argument. In the
for-loop in line 28 we iterate through all the k test- and training-set combinations (in our
case 10-times). As the NLTK classifier expects standard python arrays, not numpy arrays,
we reconvert them in lines 29 and 30.
Finally, we call our classifier function sentiment classification() by passing the training
and test dataset created in previous steps. In the following lines, we store statistics like the
minimal and maximal accuracies during the loops, calculate, and print the total average
accuracy in lines 42 and 45.
In Table 6.6 the average accuracies based on a stratified 10-fold cross-validation are illustrated as well as the minimal and maximal accuracies of the 10 training and test iterations.
Sentiment Classification and Visualization of Product Review Data
Product
Category
Video Surveillance
TVs
Tablets
Mobile Phone
Laptops
Cameras
Average
Accuracy
0.73
0.63
0.76
0.84
0.72
0.85
Minimal
Accuracy
0.54
0.31
0.55
0.80
0.46
0.82
145
Maximal
Accuracy
0.87
0.91
0.83
0.89
0.91
0.90
TABLE 6.6: Sentiment classification accuracies per product category.
Predicted Sentiment
Positive
Negative
Actual Sentiment
Positive Negative
519.4
82.3
21.8
53.3
TABLE 6.7: Resulting average confusion matrix for the mobile phone category.
The average consistancy table resulting from the cross-validation process is demonstrated
in Table 6.7. This table is built by summing all entries of the consistency table per crossvalidation iteration, and by dividing this sum by the number of folds (10 in our case). This
table shows that positive reviews are classified with an accuracy of 96%, and the negative
ones with an accuracy of 61%. This accuracy is calculated by the following formula:
P
TN
accuracypositive = T PT+F
P , and accuracynegative = T N +F N .
For increasing the accuracy of the negative classifications, we suggest to experiment with
resampling techniques of the data set such as oversampling or undersampling to reduce the
inbalance of the classes. The idea of oversampling is to randomly duplicate entries of the
minority class, whereas in undersampling entries of the majority class are randomly removed
[15].
6.2.5
Step 5: Visualization of Review Data
In the first four steps of the process, we trained a classifier and managed to achieve
an overall accuracy of approximately 75%. We now know that the category mobile phone
has approximately 80% positive and 20% negative sentiment. Therefore, we are interested
to gain more insight and understand the context behind such sentiment. What do the
customers like or dislike about the phones we sell? To achieve this, commercial tools as
Sysomos are using different text visualization options:
• Word clouds: representation of most frequent words occurring within a text
• Buzz graphs: not only the keywords, but also the relationships between them – how
often words are mentioned in relation to each other
• Popular phrases within the text
• Entities: categorization of words (e.g., name, place, organization, person) [16]l
146
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
Text Mining and Visualization: Case Studies Using Open-Source Tools
#i n i t i a l i z e s t a t i s t i c s
maxAccuracy = 0
minAccuracy = 1
averageAccuracy = 0
average TP = 0
average TN = 0
average FP = 0
average FN = 0
variables
#number o f f o l d s used i n c r o s s v a l i d a t i o n
kFold = 10
o p i n i o n D a t a s e t = r "C: \ p r o j e c t \ AmazonReviews \ mobilephone \∗"
p o s r e v i e w s , n e g r e v i e w s = readJSONdata ( o p i n i o n D a t a s e t )
p o s f e a t s = p r e p r o c e s s i n g ( p o s r e v i e w s , ’ pos ’ )
n e g f e a t s = p r e p r o c e s s i n g ( n e g r e v i e w s , ’ neg ’ )
data = p o s f e a t s + n e g f e a t s
#p r e p a r e l a b e l a r r a y i n d i c a t i n g p o s i t i v e r e v i e w s with 1 , n e g a t i v e with −1
l a b e l = np . a r r a y ( [ 1 ] ∗ l e n ( p o s f e a t s ) + [ − 1 ] ∗ l e n ( n e g f e a t s ) )
#c r e a t e i n d i c e s f o r k c r o s s v a l i d a t i o n c o m b i n a t i o n s
c r o s s V a l i d a t i o n I t e r = c r o s s v a l i d a t i o n . S t r a t i f i e d K F o l d ( l a b e l , n f o l d s=
kFold )
#p e r f o r m s t r a t i s f i e d k−f o l d c r o s s −v a l i d a t i o n
for train index , test index in crossValidationIter :
t r a i n i n g D a t a = [ data [ i ] f o r i i n t r a i n i n d e x ]
t e s t D a t a = [ data [ i ] f o r i i n t e s t i n d e x ]
(TN, TP, FN, FP) = s e n t i m e n t c l a s s i f i c a t i o n ( t r a i n i n g D a t a , t e s t D a t a )
a c c u r a c y = (TP + TN) / (TN + TP + FN + FP)
average TP += TP
average TN += TN
average FN += FN
average FP += FP
i f a c c u r a c y > maxAccuracy : maxAccuracy = a c c u r a c y
i f a c c u r a c y < minAccuracy : minAccuracy = a c c u r a c y
a v e r a g e A c c u r a c y += a c c u r a c y
#p r i n t a v e r a g e c l a s s i f i c a t i o n a c c u r a c y
p r i n t ( " Average Accuracy : " + s t r ( ( a v e r a g e A c c u r a c y /kFold ) ) + "\ nMinimal
Accuracy : "+ \
s t r ( minAccuracy ) + "\nMaximum Accuracy : " + s t r ( maxAccuracy ) )
#p r i n t a v e r a g e c o n f u s i o n m a t r i x
p r i n t ( "\ nAverage C o n f u s i o n Matrix \nTP : " + s t r ( average TP /kFold ) + " FP :
" + s t r ( average FP /kFold ) + \
"\nFN : " + s t r ( average FN /kFold ) + " TN: " + s t r ( average TN/kFold ) )
Listing 6.5: Evaluation of the classifier.
Although these techniques might differ slightly in terms of the way they are named, in
most cases they work on the same principle. A word cloud will only show the frequency
of the words appearing in the reviews, which is not enough to understand the context
behind the sentiment. Similarly, the buzz graphs show the relationship between words, but
not the sentiment context. Therefore, it is more useful when different visualizations are
used in combination rather than independently. There are currently some freely available
tools for text visualization like Wordle,4 Voyant Cirrus, and5 Jigsaw.6 They are easy and
4 http://www.wordle.net.
5 http://voyant-tools.org/tool/Cirrus.
6 http://www.cc.gatech.edu/gvu/ii/jigsaw/index.html.
Sentiment Classification and Visualization of Product Review Data
147
straightforward to use, mainly requiring a simple copy and paste of your data to display
the visualization.
1
2
3
4
5
6
7
import
import
import
import
json
glob
io
re
o p i n i o n D a t a = r "E : \ t e s t −data \ AmazonReviews \ l a p t o p s \∗"
posCommentsFile = r "E : \ t e s t −data \ AmazonReviews \ o n l y c o m m e n t s \
laptops pos . txt"
8 negCommentsFile = r "E : \ t e s t −data \ AmazonReviews \ o n l y c o m m e n t s \
laptops neg . txt"
9
10 a l l f i l e s = g l o b . g l o b ( o p i n i o n D a t a )
11 p o s r e v i e w s = [ ]
12 n e g r e v i e w s = [ ]
13
14 f p o s = i o . open ( posCommentsFile , ’ w ’ , e n c o d i n g=" u t f 8 " )
15 f n e g = i o . open ( negCommentsFile , ’ w ’ ,
e n c o d i n g=" u t f 8 " )
16
17 f o r elem i n a l l f i l e s :
18
with open ( elem ) a s j s o n f i l e :
19
json data = json . load ( j s o n f i l e )
20
i f j s o n d a t a . g e t ( " Reviews " ) :
21
i f " Content " i n j s o n d a t a [ " Reviews " ] [ 0 ] . k e y s ( ) :
22
i f " O v e r a l l " i n j s o n d a t a [ " Reviews " ] [ 0 ] . k e y s ( ) :
23
i f ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 5 . 0 " o r j s o n d a t a [ "
Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 4 . 0 " ) and j s o n d a t a [ " Reviews "
] [ 0 ] [ " Content " ] and not s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ "
Content " ] ) . s t a r t s w i t h ( "<!−−" ) :
24
tmp = s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] )
25
i f not tmp i n p o s r e v i e w s :
26
f p o s . w r i t e ( ( tmp+s t r ( ’ \ n\n ’ ) ) )
27
p o s r e v i e w s . append ( tmp )
28
i f ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 1 . 0 " o r j s o n d a t a [ "
Reviews " ] [ 0 ] [ " O v e r a l l " ] == " 2 . 0 " ) and j s o n d a t a [ " Reviews "
] [ 0 ] [ " Content " ] and not s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ "
Content " ] ) . s t a r t s w i t h ( "<!−−" ) :
29
tmp = s t r ( j s o n d a t a [ " Reviews " ] [ 0 ] [ " Content " ] )
30
i f not tmp i n n e g r e v i e w s :
31
f n e g . w r i t e ( ( tmp+s t r ( ’ \ n\n ’ ) ) )
32
n e g r e v i e w s . append ( tmp )
Listing 6.6: Data extractor.
To demonstrate some of the visualization techniques, we decided to use the category
mobile phones and the open-source visualization tools Wordle (for word cloud) and Jigsaw
(for word tree).
In a first step, however, it is necessary to extract the content of the reviews. As seen
in the data set description, each review file contains not only the content but also other
attributes like ID and author. To achieve good results, we need to remove all other attributes
and keep the relevant text content only. With the help of a small script (see Listing 6.6),
we extracted all positive and all negative reviews per category and saved them in separate
text files accordingly by using the following scheme: product category neg.txt and product
category pos.txt, where product category is replaced by laptops or cameras, for example.
On the Wordle homepage, just click “Create your own” link, and you will see a screen
where you can paste your text. We pasted the clean data from both the negative and positive
reviews from this category.7 Click on the “Go” button. Wordle will generate a word cloud
7 Use
files mobilephone neg.txt and mobilephone pos.txt
148
Text Mining and Visualization: Case Studies Using Open-Source Tools
that can be customized to a certain extent by changing the color, the layout, font, and, most
importantly, excluding stopwords in several languages. The initial word cloud will show that
the word “phone” was used most often. This is understandable and not really useful for the
analysis. Wordle allows the exclusion of any word from the generated word cloud. Right
click on the word “phone” and choose “Remove phone”. You should now see a word cloud
as in Figure 6.2.
FIGURE 6.2: Word cloud for the positive and negative reviews of the mobilephone category.
From Figure 6.2 it becomes clear that users have mostly discussed the words “screen”,
“battery”, and “camera”. This is a valuable and interesting information, but it is unfortunately not detailed enough since we cannot understand the context behind those words.
To achieve this, we would benefit from a visualization technique called word tree. In the
following, we generate a word tree and explain the insights it brings to the analyst.
The word tree visualization is a more complex visualization than a word cloud, and gives
the best context overview of any word or phrase you want to search for and that is mentioned
in the reviews. To demonstrate this, we will use Jigsaw, a tool for visual exploration of text
data. It is available as a free download.
After you start the tool, you see the welcome screen, as shown in Figure 6.3. To import
FIGURE 6.3: Jigsaw’s welcome screen.
the data, click on “File” in the upper menu and then “Import”. You will see a screen as in
Figure 6.4.
Sentiment Classification and Visualization of Product Review Data
149
FIGURE 6.4: Import screen in Jigsaw.
Click “Browse” and locate the folder only comments where we stored the cleaned
product data. You can choose to import one or more TXT files to import in a single step.
We chose to import the product reviews for the category mobile phone with both positive
and negative sentiment. These are the files mobilephone pos.txt and mobilephone neg.txt.
Click “Import”. You will see a screen as in Figure 6.5.
FIGURE 6.5: Entity identification screen in Jigsaw.
Since for this visualization it is not necessary to identify entities, select “None”, and
click “Identify”. After the data import is finished, the welcome screen appears with a short
confirmation that there are two documents in the project. To create a word tree, click on
“Views”, and select “Word Tree”.
Initially, you see a relatively blank screen, containing a menu and a search bar. This is
because the visualization works only when the user enters a word or a phrase. From the
generated word cloud we now know which words are mentioned more often in the reviews.
The word tree can reveal the reason and perhaps the sentiment behind this. If we enter
“screen” in the search box, we get a branch like structure as in Figure 6.6, where screen is
the root of the tree and the branches are the different contexts in which this word appears
in the text.
150
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 6.6: Word tree view for “screen”.
Some of the words in Figure 6.6 are in larger font size. This means that they appear
more frequently in the reviews. To investigate more closely the branch “is”, we can click on
it. The screen zooms in and displays only that branch, as in Figure 6.7.
From the word tree visualization it can be concluded that users mention “screen” mostly
in a positive context. Some of the sentences that appear are:
“screen is very responsive.”
“screen is bright and clear.”
“screen is small but everything comes out in great detail.”.
To inspect specifically the negative context in which screen is mentioned, if any, it would
be easier to import the file with negative reviews only and investigate further. Furthermore,
Jigsaw initially displays a word tree based on part of the available data for the chosen word
or phrase. You can change this setting in the prunning section at the bottom of the word
tree view window (See Figure 6.7).
This type of visualizations can contribute significantly to understanding the reason behind the sentiment. It is certainly much easier to use a professional tool, which offers diverse
text visualizations. However, these tools are usually quite expensive and a perfect solution
is yet to be developed because of the many disadvantages of text data. It is difficult for
computers to understand language the way humans do. On the other hand, due to social
media, text data are growing exponentially and the need to develop more sophisticated
tools and techniques for text analytics is still present. The main disadvantage of relying on
open-source tools is that one may need to use several tools to get the desired results.
Sentiment Classification and Visualization of Product Review Data
151
FIGURE 6.7: Word tree view for “screen is”.
6.3
Summary
In this chapter, we presented a framework approach for training a sentiment classifier
on product review data. It uses a preexisting overall rating score as a proxy for positive or
negative sentiment and is able show some relationship between this and the text contained
within the review. We also demonstrated an approach for determining the context behind the
sentiment by applying different visualization techniques. We now give recommendations for
further improvement of our classifier. Depending on the data set, it is advisable to perform
several experiments by changing and applying different variables and settings:
• Emoticons: by further defining more classes with more emoticons it is expected to
have a positive influence to the classification performance.
• Tokenizer: via the NLTK other tokenizers can be employed. For sentiment analysis the
reader can try the Regexp Tokenizer where individualized rules for the token creation
can be defined. Another option is the Stanford tokenizer.
• Classifier: we used the Naı̈ve Bayes classifier because of its good accuracy and performance in the context of sentiment classification. Nevertheless, the reader in encouraged
to try further classification algorithms like the maximum-entropy or support vector
machines, which are provided in the NLTK and scikit-learn module.
• Data sampling: because of the unbalanced nature of the used data set the prediction
accuracy of the class of a smaller sample size is lower than the predictions of the
class which has a larger sample size. By applying oversampling or undersampling
techniques, the prediction quality of the smaller class may be increased (see Section
6.2.4).
152
Text Mining and Visualization: Case Studies Using Open-Source Tools
Bibliography
[1] B. Liu. Sentiment analysis and opinion mining, volume lecture #16 of Synthesis lectures on human language technologies. Morgan & Claypool, San Rafael, CA, 2012.
[2] Q. E. McCallum. Bad data handbook. O’Reilly Media, Sebastopol, CA, 2012.
[3] H. Wang, Y. Lu, and C. X. Zhai, editors. Latent aspect rating analysis without aspect
keyword supervision. ACM, 2011.
[4] A. Hogenboom, D. Bal, F. Frasincar, N. Bal, F. de Jong, and U. Kaymak. Exploiting
emoticons in polarity classification of text. J. Web Eng., 14(1&2):22–40, 2015.
[5] R. Feldman and J. Sanger. Text mining handbook: advanced approaches in analyzing
unstructured data. Cambridge University Press, 2006.
[6] S. M. Weiss, N. Indurkhya, and T. Zhang. Fundamentals of predictive text mining.
Springer-Verlag, 2010.
[7] G. S. Ingersoll, T. S. Morton, and A. L. Farris. Taming text: How to find, organise,
and manipulate it. Manning and Pearson Education [distributor], 2013.
[8] G. Miner. Practical text mining and statistical analysis for non-structured text data
applications. Academic Press, 2012.
[9] J. Perkins. Python 3 text processing with NLTK 3 cookbook: Over 80 practical recipes
on natural language processing techniques using Python’s NLTK 3.0. Packt Pub., 2014.
[10] B. Pang and L. Lee. Opinion mining and sentiment analysis. Found. Trends Inf. Retr.,
2(1-2):1–135, January 2008.
[11] M. J. Zaki and W. Meira. Data mining and analysis: Fundamental concepts and algorithms. Cambridge University Press, 2014.
[12] B. Lantz. Machine Learning with R: Learn how to use R to apply powerful machine
learning methods and gain an insight into real-world applications. Packt, 2013.
[13] S. Bird, E. Klein, and E. Loper. Natural language processing with Python. O’Reilly,
2009.
[14] J. Han and M. Kamber. Data mining: concepts and techniques. Morgan Kaufmann,
2006.
[15] B. W. Yap, K. A. Rani, H. A. A. Rahman, S. Fong, Z. Khairudin, and N. N. Abdullah.
An application of oversampling, undersampling, bagging and boosting in handling imbalanced datasets. In Proceedings of the First International Conference on Advanced
Data and Information Engineering (DaEng-2013), pages 13–22. Springer, 2014.
[16] Nilesh Bansal.
Sysomos Text Analytics Roundup: Making Sense of
Data [ONLINE] Available at http://blog.sysomos.com/2011/04/18/
sysomos-text-analytics-roundup-making-sense-of-data/. Accessed 01
August, 2015.
Chapter 7
Mining Search Logs for Usage Patterns
Tony Russell-Rose
UXLabs, UK
Paul Clough
Information School, University of Sheffield, UK
7.1
7.2
7.3
7.4
7.5
7.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Getting Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Using Clustering to Find Patterns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.1
Step 1: Prepare the Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.2
Step 2: Cluster Using Weka . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
7.3.3
Step 3: Visualise the Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Replication and Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
153
154
158
159
159
161
164
168
170
Introduction
One of the greatest opportunities and challenges of the 21st century is the ever-increasing
significance of data. Data underpin our businesses and our economy, providing awareness
and insight into every sphere of life, from politics to the environment, arts, and society. The
everyday interactions between people and devices can be harnessed to power a new generation of products and services, allowing us to better understand human needs, aspirations,
and behaviour.
One particular area that has received considerable interest in recent years is search or
transaction log mining [1]. Valuable insights can be gained from analysing the traces people
leave when they search for digital information. In browsing the Web, people reveal something
about their behaviour and habits but little about their intent. By contrast, when people
search for information, they express in their own words their explicit needs and goals. This
data represent a unique resource that offers extraordinary potential for delivering insights
that can drive the next generation of digital services and applications.
Various studies have been undertaken to understand how and why people interact with
search engines. Such studies have led to the creation of frameworks that describe distinct
patterns of use, ranging from individual queries to entire information seeking episodes.1
These patterns may focus on information-seeking behaviour [3], the types of search tasks
that users perform [4], their goals and missions [5], their task-switching behaviour [6], or the
tasks, needs, and goals that they are trying to address when using search systems [4, 7]. In
this article we use search logs to analyse search patterns: the actions (sometimes referred to
1 A search episode is defined as “the activity in between the initiation and termination of interacting
with a particular external information resource” [2] (p. 394).
153
154
Text Mining and Visualization: Case Studies Using Open-Source Tools
as moves) taken in conducting a search. The search patterns are behavioural manifestations
of users’ search strategies: the approaches people take in searching for information [8].
It is, however, important to recognise that the logs only capture what people do (i.e.,
their search patterns), not why they do it (i.e., their search strategies or mental models).
As Jansen [9] clearly points out, “ [Query log data] cannot account for the underlying situational, cognitive or affective elements of the searching process”. This is a known limitation
of relying on evidence from search logs alone to characterise and establish search behaviour
where additional techniques, such as interviews, can be used to capture more subjective and
in-depth data [10, 11].
Indeed, the academic community is not alone in showing an interest in mining search logs.
Two highly influential commercial organisations, ElasticSearch and LucidWorks, have both
recently released independent log file analysis platforms (Kibana2 and SiLK3 , respectively).
What unites all of these efforts is the belief that finding distinct, repeatable patterns of
behaviour can lead to a better understanding of user needs and ultimately a more effective
search experience.
In this chapter, we explore the use of data mining techniques to find patterns in search
logs, focusing on the application of open source tools and publicly available log data. To
start with we outline the fundamentals of search log analysis, examining the basic anatomy
of search logs, search sessions and the features that characterise them (Section 7.2). We
then explore the use of clustering to find patterns in search logs, and describe a number of
experiments in which the Weka platform [12] was used to analyse sessions from a publicly
available query log (Section 7.3). We then discuss role of validation and the extent to which
different approaches can reveal different patterns in the data, and consider what this might
mean for the replication of previous published studies (Section 7.4). Finally, we draw these
threads together and reflect on the outputs from the analysis and our conclusions regarding
the methodology itself (Section 7.5).
7.2
Getting Started
Search logs are created when a search engine records its interactions with other agents.
Typically, such agents are people interacting with the search engine via a Web browser
or other type of client application. However, automated agents such as bots, spiders, and
indexing agents can also interact with a search engine, and their activity may be recorded
in the logs. Each interaction results in the addition of one or more records to the log, so
these files can eventually grow to several gigabytes in size.
Since there are many different types of search application, serving different audiences
with different content for different needs, the type of information that gets stored in search
logs also varies. Some logs can be relatively simple, storing just minimal information,
whereas others can be highly detailed, with extensive metadata describing each interaction. No standard format exists for what should be recorded in a search log, or indeed how
it should be formatted. But as a minimum a search log should store the content of the
query strings submitted by the user and the items that they click on in response. Jansen [9]
provides a useful conceptual model for a typical search transaction log consisting of queries
and page views that can be realised as a relational database.
Since logs can vary so widely, it is prudent to start with one that is well-known and
2 http://www.elasticsearch.org/overview/kibana/
3 http://www.lucidworks.com/lucidworks-silk/
Mining Search Logs for Usage Patterns
155
publicly accessible. One such example is the AOL search log.4 released in 2006 [13]. It
consists of a compressed text file containing 21,011,340 queries by 657,426 users over a 3month period. This log has a somewhat infamous history, due to concerns that users could
potentially be identified by their searches.5 However, it is useful as it is relatively large
and the records contain a variety of fields that offer a range of data mining possibilities. A
sample from the log is shown in Figure 7.1.
FIGURE 7.1: A sample of records from the AOL log.
The data include the following fields (AnonID, Query, QueryTime, ItemRank, ClickURL):
• AnonID: an anonymous user ID number.
• Query: the query issued by the user, case shifted with most punctuation removed.
• QueryTime: the time at which the query was submitted for search.
• ItemRank: if the user clicked on a search result, the rank of the item on which they
clicked is listed.
• ClickURL: if the user clicked on a search result, the domain portion of the URL in
the clicked result is listed.
The search log in Figure 7.1 represents users’ actions with the search engine (either a
query or clicked search result) using a customised format. However, it is common that the
starting point is the raw web server log that will contain all request/responses between
the web browser and server resulting in more lines in the log. This requires filtering to
identify only those lines in the log that reflect user actions rather than system requests,
such as elements to create a web page (e.g., requests for images and CSS files). In addition,
non-human requests (e.g., from bots) must be excluded from the dataset for analysis as
the focus in search log analysis is the user–system interactions. In his book on Web Data
Mining, Bing Liu [14] (pp. 527–594) provides a comprehensive discussion of methods that
can be used to pre-process (also known as data preparation) and analyse transactional data,
including search logs.
4 The AOL log can be downloaded from: http://www.gregsadetsky.com/aol-data/. Other logs
can be found at http://jeffhuang.com/search_query_logs.html and http://faculty.ist.psu.
edu/jjansen/academic/transaction_logs.html.
5 http://en.wikipedia.org/wiki/AOL_search_data_leak.
156
Text Mining and Visualization: Case Studies Using Open-Source Tools
When analysing search logs it is helpful to think of them in terms of user sessions:
periods of continued interaction between a specific user and a search application during
one visit to the site [9, 15, 16, 17]. A session typically begins with a query and ends with a
webpage or some arbitrary time-based cut off, such as 30 minutes of inactivity, and contains
a continuous sequence of user actions with the system [18].
By convention, a session captures the interactions between a specific IP address and
search system. The IP address can be used as a proxy for user to identify repeating patterns
across multiple sessions (anonymised user ID in the log described in Figure 7.1). However,
this can be problematic for a number of reasons. For example, Internet Service Providers
may assign different IP addresses to the same user as they browse the web, and multiple
users may share a single IP address (e.g., the users of a library catalogue situated in a
publicly accessible location). More reliable methods include using client-side cookies or user
authentication if available; alternatively if using IP address then additional information,
such as user agent and referrer, can also be used to identify users more reliably.
Various methods exist for segmenting the search logs into sessions (commonly referred to
as sessionisation) [15]. The simplest methods are time-based and rely on timstamping each
user action (typically starting with the initial query). One approach assumes that sessions
do not exceed some pre-defined length (e.g., no longer than one hour); another approach
involves computing the times between consecutive lines in the log from a single user and
segmenting sessions based on a maximal time gap (e.g., 30 minutes). The assumption behind
this is that long periods of inactivity will signal the start and end of different sessions. Of
course these assumptions can be broken: users may engage in very long sessions beyond
some pre-defined maximum and users may re-convene sessions following a break. Other,
more sophisticated methods make use of referrer information, site structure and semantic
information derived from queries to infer session breaks [15].
By way of example, taking a simple time-based approach based on segmenting sessions
if the time difference between lines from the same user exceeds 30 mins (a commonly used
time-out period), in the previous sample we might infer that there are seven individual
sessions, as illustrated in Figure 7.2.
FIGURE 7.2: A sample from the AOL log divided into sessions.
Evidently, the 30-minute cut off is only a heuristic, as it can indicate two sessions where
one would perhaps be more reasonable (e.g., sessions 3 and 4), and vice versa. During a given
session a user may work on multiple tasks (often referred to in the literature as episodes
— a subset of semantically or functionally related consecutive actions [14] (p. 537)), such
Mining Search Logs for Usage Patterns
157
as researching holiday destinations whilst simultaneously comparing alternative weather
forecasts. Likewise, a given task may be performed across multiple sessions [5, 6] or across
multiple channels such as desktop and mobile [19].
To characterise search patterns, we need to identify and quantify suitable features. These
features could capture aspects of the user’s query formulation and reformulation strategies,
resources selected and viewed and specific features of the search system that are used during the session. These features can be used to investigate aspects of users’ search strategies.
Common approaches to analysing search patterns include using information mined from
the logs to answer specific questions or hypotheses; alternatively a more inductive approach
could be used based on exploratory analyses. Search patterns mined from logs can be used
(with care) to predict search behaviour. For example, researchers have used evidence gathered from search logs (e.g., the rank position of clicked items, number of query reformulations
and length of session) to predict whether a search was likely to have been successful or not
[20]. Patterns of users’ click behaviour (e.g., the time spent viewing a page — dwell time —
and subsequent actions) have also been shown to be good predictors of document relevance
[21].
Measures based on quantifying feature occurrences can be computed for the session (e.g.,
total number of queries issued, total number of items selected, the ratio of queries to clicks)
offering a descriptive summary. Analysis can also be carried out on sequences of actions, for
example by computing a transition matrix that computes the probability of moving from
one action to another (e.g., the likelihood of a click following a query is typically very high).
Analysis may also be carried out to study patterns of query reformulation that may also
suggest common strategies for refining a search [22, 23, 24, 25].
In analysing users’ search strategies, Vakkari [26] gathers information about three aspects
of a search session: (i) the search terms used (e.g., variety in vocabulary, length of queries
and specificity of search terms); (ii) the search tactics used (e.g., how users start a session,
patterns of query reformulation, use of operators including Boolean, use of advanced search
features), and (iii) how users make relevance judgments (e.g., what items are selected as
relevant). It is obviously easier to capture some of these features from a log file than others
(e.g., search terms used vs. items judged as relevant). Although there is no agreed list of
features that can be used to characterise search patterns and any features could be selected,
it is worth asking whether a feature can be explained and how the feature can be captured
from a search log and quantified. Not all features identified may be suitable at characterising
search patterns and some element of feature selection could be applied.
To find patterns in the AOL log, we identified a number of descriptive features that
had been used in previous work when clustering search patterns into distinct behavioural
patterns:
1. Query length [17, 27]: the mean number of terms per query
2. Session duration [16, 27, 28]: the time from first action to session end
3. Number of queries [16, 17, 27, 28]: the number of keyword strings entered by the
user
4. Number of viewed hits / items retrieved [16, 27, 28]: the number of links clicked
by the user
5. Number of requested result pages [27]: the number of explicit page requests
6. Number of activities [27]: the sum of all the interactions with the system
These features are all reasonably straightforward to extract from a log such as AOL
158
Text Mining and Visualization: Case Studies Using Open-Source Tools
using Python or any other scripting language, and the output can be represented as a set
of feature vectors for each session, as shown in Figure 7.3.
FIGURE 7.3: A set of feature vectors from the AOL log.
There are of course a great many other features we could extract or derive from this log,
and the process of feature selection is something to which careful thought should be given.
But for now let us consider another key question: What are the most effective techniques
for revealing common patterns in such data? One such approach is to use an unsupervised
learning technique, such as clustering [16, 17, 27]. Categorising search patterns across users
into groups is useful in identifying common and distinct manifestations of search strategies
and behaviours. These strategies might help to improve search systems. For example, if a
common reoccurring pattern was a mis-spelt query followed by users correcting the query
then adopting automated spell-checking detection and correction might improve the search
experience. It is also commonplace for search providers to run online tests where search
patterns are analysed both with and without a new search feature and differences in behaviours captured and investigated to establish whether the new feature is likely to benefit
users or not.
7.3
Using Clustering to Find Patterns
In unsupervised learning there is no right answer in the sense that different outputs
can be obtained and no single output is necessarily more correct than the others (although
the patterns should be shown to be stable and repeatable, as discussed later). The outputs
should also make sense and be explainable in the light of what is known from previous
literature about search strategies and behaviours.
Instead, the utility of a particular clustering depends on the extent to which it provides
useful insight into the phenomenon of interest (e.g., user behaviour). Consequently, it is
prudent to adopt an exploratory strategy at the outset and try a range of approaches. A
machine learning platform such as Weka6 [12] provides a useful starting point as it offers an
extensive range of algorithms for unsupervised learning and a helpful (but somewhat less
polished) set of routines for result visualisation.
Let us start by taking a random sample of 100,000 sessions from the AOL log and then
applying feature scaling so that values are normalised over a given range (e.g., a standard
deviation of 1 and mean of zero). This ensures that features with particularly wide ranges in
the raw data do not disproportionately influence the clustering process [29] (pp. 321-352).
We can then use Weka to apply a clustering algorithm such as expectation maximization —7
6 http://www.cs.waikato.ac.nz/ml/weka/.
7 http://en.wikipedia.org/wiki/Expectation\%E2\%80\%93maximization_algorithm.
Mining Search Logs for Usage Patterns
159
and then visualise the results. The steps to perform this are as follows, with sample code
to perform these operations available from the website that accompanies this book.
7.3.1
Step 1: Prepare the Data
1. Extract the feature vectors (note that it may be prudent to experiment with a subset
of the AOL log to begin with, e.g., the first million lines or so):
> python parse14.py -v fullcollection.1000000
2. Select a sample from the output:
> python selectRandomVectors.py output.log -s 100000 >
100000.sample1
3. Prepend this with the Weka ARFF header8 :
> cat ARFFheader 100000.sample1 > 100000.sample1.arff
4. Apply feature scaling:
> python normalise_sessions.py 100000.sample1.arff >
100000.sample1.norm.arff
7.3.2
Step 2: Cluster Using Weka
1. Start Weka and choose the Experimenter interface (Figure 7.4)
FIGURE 7.4: The Weka GUI chooser.
2. Select the Preprocess tab and open the output of Step 1 in Weka (Figure 7.5)
3. Select the Cluster tab and choose the Expectation Maximization (EM) algorithm
(Figure 7.6)
4. Select the Ignore attributes button and make sure that any attributes other than the
six above are ignored.
5. Select the option to Store clusters for visualisation, and then run the EM algorithm
by selecting the Start button.
8 http://www.cs.waikato.ac.nz/ml/weka/arff.html
160
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 7.5: Loading the data into Weka.
FIGURE 7.6: Configuring the EM algorithm.
Mining Search Logs for Usage Patterns
7.3.3
161
Step 3: Visualise the Output
1. Right click on the last item in the Results list, and select the option to Visualize
cluster assignments.
2. In the pop-up window, set the x axis to queries and the y axis to clicks, and apply
a little jitter to smooth out the points (Figure 7.7). The output should be a chart
similar to that shown in Figure 7.8.
FIGURE 7.7 (See color insert.): Configuring the visualization.
Figure 7.8 reveals four clusters (shown by the different coloured points), projected onto a
two-dimensional space defined by the number of queries and the number of clicks. Note that
there is nothing special about the number four; if you take a different log or use different
features you will almost certainly get a different result. So it is important to perform multiple
iterations, to ensure that the patterns are stable (i.e. replicable across different samples from
the same population).
Let us focus for a moment on the image in Figure 7.8. What does it tell us? Well,
not a lot so far: the presence of 4 clusters may be significant, but the projection does not
deliver much insight into that. Instead, we need a visualisation that reveals the difference
in feature values between clusters. If we take the output of Step 2 above (which is stored
in the Clusterer Output window) and enter it into a suitable charting tool, we obtain an
output similar to that of Figure 7.9.
This shows how the mean values for the four clusters vary across the six features. But
now it is the data that are somewhat underwhelming: it suggests that we have a group
of users who are relatively active (i.e., demonstrate a lot of interactions), a group who do
relatively little, and two other groups in between. This is hardly the kind of insight that
will have market research professionals fearing for their jobs.
162
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 7.8 (See color insert.): 100,000 AOL sessions, plotted as queries vs. clicks.
FIGURE 7.9 (See color insert.): Four clusters based on six features.
Mining Search Logs for Usage Patterns
163
So at this point it may be prudent to review our choice of features; after all we have just
six so far that were selected in a relatively informal manner. To be maximally useful (e.g.,
indicative of a latent variable such as behaviour type), we would want them to be relatively
independent, i.e. uncorrelated with each other. But if we run a Pearson correlation across
the ones above we find the opposite: most of them are actually highly correlated, particularly
pairs such as All Actions and Duration. In this instance, these features may be “drowning
out” interesting patterns that we might otherwise see. Therefore, we can try dropping the
‘All Actions’ feature and adding in two new features, which are less likely to be correlated
with overall activity levels:
• Term use frequency [17]: the mean frequency of usage of each query term within
the session
• Query interval [17]: the mean time between query submissions within a session
To do this we simply repeat Steps 1-3, selecting the appropriate seven attributes in Step
2.4. Now the results seem a bit more interesting (Figure 7.10). We could interpret these
features as follows:
• Cluster 0: seems to be a relatively small group users who are highly engaged, in long
sessions but issuing queries that are diverse/heterogeneous (possibly changing topic
repeatedly)
• Cluster 1: seems to be a large group of users who engage in short sessions but often
repeat their query terms and do more paging than querying
• Cluster 2: seems to be a middle or general group, whose defining characteristics we’ll
know more about when we’ve extracted more features.
FIGURE 7.10 (See color insert.): Three clusters based on seven features.
164
7.4
Text Mining and Visualization: Case Studies Using Open-Source Tools
Replication and Validation
From this point there is evidently a number of different ways we could extend the analysis. But instead of just adding more features and applying ever more exotic visualisations
to new data sources, it is perhaps more prudent to take a moment to reflect on the process
itself and confirm that what we are doing really is valid and repeatable. One way to do this
is to replicate other published studies and compare the results. For example, Wolfram et al.
[17] used the following features:
1. Session size: the number of queries submitted during the session
2. Terms used per query: the average number of terms used per query over the session
3. Term popularity: the average frequency of usage of each query term within the
session when measured across all users over the dataset
4. Term use frequency: the average frequency of usage of each query term within the
session
5. Query interval: the average time between query submissions within a session
6. Pages viewed per query: the average number of page requests per query within a
session
Wolfram et al. [17] applied these to a number of search logs and showed (among other
things) evidence of four distinct behavioural patterns in a sample of approximately 65,000
Web search sessions from the Excite 2001 data set [30]. If we can replicate their results,
then we not only vindicate the specific conclusions they reached, but more importantly,
provide evidence that our approach is valid and scalable to new data sources and research
questions.
Note, however, that to calculate term popularity we need first to compute the mean
frequency of usage of each term across the entire data set. The steps to perform this are
as follows, with sample code to perform these operations available from the website that
accompanies this book.
1. Split all the query terms onto their own line:
> awk ’{ for (i=1; i<=NF; i++) print $i }’ aolquerylist.txt >
temp
2. Sort the output, merge and count the lines, then sort again in descending order:
> sort temp | uniq -c | sort rn > temp2
3. Reformat as csv to make subsequent processing easier:
> awk -f format_as_csv.awk temp2 > aolquerylist.csv
Once this is complete, we can repeat Steps 1-3 in Section 7.3, selecting the appropriate six
attributes in Step 2.4. Note that running the EM algorithm on large data samples can take
some time to complete, so it may be prudent to experiment with smaller samples first (e.g.,
by selecting a sample of 10,000 lines in Step 1.2. The output is shown in Figure 7.11. This
Mining Search Logs for Usage Patterns
165
time, we find four clusters as Wolfram et al. [17], but the patterns are very different. (Note
that display order of the features is changed here to facilitate comparison with Wolfram et
al.’s results, and their labels have been simplified). But crucially, the results are not even
replicable within themselves: a further three samples of 10,000 sessions produces widely
different outcomes (5, 6, and 7 clusters, respectively). Even increasing the sample size to
100,000 seems to make little difference, with 7, 13, 6, and 6 clusters produced on each
iteration (despite the suggestion in Wolfram et al.’s paper that subsets of 50k to 64k sessions
should produce stable clusters).
FIGURE 7.11 (See color insert.): Applying EM using Wolfram et al.’s 6 features to 10,000
sessions from AOL.
So why are we seeing such different results? One interpretation may be of course that
these insights are indeed an authentic reflection of changes in user behaviour due to differences in context (e.g., a different search engine, time period, demographic)
But let’s pause for a moment and examine the pattern in more detail. There is something
very odd happening with term popularity now: we see a small cluster (just 3% of the sessions)
where this feature seems to be something of an outlier, compressing the remaining traces
into a narrow band. Indeed, the phenomenon becomes even more pronounced when we take
a sample of 100,000 sessions. (To do this, we simply repeat Steps 1-3, passing 100,000 as a
parameter to the sampling process in Step 1.2). The output is shown in Figure 7.12.
Perhaps this is an artifact of the clustering algorithm? We could try XMeans9 in Step 2.3
instead (which is a variant of kMeans,10 where the value for k is determined automatically).
For our data sample, XMeans finds a local optimum at k = 10, so the number of clusters
is now different. But the overall pattern, with a small cluster (1% of sessions) representing
outlier values for term popularity, is again clearly visible (Figure 7.13).
So something else must be at play. It turns out that there is indeed an artifact of the
data which is causing this: there is a small number of sessions that contains just a single
query, consisting solely of the character ‘-’. Precisely why they are present in the log is
a matter for speculation: they may have been the default query in some popular search
application, or a side-effect of some automated service or API. But sessions like these, along
9 http://weka.sourceforge.net/doc.packages/XMeans/weka/clusterers/XMeans.html.
10 http://en.wikipedia.org/wiki/K-means_clustering.
166
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 7.12 (See color insert.): Applying EM using Wolfram et al.’s 6 features to 100,000
sessions from AOL.
FIGURE 7.13 (See color insert.): Applying XMeans (k <= 10) and Wolfram et al.’s 6
features to 100,000 sessions from AOL.
Mining Search Logs for Usage Patterns
167
with other robot-generated sessions, are not generally helpful when trying to understand
human behavioural patterns. Instead, they are best removed prior to analysis. Of course,
there are no 100% reliable criteria for differentiating robot traffic from human, and what
should be removed is a matter for judgement, often on a case-by-case basis [31]. In this
case, including these single-character queries appears to be counter productive. A simple
shell script or single line of Awk can do this for us:
> awk ’!/\t-\t/’ fullcollection.1000000 >
fullcollection.1000000.filtered
With a new sample of 100,000 sessions excluding these outlier queries, we see Steps 1-3
using EM produce a very different output (Figure 7.14).
FIGURE 7.14 (See color insert.): Applying EM and Wolfram et al.’s 6 features to 100,000
filtered sessions from AOL.
This pattern is much more stable, with four iterations producing 7, 7, 7, and 9 clusters
respectively. At this point we can start to speculate on what these patterns may be telling
us. For example:
• Cluster 6: appears to be a group of users that engage in longer sessions, with many
queries and many page views (clicks) but few repeating terms
• Cluster 4: appears to be a smaller group who seem to specialise in relatively long
but popular queries, also with few repeating terms
• Cluster 3: appears to be a relatively large group who make greater use of repeated
terms but are otherwise relatively unengaged (with shorter sessions and fewer page
views)
And so on. Evidently, the previous patterns are somewhat hard to interpret due to the
larger number of clusters and lines on the chart. What would happen if we tried to determine
the optimum number ourselves, rather than letting XMeans find one for us? One way of
investigating this is to specify different values for k a priori, and see how the within-cluster
sum of squared errors (which is calculated by Weka as part of its output) varies on each
168
Text Mining and Visualization: Case Studies Using Open-Source Tools
iteration. To do this, simply repeat Step 2 nine times, selecting XMeans in Step 2.3 and
varying k from 2 to 10 on each iteration. When we plot the output as a chart, we see a
function similar to that shown in Figure 7.15.
FIGURE 7.15: Sum of squared errors by k for 100,000 filtered sessions from AOL.
As we can see, there is an elbow around k = 4 and another around k = 7. This implies
that either of these two values may be good choices for a local optimum. We have already
seen the output for k = 7 (which is the optimum that XMeans found), so now we repeat
Step 2 with kMeans and k = 4. The output is shown in Figure 7.16.
This time the groups are somewhat easier to differentiate. For example, we might suggest
that:
• Cluster 3: represents a baseline or generic cluster of users who hover around the
average for all measures
• Cluster 4: represents a relatively large group of users who engage in longer sessions
(with more queries and page views) but are diverse in their interests, with few repeated
terms
• Cluster 1: represents a smaller group who are the converse to cluster 4, engaging in
shorter sessions but with more repeated terms
• Cluster 2: represents a tiny group (2%) of users who are similar to cluster 1 but
focus on highly popular queries
7.5
Summary
Evidently, there are other ways we could analyse these data, and there are other ways
we could interpret the output. What has been presented is really just a starting point for
Mining Search Logs for Usage Patterns
169
FIGURE 7.16 (See color insert.): Applying kMeans (k = 4) and Wolfram et al.’s 6 features
to 100,000 sessions from AOL.
exploration. So for now, let us draw some of the threads together and review what has been
covered.
• Start with the end in mind: Mining search logs for usage patterns is inherently an
exploratory activity, so it makes sense to consider a number of different approaches.
In each case, the utility of the results should be judged by the extent to which they
provide useful insight into the phenomenon of interest (e.g., user behaviour).
• Results may not generalise: it is possible that the insights gained from one search
log may not generalise to logs from other systems as users’ patterns of interaction
will be influenced by many factors including differences in system content, functionalities, interfaces, and user groups [11]. However, analysis of search logs can highlight
key aspects of user–system interaction, thereby potentially “enlighten[ing] interface
development, and devising the information architecture for content collections” [9] (p.
407).
• There is no right answer: As in many investigations of naturalistic phenomena,
there is a tendency to look for patterns that align with our expectations. However,
those expectations themselves are a subjective, social construct. The fact that we can
produce multiple interpretations of the same data underlines the need for a common
perspective when comparing patterns in search logs and to apply recognised models
of information-seeking behaviour when interpreting the outputs.
• Look for orthogonal features: The hypothesis underlying much search log analysis
of this type is that behaviour type is a latent variable whose influence can be measured
indirectly via features such as those above. To be maximally useful these features
should be highly correlated with the behaviours we wish to observe but relatively
independent of each other.
170
Text Mining and Visualization: Case Studies Using Open-Source Tools
• Apply Occam’s razor: It is tempting to select features based on whatever a particular data source offers and to include as many as possible in the learning process.
But not all are equally useful, and some can indeed drown out the influence of more
important signals. So rather than starting from what the data can offer, identify the
information-seeking behaviours you’d like to explore, and try to find the features that
most closely align with them.
• Replicate to validate: As researchers, our instincts are to explore the unknown, to
solve the unsolvable, and to favour novelty over repetition. But sometimes it benefits
us to focus on the replication: by applying new techniques to old data, we validate
our methodology and build a more reliable baseline for future experimental work.
Acknowledgments. Work partially funded by a Google Research Award project: “Developing a Taxonomy of Search Sessions”.
Bibliography
[1] F. Silvestri. Mining query logs: Turning search usage data into knowledge. Found.
Trends Inf. Retr., 4(1–2):1–174, January 2010.
[2] S.J Lin and N. Belkin. Validation of a model of information seeking over multiple
search sessions. J. Am. Soc. Inf. Sci. Technol., 56(4):393–415, February 2005.
[3] M. Kellar, C. Watters, and M. Shepherd. A field study characterizing web-based
information-seeking tasks. J. Am. Soc. Inf. Sci. Technol., 58(7):999–1018, May 2007.
[4] Y. Li and N. J. Belkin. A faceted approach to conceptualizing tasks in information
seeking. Inf. Process. Manage., 44(6):1822–1837, November 2008.
[5] R. Jones and K. L. Klinkner. Beyond the session timeout: Automatic hierarchical
segmentation of search topics in query logs. In Proceedings of the 17th ACM Conference
on Information and Knowledge Management, CIKM ’08, pages 699–708, New York,
NY, USA, 2008. ACM.
[6] A. Spink, M. Park, B. J. Jansen, and J. Pedersen. Multitasking during Web search
sessions. Inf. Process. Manage., 42(1):264–275, January 2006.
[7] A. Broder. A taxonomy of web search. SIGIR Forum, 36(2):3–10, September 2002.
[8] G. Marchionini. Information-seeking strategies of novices using a full-text electronic
encyclopedia. J. Am. Soc. Inf. Sci., 40(1):54–66, January 1989.
[9] B. J. Jansen. Search log analysis: What it is, what’s been done, how to do it. Library
& Information Science Research, 28(3):407 – 432, 2006.
[10] C. Grimes, D. Tang, and D. M. Russell. Query logs alone are not enough. In Proceedings of the WWW 2007 Workshop on Query Logs Analysis: Social and Technological
Challenges, 2007.
[11] M. Agosti, F. Crivellari, and G.M. Di Nunzio. Web log analysis: a review of a decade of
studies about information acquisition, inspection and interpretation of user interaction.
Data Mining and Knowledge Discovery, 24(3):663–696, 2012.
Mining Search Logs for Usage Patterns
171
[12] M. Hall, Eibe F., G. Holmes, B. Pfahringer, P. Reutemann, and I.H. Witten. The weka
data mining software: An update. SIGKDD Explor. Newsl., 11(1):10–18, November
2009.
[13] G. Pass, A. Chowdhury, and C. Torgeson. A picture of search. In Proceedings of the 1st
International Conference on Scalable Information Systems, InfoScale ’06, New York,
NY, USA, 2006. ACM.
[14] B. Liu. Web data mining exploring hyperlinks, contents, and usage data. SpringerVerlag New York, Inc., Secaucus, NJ, USA, 2011.
[15] D. Gayo-Avello. A survey on session detection methods in query logs and a proposal
for future evaluation. Inf. Sci., 179(12):1822–1843, May 2009.
[16] I. Weber and A. Jaimes. Who uses Web search for what: and how. In Proceedings of
the Fourth ACM International Conference on Web Search and Data Mining, WSDM
’11, pages 15–24, New York, NY, 2011. ACM.
[17] D. Wolfram, P. Wang, and J. Zhang. Identifying web search session patterns using
cluster analysis: A comparison of three search environments. Journal of the American
Society for Information Science and Technology, 60(5):896–910, 2009.
[18] D. He and A. Göker. Detecting session boundaries from web user logs. In Proceedings of
the BCS-IRSG 22nd annual colloquium on information retrieval research, pages 57–66,
2000.
[19] Y. Wang, X. Huang, and R. W. White. Characterizing and supporting cross-device
search tasks. In Proceedings of the Sixth ACM International Conference on Web Search
and Data Mining, WSDM ’13, pages 707–716, New York, NY, 2013. ACM.
[20] H. Wang, Y. Song, M.-W. Chang, X. He, A. Hassan, and Ryen W. White. Modeling action-level satisfaction for search task satisfaction prediction. In Proceedings of
the 37th International ACM SIGIR Conference on Research &#38; Development in
Information Retrieval, SIGIR ’14, pages 123–132, New York, NY, 2014. ACM.
[21] G. Dupret and C. Liao. A model to estimate intrinsic document relevance from the
clickthrough logs of a web search engine. In Proceedings of the Third ACM International
Conference on Web Search and Data Mining, WSDM ’10, pages 181–190, New York,
NY, 2010. ACM.
[22] M. Whittle, B. Eaglestone, N. Ford, V. J. Gillet, and A. Madden. Data mining of
search engine logs. J. Am. Soc. Inf. Sci. Technol., 58(14):2382–2400, December 2007.
[23] J. Huang and E. N. Efthimiadis. Analyzing and evaluating query reformulation strategies in web search logs. In Proceedings of the 18th ACM Conference on Information and
Knowledge Management, CIKM ’09, pages 77–86, New York, NY, USA, 2009. ACM.
[24] B. J. Jansen, M. Zhang, and A. Spink. Patterns and transitions of query reformulation
during web searching. International Journal of Web Information Systems, 3(4):328–
340, 2007.
[25] S. Y. Rieh and H. Xie. Analysis of multiple query reformulations on the web: the
interactive information retrieval context. Inf. Process. Manage., 42(3):751–768, May
2006.
172
Text Mining and Visualization: Case Studies Using Open-Source Tools
[26] P. Vakkari. A theory of the task-based information retrieval process: a summary and
generalisation of a longitudinal study. Journal of Documentation, 57(1):44–60, 2001.
[27] D. Stenmark. Identifying clusters of user behavior in intranet search engine log files.
Journal of the American Society for Information Science and Technology, 59(14):2232–
2243, 2008.
[28] H.-M. Chen and M. D. Cooper. Using clustering techniques to detect usage patterns
in a web-based information system. Journal of the American Society for Information
Science and Technology, 52(11):888–904, 2001.
[29] O. Maimon and L. Rokach. Data Mining and Knowledge Discovery Handbook. Series
in Solid-State Sciences. Springer, 2010.
[30] A. Spink, B. J. Jansen, D. Wolfram, and T. Saracevic. From e-sex to e-commerce: Web
search changes. Computer, 35(3):107–109, March 2002.
[31] Y. Zhang and A. Moffat. Some observations on user search behavior. In Proceedings
of the 11th Australasian Document Computing Symposium. Brisbane, Australia, 2006.
Chapter 8
Temporally Aware Online News Mining and
Visualization with Python
Kyle Goslin
Department of Computer Science, College of Computing Technology, Dublin, Ireland
8.1
8.2
8.3
8.4
8.5
8.6
8.7
8.8
8.9
8.10
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.1
Section Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.1.2
Conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
What You Need . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.1
Local HTTP Host . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.2
Python Interpreter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.3
PIP Installer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.4
lxml and Scrapy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.5
SigmaJS Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.6
News Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.2.7
Python Code Text Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Crawling and Scraping — What Is the Difference? . . . . . . . . . . . . . . . . . . . . . . . . . . .
What Is Time? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Analyzing Input Sources . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.5.1
Parameters for Webpages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.5.2
Processing Sections of Webpages vs. Processing Full Webpages . . . . . .
Scraping the Web . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.1
Pseudocode for the Crawler . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.2
Creating a Basic Scraping Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.3
Running a Scraping Tool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.4
News Story Object . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.5
Custom Parsing Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.6
Processing an Individual News Story . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.6.7
Python Imports and Global Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Generating a Visual Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.7.1
Generating JSON Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.7.2
Writing a JSON File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Viewing the Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Additional Concerns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.9.1
Authentication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
8.9.2
Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
174
174
174
174
175
175
176
177
177
178
178
179
179
180
182
182
183
183
184
184
185
185
186
187
188
189
191
193
194
196
196
197
197
173
174
8.1
Text Mining and Visualization: Case Studies Using Open-Source Tools
Introduction
The quantity of information rich data on the Web has been increasing year on year
providing a further burden on users to decipher what is relevant to them. The use of
aggregation services to define what is considered relevant and useful have become the norm
during the Web browsing process. A good example of an aggregation service pulling content
from a wide number of sources is Google News.
However, the quantity of information being presented to the user is still quite vast and
also time sensitive, for instance, news stories changing every few minutes on Google News.
The quantity of data creates the need for further processing of aggregated data to present
the content to the user in a easily consumable format.
News stories that have been aggregated often come clearly marked with title text, description text, date, and timestamps as standard fields during the publication process. One
of the most useful fields is the timestamp associated with data, which, although simple, can
be an effective measure of document relevance at the current moment in time.
This chapter aims to provide a detailed overview of the development of a temporally
aware news mining tool with visualization using free and open-source software. By providing
a temporal window on the data an additional level of depth can be achieved during the
selection process allowing a quick decision-making process before retaining resources for
further processing.
The process of deciding text value is not described in this chapter but can be found in
other chapters of this book. All of the code samples in this chapter can be downloaded from
the official text mining book webpage at http://www.text-mining-book.com.
8.1.1
Section Outline
The aim of this chapter is to provide an example of the scraping and selection process
needed for extracting data from a news aggregation website. Once these data have been
extracted, they can then be assessed for relevance and visualized using free and open-source
visualization tools.
Section 8.1 provides a brief introduction to the area of scraping a visualization. Section 8.2 outlines the tools that are needed for this work. Section 8.3 outlines the difference
between crawling and scraping. Section 8.4 outlines the concept of time and issues that can
arise. Section 8.5 outlines the process of analyzing input sources. Section 8.6 outlines the
process of scraping a webpage. Section 8.7 outlines the process of generating a visual representation with JSON. Section 8.8 outlines the viewing process using SigmaJS. Section 8.9
outlines additional concerns during the scraping process. Section 8.10 concludes this chapter
with a summary of what was covered in this chapter.
8.1.2
Conventions
In this chapter, commands that should be typed out in the operating system command
line or terminal will be proceeded with the following syntax:
>>
The command that must be typed will follow these two characters do not type the two
>> characters.
Temporally Aware Online News Mining and Visualization with Python
8.2
175
What You Need
Before starting, this section outlines each of the tools that are needed to develop the
application outlined in this chapter. The examples outlined in this chapter are performed on
a Microsoft Windows-based operating system. However, due to the cross-platform nature
of Python [1] and JavaScript, each of the steps outlined here can be easily ported to the
operating system of your choice.
To get the most out of this chapter, having a basic knowledge of the Python programming
language and how to develop applications locally will help. If you do not have a knowledge
about the Python programming language, a wealth of knowledge is available at Python.org.
8.2.1
Local HTTP Host
For the visualization process we will be using a JavaScript library for rendering. To avoid
any issues during this process, we first need to download a local HTTP server where we can
locally host our visualization code. During this tutorial we will be using XAMPP [2] as it
provides all of the functionality that we need with a simple user control panel.
You can download XAMPP from https://www.apachefriends.org/index.
html and clicking on the download link. This installer is rather large and may take some
time to download. During the visualization steps, we will be looking into the details of
using this tool. XAMPP by default will attempt to run the Web server on port 80. Ensure
that you do not have any other services running on this port. If you are experiencing any
issues during this process whereby the service will not start, ensure that the firewall on
your system is turned off and that you are running the sever in administrator mode. This
can be done by right click on the control panel icon and clicking Run as Administrator. If
issues with port 80 persist, you can easily change the port number of the HTTP server to
an alternative port using the XAMPP control panel.
8.2.2
Python Interpreter
The most important tool we will be using is the Python interpreter. The Python interpreter allows you to run Python program files with a .py extension. This interpreter can
be downloaded from Python.org. Although a number of different versions of the interpreter
exist such as the later version 3.0, to aid backwards compatibility we will be working with
the stable Python 2.7 for this project chapter.
It is important that we can access our Python interpreter and Scripts folder from anywhere on our system. To do this we must add Python into our system path variable. On
Microsoft Windows-based systems, this is done by editing the environment variables on our
system to include the path to where Python was installed. This can done by:
1. Entering into the Windows control panel
2. Clicking on the System icon
3. Once the system menu has opened, on the left-hand side we can then enter into the
Advanced System Settings. The system properties menu will then open for us.
4. From here we can click the Environment Variables button.
5. In the collection of System Variables that are available, locate the PATH variable and
click the Edit button.
176
Text Mining and Visualization: Case Studies Using Open-Source Tools
6. Insert a new entry at the end of the value for this variable (not replacing the current
string!); The values that must be appended to this string are:
;C:\Python27;C:\Python27\Scripts This should reflect the path where your Python
installation is located on your system. You may need to alter this to reflect the drive
on your system where Python is installed.
7. To finish click Ok on each of the three open windows.
It is important to ensure that this string is proceeded by a semicolon, if this is omitted then
your path variable will not be updated correctly.
If you are working in a Linux environment, Python is often installed as a default as
part of your installation. To check what version is installed on your installation, from the
command line type the following command:
>> python -V
It is important to remember that Python is space sensitive. Correct indentation must
be added to your code to ensure it will run correctly. This can be difficult to users who have
no experience with Python code as their code may look correct but will not correctly be
interpreted.
8.2.3
PIP Installer
When working with a number of different libraries, the process of downloading each
individual library and keeping track of where they are located can become a monotonous
task. As a solution to this process, a tool was developed called PIP [3], which is used for
downloading and installing different Python libraries quickly and easily on your system.
To install PIP, we first need to download the get-pip.py script from https://
bootstrap.pypa.io/get-pip.py. This can be done by opening the file in a Web
browser and then right-clicking the file in the Web browser and clicking Save As. This
file should be saved to your desktop for easy access in later steps.
Once you have download the file, you need to run the PIP installer script. This is done
by opening up the Windows command line for your machine. To do this click the Start
button and then type the command cmd into the search box and hit the enter key. The
command line will open at your home directory shown in Figure 8.1.
FIGURE 8.1: Windows command prompt.
Once the command line has been opened, we need to navigate to the desktop by typing
in the following command:
Temporally Aware Online News Mining and Visualization with Python
177
>> cd Desktop
After you have moved to the desktop, we use the following command to run the get-pip.py
installer script.
>> python get-pip.py
This will then begin the installation process for the PIP installer tool. A message will appear
in the console telling you the current progress of your download and installation. At the
end of this process a message will be presented telling you if your installation succeeded or
failed. If you get a message saying that Python cannot be found, this may be an indication
that you have not installed Python or have not updated the path variable. If this happens
check to see if you have installed Python and added the Python directory to your path
variable by reviewing the steps in Section 8.2.2.
8.2.4
lxml and Scrapy
Scrapy [4] is a Python-based application that allows us to develop a Web scraping
application without the need to focus on the finer details of the scraping process. We will
use PIP which we have previously installed to install Scrapy. However, to use Scrapy, we
must first download and install the lxml [5] parser.
To use the PIP that we previously installed, we must navigate into the folder where PIP
is located on your local machine. PIP can be found inside the Scripts folder inside of the
directory where Python has been installed. To navigate into this folder, type the following
command into the command line:
>> cd C:\Python27\Scripts
Once we are inside the Scripts folder, we can the run the following command to install
lxml using PIP:
>> pip install lxml
Like the previous process, a message will be printed to the console telling you if your
installation process was successful or not. After we have installed lxml, we can then begin
the process of installing Scrapy. From the directory we are in type the following command
into the command line:
>> pip install scrapy
After this process has completed you should have both lxml and Scrapy successfully
installed on your local machine.
8.2.5
SigmaJS Visualization
The application that we are building will contain visualizations of the data we collect.
To do this we will need to use an additional visualization package. Although Python is used
for a variety of different text scraping and processing tasks, we will not be using Python for
the visualization of the data. We will opt for a Web-based visualization tool called SigmaJS
[6] which is a light-weight open-source JavaScript based visualization library. To get started:
1. Downloaded SigmaJS from http://sigmajs.org by clicking on the download button. You will then be directed to the download page.
178
Text Mining and Visualization: Case Studies Using Open-Source Tools
2. On this page, under the release section, click and download the .zip file containing
SigmaJS source code. This will begin a download of a single .zip file.
3. Once this file has downloaded, you must then right-click and extract the contents of
the .zip file to your Desktop.
4. Open the extracted folder and view the contents.
Figure 8.2 shows the contents of this folder. Right-click this folder and rename this folder
sigma. In the rest of this chapter, we will treat the sigma folder as the entire visualization
application.
FIGURE 8.2: Contents of the SigmaJS folder.
Inside this folder you will see an examples folder. It contains a number of different
samples of SigmaJS in action. The one we are most interested in is the load-externaljson.html file. This gives a fully working sample of SigmaJS pulling in the JSON content
of a file for visualization. The file which it is pulling in will later be generated by our
application.
8.2.6
News Source
The focus of this chapter is temporally aware news mining and visualization, we will
need to have a site of interest to us that we will perform this mining process on. For this
chapter, we will be working with the Google News aggregation service. Each story comes
with the attributes we are interested in for this chapter such as the date, time, title, and
description text for each story.
The process followed in this chapter is not specific to Google News and can be altered
to work with any other online news source.
8.2.7
Python Code Text Editor
Although it can be tempting to get involved in using an integrated development environment, for this project it is not needed. Windows Notepad can provide all of the features you
Temporally Aware Online News Mining and Visualization with Python
179
need. However, if you would like some additional features, a simple lightweight editor such
as Notepad++ (https://notepad-plus-plus.org) can offer syntax highlighting to
make the code editing process a little bit easier.
If you are working in a Linux environment, gedit or vi will perform just as well for each
of the tasks.
8.3
Crawling and Scraping — What Is the Difference?
Although the terms crawling and scraping are often found together, they are two different
processes. The process of crawling is identifying links on a website and following through
each webpage looking for subsequent links to crawl. The tools that perform this task are
often called spiders, due to their crawling nature. Scraping, however, is performed when a
page of interest has been found. The content on the page is scraped and stored in a local
repository for further processing.
When a website of interest has been found it may contain a collection of additional
links. It is possible, due to a combination of links on the website that the crawling tool
may unintentionally revisit the same link a number of times causing additional wasted
resources, or loop on a collection of pages. An example of this is a naı̈ve crawler that may
keep revisiting the home link on a website, starting the same crawling process over and
over again. To prevent this from happening, it is often the case that a log of visited pages
is retained and utilised. Before any pages are scraped the list of links that have previously
been scraped is assessed for similarity.
The concept of depth is often cited that refers to how many links should be followed
on a webpage. A max depth M D is often set for a given webpage, P . When considering
a webpage as a tree-like structure with the index page as the root of the tree, each link
followed is considered an additional level deep or D + 1. The value for D should always
remain ≤ than M D.
Scrapy by default does not follow or gather any links. As we will be scraping the index
page of Google News, we do not need to set up any additional settings for crawling or
following links.
8.4
What Is Time?
When working with data, having a timestamp attached provides a great insight during
the relevancy decision-making process. Time, however, is a complex concept when considering a worldwide audience and variations can exist. Depending on the country of origin
the layout of the timestamp can also cause confusion. An example of this is the Americanstyle datestamp that places the month first, followed by the day and year in comparison
to the European timestamp, which places the day first followed by the month and year. A
number of other variations of these timestamps exist, such as spelling out the month (e.g.,
October-2014) compared with using numeric form (e.g., 10-2014). Other variations can exist
in the format such as the use of hyphens, slashes, or dots to break each component of the
timestamp.
As a solution to date-related problems, typically the time reference Epoch is often used
180
Text Mining and Visualization: Case Studies Using Open-Source Tools
in the Unix world. Epoch is the 1st of January, 1970. From this date the number of seconds
that have lapsed since then can easily be calculated and stored e.g. 1408975329. This allows
all timestamps to be saved in a unambiguous format without complex formatting issues.
These seconds based timestamps can then be easily converted into the desired format before
presenting the timestamp to the user.
An important element of working with time-sensitive data is the process of specifying
a time window where data can be considered relevant. We will logically break down time
windows into discrete individual periods of time. Each individual window is then given a
unique ID. An example of these windows are shown in Table 8.1.
TABLE 8.1: Time windows.
Window
1
2
3
4
Days
Today
Last 2 Days
Last 3 Days
Last 7 Days
Hours
All posts
All posts
All posts
All posts
in
in
in
in
the
the
the
the
last
last
last
last
12 hours
48 hours
72 hours
168 hours
We are using hours as the metric for time as breaking news is related in terms of hours
since the story was published. Older news stories (greater than 1 day old) do not come with
an hours based stamp but are given in terms of a traditional time/date stamp. An example
of this is when a news story on the front page of a news website becomes older than 24
hours it reverts to the complete date format (e.g., 1/11/14). We will need to apply further
transformations to convert the date stamp into a useable format for future processing. This
process is described in later sections.
8.5
Analyzing Input Sources
To get the best quality results from your scraping process it is important that the data
sources which you use are of high quality. Individual news sources can be used to gather
content such as CNN.com or a news aggregator such as Google News. Using a single news
source can have advantages as a single format of webpage is known so all implementations
can follow strict scraping rules.
We will be looking at the webpage as a single entity for this project. When viewing a
HTML page, we can note that a common structure is always followed:
1 <html>
2 <head>
3
<title> </title>
4 </head>
5
6 <body>
7 </body>
8 </html>
HTML documents can contain a mixture of HTML and CSS. The page that we will
be looking at is the Google News home page. This page consists of a standard HTML
Temporally Aware Online News Mining and Visualization with Python
181
structure, but also inside of the simple HTML document a main parent div titled contentpane-container is found. The structure of this div is as follows:
1 <div class="content-pane-container">
2 </div>
Inside this div, a number of child divs are added for each news story. Each of these contain
a title describing the news story, description text for the story, a timestamp outlining when
the story was published, and a URL address pointing to the news story. What follows is a
sample child div added to the container div:
1 <div class="blended-wrapper esc-wrapper">
2 </div>
The high-level view of the logical structure of a Google news page would look like the
following after news stories have been added:
1
2
3
4
5
6
7
8
9
10
11
12
13
<html>
<head>
<title> </title>
</head>
<body>
<div class="content-pane-container">
<div class="blended-wrapper esc-wrapper"> </div>
<div class="blended-wrapper esc-wrapper"> </div>
<div class="blended-wrapper esc-wrapper"> </div>
</div>
</body>
</html>
In this example you can see that inside the body a single content-pane-container exists,
and inside this multiple different divs with the class attribute set to blended-wrapper escwrapper can be seen. For each additional news story on the page, an additional blendedwrapper esc-wrapper is then added to this parent div. Inside the div with class type blendedwrapper esc-wrapper, the four elements are added that are of interest to us are described as
follows:
1. The div with class type esc-lead-snippet-wrapper contains information describing the
news story.
2. The Span of class type titletext is a short title for the news story.
3. A link, shown here as a <a target= blank .. is the link element which contains the
URL to the actual news story being shown.
4. A final span tag with class type al-attribution-timestamp contains a text based timestamp outlining how long ago the news story had been published. This can contain
either the number of minute(s) or hour(s).
A complete example of the contents of a blended-wrapper esc-wrapper for an individual
news story is shown as follows with header and trailing HTML removed:
1 <span class="titletext">2014 World Series Game 6 Preview: Jake
2 Peavy vs. Yordano Ventura</span>
3 ...
4
182
Text Mining and Visualization: Case Studies Using Open-Source Tools
5 <a target="_blank" class="article usg-AFQjCNHszuSN7CVG9y5hIa7zlQvEU3t-aQ
sig2-5Vjs9uBMiqZO9p3HXLiX7Q did--2408506129546916749 _tracked" href="
http://news.google.com/news/url?sr=1&amp;ct2=us%2F0_0_s_3_1_a&amp;sa=
t&amp;usg=AFQ......" url="http://www.cbssports.com/mlb/eye-onbaseball/24772127/world-series-game-6-preview-jake-peavy-vs-yordanoventura" rel="noreferrer" id="MAA4AEgDUABgAWoCdXM" ssid="s"
saprocessedanchor="true" originalhref="http://www.cbssports.com/mlb/
eye-on-baseball/24772127/world-series-game-6-preview-jake-peavy-vsyordano-ventura"><span class="titletext">2014 World Series Game 6
Preview: Jake Peavy vs. Yordano Ventura</span></a>
6
7
8 ...
9 <span class="al-attribution-timestamp">18 minutes ago</span>
10
11 ...
12 <div class="esc-lead-snippet-wrapper">The Giants and Royals are set to
resume the World Series in Kansas City’s Kauffman Stadium on Tuesday
night. San Francisco leads the series three games to two and is one
win away from their third title in the last five seasons.</div>
13
14 ...
8.5.1
Parameters for Webpages
Most websites offer the ability generally to add additional parameters in the URL. This
allows additional user specific changes to be made to the request, such as the language,
geographic region of interest or number of different resources to be returned to the caller.
An example of this for the Google News website is a parameter for setting the Edition
(shown as ned parameter).
In the following snippet, the ned parameter is being set to en ie. If we wanted to get
the American edition of Google News this parameter could be changed to us.
1 ned=en_ie
This parameter is passed in via the request URL:
1 https://news.google.com/?edchanged=1&ned=us
When crawling through webpages it is often the case that not all of the pages are directly
linked to but can be accessed by passing the correct parameter through a URL variable.
Crawlers often create variations of these parameters (e.g., where a story ID is set to the
numbers between 1 and 1,000 with the chance of discovering a webpage that has not been
directly linked to).
8.5.2
Processing Sections of Webpages vs. Processing Full Webpages
Webpages have a number of different elements such as HTML, CSS and JavaScript which
are not useful to us during our processing as we are only interested in the text content. In
addition to this, content may be added to the pages such as advertisements which are large
sections of text that are not relevant to the current task. Before any further processing is
done, we must be sure to remove these from the document.
To help us remove elements of content that are not useful to us, we can use the Selector
class which comes with Scrapy. The Selector allows us to select a div inside the HTML
Temporally Aware Online News Mining and Visualization with Python
183
document we are currently working with. This is done by identifying the class name of the
div we would like to select using XPath notation. If you are comfortable with other selection
tools such as BeautifulSoup or lxml, you can use this instead of the Selector class offered
by Scrapy.
In this particular example, we know the name of the class that we would like to remove
from the document. Generally large companies such as Google AdWords or other advertisement networks use a similar set of IDs for the divs where their advertisements are shown,
allowing easy identification and removal.
After we have selected the div we would like to remove, we can then use the Python
function shown here as body.replace, to remove the content that we have just selected from
the main document. In the following example, a sample selection by div name and removal
using body.replace is being performed on a sample HTML document.
After we have removed the content, we are then printing out the cleaned document out
using the Python print function.
1 from scrapy.selector import HtmlXPathSelector
2 from scrapy.selector import Selector
3
4 body = ’<html><body><div class="advert">This is an advert!</div><span>
good</span></body></html>’
5 selectElement = Selector(text=body).xpath(’//div[@class="advert"]’).
extract()
6
7 str = ’’.join(selectElement)
8
9 print str
10 cleaned = body.replace(str, ’’)
11
12 print cleaned
In addition to removing the advertisements, we may also want to filter out other pieces of
data that are not relevant from the document (e.g., repeated copyright notices and header
information). Depending on the website that you are working with, you may manually
identify the elements that you would like to remove from the document and remove these
sections before automating the task of content removal from the document.
8.6
Scraping the Web
The process of scraping content from the Web can be performed in two different ways.
The first of these is developing a custom Web scraping utility which makes a direct connection using a Python library such as urllib. After the connection has been made the webpage
is then downloaded and parsed manually. The second of these approaches is to use an existing library for processing. There are a number of benefits to using a predefined library,
such as handling of errors, authentication, session data management, and other smaller aspects of the scraping process that can become time-consuming when a custom scraping tool
has been developed. For these time saving benefits, we will be working with the premade
Python-based Web scraping tool Scrapy.
184
8.6.1
Text Mining and Visualization: Case Studies Using Open-Source Tools
Pseudocode for the Crawler
Because the process of scraping, filtering, and processing data is complex, before we
begin creating any code we will first step back and consider what the pseudocode for this
process will look like. Now we outline a rough guide to what steps are involved in our initial
scraping process:
1
2
3
4
5
6
7
8
Using Scrapy Connect to the URL and access the page HTML
Search the document we receive for blended-wrapper esc-wrapper div tags
using XPath notation
Process each individual news story div
For each story div tag block search for the timestamp and titletext
Analyze the timestamps to see if the text is within our current range
of interest
IF timestamp in our time window THEN
Store timestamp, title text, unique id for story as a new
NewStory object
Add story object to local storage list
8.6.2
Creating a Basic Scraping Tool
In the following sections we will begin the process of creating our custom scraping tool.
To begin, we will start the process of creating a new Python script file and placing it inside
the Python27\Scripts folder on your local machine. Using your favourite text editor, save a
blank file titled simplecrawler.py inside of your Scripts folder. From this point on, we will
assume your default directory for Python is in the C: drive of your machine. The folder
where we will place our code can be found at
C: \Python27\Scripts
After creating the empty simplecrawler.py file, we will then begin by putting in the class
for the following SimpleCrawler script:
1 class SimpleCrawler(Spider):
2
name, start_urls = ’googlenews’, [’http://news.google.com’]
3
4
def parse(self, response):
5
print ’Inside parse function’
The start url shown in the example is the website address where we will begin the
scraping process. This is not a recursive process and the scraper will only scrape the current
page that we have provided and not follow any of the links.
The parse function that can be seen in the previous example receives a response object
which contains the content from the webpage that we have just scraped. All of our processing
will be done on this response object. When the scraper runs it will trigger the parse function
to perform some tasks on the content that has been scraped. When you are finished making
the changes, save the file.
8.6.3
Running a Scraping Tool
Now that we have created a sample script, we then then begin the process of running
the utility. To do this, open up a command line and using the cd command move into the
Scripts folder:
>> cd C:\Python27\Scripts
Temporally Aware Online News Mining and Visualization with Python
185
Once inside the Python27\Scripts folder where our scraping tool is located, we then
need to call Scrapy and pass the source code file we just created as an argument by issuing
the following command:
>> scrapy runspider simplecrawler.py
If all goes to plan, your scraping tool will make a connection to the Google News website
and trigger the parse function. As we have not added any implementation specific functions
into the parse function, no actions will be performed.
8.6.4
News Story Object
As we will be working with individual news stories, to make the process a little easier
we will need to make a News Story object template. At the top of your simplecrawler.py
file, add the following code:
1 class NewsStory:
2
wordList = list()
3
timestamp = 0;
4
storyid = 0;
5
weight = 0;
Inside this class definition, you will see four different variables that each of the news
stories will have. A wordList is used for storing each of the words in the title, a timestamp
variable for storing the timestamp of the story, a storyid variable for storing the unique
story ID, and a weight variable for storing the weight of the story.
As we encounter different news stories, we will create a new NewsStory object for
each and store them locally for later processing. After you have finished save your simplecrawler.py file.
8.6.5
Custom Parsing Function
The following example shows the alterations that we have made to the parsing function.
The response object is the name given to the content currently being processed by Scrapy.
In this example, we have used the HTMLXPathSelector class on the response object that we
have just received. This class allows us to use XPath notation for identifying the elements
that we are interested in selecting. In this example, we are selecting the HTML div tags in
the response object that have a class name of blended-wrapper esc-wrapper.
The XPath notation is passed into the select function. This is telling the selector to search
for a div where the class is equal to blended-wrapper esc-wrapper. Once it has selected the
div tags, it then uses the text to collect the content. To aid the process of identification, a
storyid variable was created that increases incrementally every time a new story has been
encountered.
Update your parsing function to include the following code to do this:
1
2
3
4
5
6
7
8
def parse(self, response):
hxs = HtmlXPathSelector(response)
namesubs = hxs.select(’//div[@class="blended-wrapper esc-wrapper"]/
*’).extract()
storyid = 0;
for namesub in namesubs:
self.processNewsStory(namesub, storyid);
storyid = storyid+1;
186
Text Mining and Visualization: Case Studies Using Open-Source Tools
After we have selected each of these divs, we begin the looping process over the list of
divs using a for-loop and passing each individual div to a custom function along with a
unique story id. This custom function titled processNewsStory is where the processing for
each individual news story will be performed.
8.6.6
Processing an Individual News Story
In this section we will review each of the sections of code that go into our custom
processNewsStory function. Each individual news story is passed to the processNewsStory
function in the form of a namesub variable and a storyid variable. Once the complete
namesub variable has been received, processing is then started to extract the key information
from the story.
In the code sample below you can see that we are using XPath notation to select the
span with a class title of al-attribution-timestamp. Once we have selected this, we are then
extracting the text. After this has been done, the replace function is being used to strip out
some unwanted characters by replacing them with nothing.
Once the timestamp has been extracted, the focus is then placed upon the title text of
the news story. Again, XPath notation is used to select the span where the class attribute is
set up titletext. This value is then extracted. A small amount of cleaning is also performed on
this string to strip out unwanted characters. Update the processNewsStory function (shown
below) to include the following code and then save the file.
1
2
3
4
5
6
7
8
9
10
def processNewsStory(self, namesub, storyid):
print ’processing’
timestamp = Selector(text=namesub).xpath(’//span[@class="alattribution-timestamp"]/text()’).extract()
cleaned_string= str(timestamp).replace("[u’\u200e",’’)
cleaned_string2 = str(cleaned_string).replace("\u200e’]", ’’)
titletext = Selector(text=namesub).xpath(’//span[@class="titletext"
]/text()’).extract()
title_string= str(titletext).replace("[u’\u200e",’’)
cleaned_title = str(title_string).replace("\u200e’]", ’’)
After the initial selection of the data has been performed, processing is then started
on the timestamp to analyze which of the stories are of interest to us. The format of the
timestamp however is in a human readable format such as 1 hour ago or 1 minutes ago.
To work with these timestamps in a machine readable format, depending on the timestamp that was entered, a different block in a nest of if statements is executed to correctly
convert the timestamp. Inside each of these blocks the original text in the timestamp is
removed. The initial number which was provided is then multiplied by a factor of 60 to turn
the timestamp into the number of seconds in total since the news story was published if it
is in minutes format, or 60 * 60 if it is in hour format. This provides a common base for all
timestamps to be compared. The code for processing timestamps is shown as follows:
A variable titled number of seconds old is created to store the seconds that have lapsed
since the story was published in seconds form.
1
2
3
4
5
6
number_of_seconds_old = 0;
if ’minutes’ in str(cleaned_string2):
print ’contains minutes’
timestamp = str(cleaned_string2).replace(’minutes’,’’)
Temporally Aware Online News Mining and Visualization with Python
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
187
timestamp2 = str(timestamp).replace(’ago’,’’)
timestamp_trimmed = str(timestamp2).strip()
number_of_seconds_old = int(timestamp_trimmed) * 60
elif ’hours’ in str(cleaned_string2):
print ’contains hours’
timestamp = str(cleaned_string2).replace(’hours’,’’)
timestamp2 = str(timestamp).replace(’ago’,’’)
timestamp_trimmed = str(timestamp2).strip()
number_of_seconds_old = int(timestamp_trimmed) * 60 * 60
elif ’hour’ in str(cleaned_string2):
print ’contains hours’
timestamp = str(cleaned_string2).replace(’hour’,’’)
timestamp2 = str(timestamp).replace(’ago’,’’)
timestamp_trimmed = str(timestamp2).strip()
number_of_seconds_old = int(timestamp_trimmed) * 60 * 60
Insert this code into the end of your processNewsStory function. After information about
the timestamp has been processed, we can then begin the process of deciding if we should
keep the story or not. To do this, an if statement is created that checks to see if the variable
number of seconds old, which retains how old the current story is, is ≤ than the threshold
of 3,000 seconds. If this is true, then we will begin the storage process of the following story:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
’’’ If the number of seconds old is greater
than our set time, then keep the story
’’’
global Story_Object_Storage
if number_of_seconds_old > 3000:
print ’adding new story’
ns = NewsStory()
lower_case_title = [element.lower() for element in titletext]
ns.wordList = lower_case_title
ns.timestamp = number_of_seconds_old
ns.storyid = storyid
Story_Object_Storage.append(ns)
Insert the code above at the end of your processNewsStory function.
In this code, an instance of the NewsStory class is created, which is used for storing the
story information (timestamp, story id, title content). Each of these fields are then added
to the object, and then the object is appended onto the Story Object Storage list, which
contains a list of each of the stories that we wish to keep for the visualization process. This
list is a global variable stored at the top of our application, so it can be accessed by any
subsequent sections of code.
188
8.6.7
Text Mining and Visualization: Case Studies Using Open-Source Tools
Python Imports and Global Variables
At the very top of your simplecrawler.py file, some basic imports for the function of
Scrapy are needed. Additional imports for working with JSON, random numbers and copying are also needed. The following code should be added to the top of your file:
1
2
3
4
5
6
7
8
from scrapy import Spider
from scrapy.selector import HtmlXPathSelector
from scrapy.selector import Selector
import json
from random import randint
import random
import unicodedata
import copy
Just below this code, a collection of global variables need to be added. These include the
counter variable, which is used to count the number of nodes added; the edge counter for
counting the number of edges (connection between first story and all other stories); and the
Story Object Storage, list which is used for holding each of the NewsStories in memory in
an easily accessible format. The addedEdgeId is used for keeping track of the IDs associated
to each edge created.
Add the following piece of code just below your imports in your simplecrawler.py file.
1
2
3
4
5
6
7
counter = 1
’’’ Edge Counters ’’’
edge_counter = 0;
Story_Object_Storage = list()
addedEdgeId = 1;
8.7
Generating a Visual Representation
When creating a visualization of any data typically the data are first converted into a
format that can be read by a visualization application. These formats are often XML, JSON,
or CSV formatted files allowing for the easy manipulation of data. The data visualization
package that we are using is the SigmaJS library, which takes JSON structured documents
with defined nodes and edges.
The following sample shows an example of the structure of a JSON that is read in by
SigmaJS.
1 {
2 "nodes": [
3
{
4
"id": "n0",
5
"label": "A node",
6
"x": 0,
7
"y": 0,
8
"size": 3
9
},
10
{
11
"id": "n1",
Temporally Aware Online News Mining and Visualization with Python
189
12
"label": "Another node",
13
"x": 3,
14
"y": 1,
15
"size": 2
16
},
17
{
18
"id": "n2",
19
"label": "And a last one",
20
"x": 1,
21
"y": 3,
22
"size": 1
23
}
24 ],
25 "edges": [
26
{
27
"id": "e0",
28
"source": "n0",
29
"target": "n1"
30
},
31
{
32
"id": "e1",
33
"source": "n1",
34
"target": "n2"
35
},
36
{
37
"id": "e2",
38
"source": "n2",
39
"target": "n0"
40
}
41 ]
42 }
In the JSON document you can see that a number of different nodes are defined. Between
each of these nodes a number of different edges can be seen. The source and target for each
edge must correspond to one of the nodes that have been defined. In our application we will
be representing each individual news stories as a single node. Edges between the nodes are
the outlined as the strength of similarity between the first node and all of the other nodes
in the application.
SigmaJS comes with a number of different visualization examples as part of the default
code sample. The example that we are most interested in is the example that has the ability
to read in JSON files directly from the local hard drive. This example can be found inside
the sigma/examples/load-external-json.html file.
8.7.1
Generating JSON Data
After each of the news stories that we are interested in has been added to the
Story Object Storage class during the parsing process, we now can begin the process of
generating the JSON file using these data. A class titled JSONMaker is added which has
two functions: one for creating nodes and one for creating edges between nodes in JSON
format.
The following code should be added to the end of your simplecrawler.py file.
1 ’’’
2 JSON Output generator
3 ’’’
190
Text Mining and Visualization: Case Studies Using Open-Source Tools
4 class JSONMaker:
5
6
def createEdge(self, targetid, weight):
7
global edge_counter
8
global addedEdgeId
9
global Story_Object_Storage
10
11
data = {}
12
data[’id’] = ’’ +str(addedEdgeId)+’’
13
data[’source’] = ’’+str(Story_Object_Storage[1].storyid)+’’
14
data[’target’] = ’’+str(targetid)+’’
15
data[’color’] = ’#ccc’
16
data[’size’] = ’’+str(weight+10)+’’
17
18
addedEdgeId = addedEdgeId + 1
19
json_data = json.dumps(data, sort_keys=True,indent=4, separators=(’
,’, ’: ’))
20
21
if addedEdgeId == 2:
22
json_data +=str(’,’)
23
24
if edge_counter != len(Story_Object_Storage)-1 and addedEdgeId !=
2:
25
json_data +=str(’,’)
26
27
edge_counter = edge_counter+1
28
return json_data
29
30
31
def createNode(self, id, value):
32
global counter
33
34
data = {}
35
data[’id’] = ’’+str(id)+’’
36
data[’label’] = ’’+str(value)+’’
37
data[’x’] = randint(10,50) #Inclusive
38
data[’y’] = randint(1,50) #Inclusive
39
data[’size’] = 1
40
data[’color’] = ’#’+’’.join([random.choice(’0123456789ABCDEF’) for
x in range(6)])
41
42
json_data = json.dumps(data, sort_keys=True,indent=4, separators=(’
,’, ’: ’))
43
44
if counter != len(Story_Object_Storage):
45
json_data +=str(’,’)
46
47
counter = counter+1
48
return json_data
49
50 mk = JSONMaker()
As each of the stories must be plotted on an X and Y axis, for this example these values
are generated using the randint random number generating function. A random color code
generator is also used to give each of the nodes a unique color. Towards the end of this
code, a check is performed to see if the current counter number is not the length of the
Temporally Aware Online News Mining and Visualization with Python
191
Story Object Store list. If it is, then no comma is added to the JSON that is being created
as the last element does not include a trailing comma.
To finish, the json data string containing the JSON for that particular node is returned
back to the caller of the function to help build the final JSON output.
8.7.2
Writing a JSON File
To generate the final output, inside the original parse method of the crawler, a pipe is
opened to create a data.json file that can be read into the SigmaJS example visualization.
For the JSON document to be fully valid, additional sections must be added such as the
header of the JSON document and to the footer of the JSON document.
In the middle of this section of code, a loop is performed over the Story Object Storage
list to pull down each of the NewsStory objects that have been stored. Each story is then
passed to the JSONMaker outlined in the previous section to generate the JSON output
for each story. This JSON output that is created is then embedded between the header and
the footer of the JSON document to make a complete valid JSON document.
In the following sample you can see the file is being opened using the open() function,
which is being passed the w flag to indicate to write to the file.
Just after the first for-loop inside the original parse function, the following code should
be added:
1
2
3
4
5
6
7
8
9
10
11
’’’ Loop over nodes ’’’
f = open(’data.json’, ’w’)
header = ’’’
{
"nodes": [ ’’’
f.write(header)
footer = ’’’ ], ’’’
The following sample shows the process of looping over each of the stories stored inside
of the Story Object Store. For each of the stories, the createNode function is being called to
create a new node in the JSON file.
After the nodes have been created in JSON, they are then written to the data.json file
using the f.write command. After generating the JSON for the nodes a JSON header for
the edges is then output and then the first record from the object store is then pulled down
and stored inside the firstRecord variable.
The first record will be treated as the single object that all other news stories will
be compared against. We then loop over each of the news stories and call the createNode
function to generate the associated JSON.
Add the following code to the end of your parse function:
1
2
3
4
5
6
7
8
9
global mk
for story in Story_Object_Storage:
f.write(mk.createNode(story.storyid, story.wordList))
f.write(footer)
edges = ’’’ "edges": [ ’’’
f.write(edges)
192
10
11
12
13
14
15
Text Mining and Visualization: Case Studies Using Open-Source Tools
’’’
Generate a list of edges between each of the nodes
if a similar words is found.
’’’
firstRecord = str(Story_Object_Storage[0].wordList)
In the following code we check the text in the first news story word by word to the text
in the rest of the news stories. The similarity strength between the first news story and
all other stories is built up here by calculating how often the terms from the first story
appeared in the other news stories.
For each time a term from the first record is found in record N, the story.weight integer
is incremented by 1. To aid the readability the text found a match! is outputted to the
console.
Add the following code to the end of your parse function:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
’’’
Calculate matches between the first record
and all the other records.
’’’
for story in Story_Object_Storage:
for word in story.wordList:
for term in firstRecord.split(’ ’):
originalTerm = term.encode(’ascii’, ’ignore’).decode(’
ascii’)
sentence = word.encode(’ascii’, ’ignore’).decode(’ascii’)
if originalTerm in sentence:
print ’found a match!’
’’’ Increment the weight of the story’’’
story.weight = story.weight+1
Once each of the weights indicating the strength between the first news story and all
other stories is created, the final step is output the edge JSON. The following code outlines
the process of skipping the first record and then calling the createEdge function to generate
the associated JSON. This JSON is then written to a file using the f.write command.
Add the following code to the end of your parse function:
1
2
3
4
5
6
7
8
9
10
’’’
Generate the edges
’’’
for story in Story_Object_Storage:
if story.storyid != ’0’:
print story.storyid
print story.weight
print ’----’
f.write(mk.createEdge(story.storyid, story.weight))
To finish the end of the JSON file is output. Add the following code to the end of your
parse function to complete the JSON document:
1
2
3
4
edges_footer = ’’’
]
}’’’
Temporally Aware Online News Mining and Visualization with Python
5
6
7
193
f.write(edges_footer)
f.close()
If at any stage of developing a JSON document you are getting an unexplained error,
you can use one of the online JSON validation tools such as http://jsonlint.com to
check that it is not stemming from a malformed JSON document.
You can now run this complete application by following the steps outlined in Section 8.6.3. This will output the data.json file for you. We will be using this data.json file in
the next section.
8.8
Viewing the Visualization
The online nature of JavaScript applications often requires the code to be running in a
Web browser from a Web address such as localhost. To prevent any issues arising with our
code we will utilise a simple HTTP server on our local machine for testing.
Section 8.2.1 outlined the process of downloading XAMPP. After installing XAMPP, we
need to start the Web server. This is done by first opening up the XAMPP control panel
shown in Figure 8.3.
FIGURE 8.3: XAMPP control panel.
After starting XAMPP, click on the Start button beside Apache. Status messages will
appear telling you if Apache has started and what port it is bound to. Next click on the
Explorer button to open the root directory of the Web server. Place your sigama folder
inside of the htdocs folder in the root of your XAMPP installation directory. By doing
this we will sidestep the XMLHttpRequest error outlined earlier from appearing in the Web
browser when running the code.
Before moving any further, we must move the data.json file that was output earlier into
the htdocs \sigma\examples\data folder.
A slight alteration to the load-external-json.html file is made to allow the edge weights
194
Text Mining and Visualization: Case Studies Using Open-Source Tools
which have been added to be shown. The path to the file we have created is also changed
to point to data/data.json that we output earlier.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
....
sigma.parsers.json(’data/data.json’, {
renderer: {
// IMPORTANT:
// This works only with the canvas renderer, so the
// renderer type set as "canvas" is necessary here.
container: document.getElementById(’graph-container’),
type: ’canvas’
},
settings: {
minNodeSize: 1,
maxNodeSize: 10,
minEdgeSize: 0.1,
maxEdgeSize: 40,
}
});
CustomEdgeShapes.init(sigma);
sigma.refresh();
....
After making the alterations and saving the file we can then open our Web browser and
type the following path:
http://localhost/sigma/examples/load-external-json.html
Due to the quiet nature of JavaScript when something goes wrong it can be difficult
to tell when an error has occurred. A great way to check what has error has occurred is
by using the console that is available inside Google Chrome by right-clicking anywhere and
then selecting Inspect Element from the menu that opens and then selecting Console. This
will outline any of the errors on the page.
Figure 8.4 shows each of the nodes representing individual news stories. The edges
between each of these nodes outlines the strength of the similarity between the first news
story and each of the other news stories that have been saved. In some cases, as can be seen
at the top of this figure, some news stories will have no connection to the first news story
if no similarity is found. All stories not in the required time frame have been removed from
the visualization before the JSON file has been generated.
Figure 8.5 shows a closer view of the news stories. Each of the labels attached to each
of the nodes can be seen. In this view, the terms that make up the news story title can
be inspected. This image outlines a view of the labels attached to each of the news stories
when zoomed in closely.
8.9
Additional Concerns
Web content sources (although freely available to the user) may have a different policy
when automated services are accessing their content. Generally most services offer a paid
subscription to the service when any automatic processes are concerned due to the drain
on resources often attributed to Web scraping utilities. Remember that a scraping utility
Temporally Aware Online News Mining and Visualization with Python
FIGURE 8.4: News stories.
FIGURE 8.5: Closer view of news stories.
195
196
Text Mining and Visualization: Case Studies Using Open-Source Tools
often has a one-track mind when retrieving content (e.g., keep scraping)! To prevent the
misuse of a service, it is advised to read the terms and conditions of the site that you will
be scraping to ensure they approve of the process.
To prevent any bad feeling between your scraping tool and the service, additional timeout
periods can be established to allow the tool to automatically take a pause after a certain
number of requests are made. The least desired result is for the content provider to blacklist
your IP address due to overuse.
A plain text file added by the system administrator named robots.txt is often added to
websites. The purpose of this file is to define how automated services should interact with
the Web server and the content that it provides. Adhering to this policy is a responsibility
given to the owner of the scraping utility. All applications that interact with Web services
also have the ability to identify themselves to a server. This allows the Web server to allow
or deny the scraper. It is often been the case that a user can access a webpage from their
Web browser, but the Web crawler cannot access the page.
The fingerprint which is given by the Web browser may be accepted, (e.g., Internet
Explorer or Google Chrome), but the fingerprint from a homebrew scraping utility may
not be. This is often a quick way to gauge the reception of a content providers attitude
towards automatic scraping utilities. To counteract this, scraping utilities often come with
the ability to change the fingerprint they will leave to become a generic Web browser when
they are asked to be identified by a Web server.
8.9.1
Authentication
When interacting with webpages, the views are often broken up into two main views.
The first of these is the anonymous view whereby the user does not need to be logged into
the system. This allows all of the content on the system to be available with ease. The
second of these is the authenticated user. These user accounts require the user to have an
account with the content provider to allow the content to be accessed. If the user does not
have an account he or she is often just redirected to another page requesting the user to log
into the page. The process of creating an account with a content provider is often designed
so that a human user must authenticate with the service. This approach is used to help
prevent the automatic creation of accounts and the spreading of spam.
When a tool is interacting with a content provider, a session is often created on the
server side, and a token is stored on the client side. Each time a request is made, a token
from the user is passed to the server so the server side session can be matched to the user
again on subsequent requests.
8.9.2
Errors
Like most pieces of software, errors are not accepted very well when the constraints of
real-time Web content processing is concerned. The large volumes of data that can quickly
pass through the tool provide a one chance pass to ensure everything goes as planned
without the need for reprocessing to occur. To maximize the uptime of the crawling tool,
it is strongly encouraged that the native logging features of Python are utilized to create a
comprehensive log of wherever errors have occurred during processing.
The concept of a graceful failure should always be kept in mind. If a website is currently
offline or producing some undesirable output causing an error to be thrown, the error should
simply be logged and the application should move onto the next website. This ensures that
long, time-consuming Web crawls will not be in vain if worst does come to worst.
When interacting with Web servers, a number of different server response status codes
[7] are often returned to the caller (e.g., the Web scraping utility), outlining what the status
Temporally Aware Online News Mining and Visualization with Python
197
of the webpage or resource currently is. These status codes can be of vital importance as
they give a quick summary of the webpage status, without the need for the entire website
to be crawled. Two examples of status codes that can be an early warning of issues are the
307 Temporary Redirect and 308 Permanent Redirect codes. These outline that the page
the crawler is looking for is not there and telling the crawler to move to a different page.
If the crawler is a self-built tool then the tool may not have the capabilities required to
follow to the new page. Other codes such as the 403 Forbidden, 404 Not Found and 408
Request Timeout are simple examples of status codes that can easily be found in everyday
Web scraping and crawling.
8.10
Summary
With the ever-increasing amount of data on the Web competing for a user’s viewing
time, emphasis has been placed upon developers to create unique visualizations to find data
most relevant to the current user. This chapter outlined the process of scraping data from
the Web using the free and open-source Web scraping tool Scrapy. The HTML structure of
the Google News aggregation webpage was described outlining fields that are of interest to
us, mainly the timestamp and description text.
After the data had been scraped, a custom filtering process was created that applies
time windows to the data and retaining and visualizing only the relevant information given
a set time window. Additional filtering was described outlining how common features such
as advertisement text and copyright notices can be removed from the data.
A custom JSON file was generated containing a unique node for each news story and
edges identifying the similarity from the first news story to all the other news stories. The
SimgaJS visualization library was used to show how a lightweight viewer can be created to
visualise the complex connections between news stories.
Additional concerns such as authentication, server errors, and common scraping rules
were outlined towards the end of the chapter. This chapter shows that creating a Web
scraping and visualization utility can be quickly and easily developed with little overhead
to the developer.
Bibliography
[1] Python. https://www.python.org/. Accessed: 2015-04-22.
[2] XAMPP. https://www.apachefriends.org/index.html. Accessed: 2015-0422.
[3] PIP. https://pypi.python.org/pypi/pip. Accessed: 2015-04-22.
[4] Scrapy. http://scrapy.org/. Accessed: 2015-04-22.
[5] lxml. https://www.gephi.org/. Accessed: 2015-04-22.
[6] SigmaJS. http://sigmajs.org/. Accessed: 2015-04-22.
198
Text Mining and Visualization: Case Studies Using Open-Source Tools
[7] Status Codes List of HTTP Status Codes. http://en.wikipedia.org/wiki/
List_of_HTTP_status_codes. Accessed: 2015-09-30.
Chapter 9
Text Classification Using Python
David Colton
IBM, Dublin, Ireland
9.1
9.2
9.3
9.4
9.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.1
Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.2
The Natural Language Toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.3
scikit-learn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.4
Verifying Your Environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.5
Movie Review Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.6
Precision, Recall, and Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.1.7
G-Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Modelling with the Natural Language Toolkit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.1
Movie Review Corpus Data Review . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.2
Developing a NLTK Naı̈ve Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.3
N-Grams . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.4
Stop Words . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.2.5
Other Things to Consider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Modelling with scikit-learn . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
9.3.1
Developing a Naı̈ve Bayes scikit-learn Classifier . . . . . . . . . . . . . . . . . . . . . .
9.3.2
Developing a Support Vector Machine scikit-learn Classifier . . . . . . . . .
Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
199
199
200
200
201
201
202
203
204
204
204
208
209
210
212
213
218
219
219
Introduction
This chapter will examine how Python, the Natural Language Toolkit (NLTK), scikitlearn, and Numpy can be used to classify text. Using a corpus of previously classified positive
and negative movie reviews, a number of classification models will be developed to predict
whether unseen movie reviews can be correctly identified as either positive or negative in
nature.
This chapter is not a Python tutorial. It will cover just enough to let you run through
the samples described. Nor does this chapter provide an in-depth analysis of the tools
and packages described. There are many excellent books and online tutorials that can give
you, the reader, a more thorough explanation of the concepts introduced here if required.
What this chapter does provide, however, is an introduction to the packages mentioned and
examples of how text can be classified.
199
200
Text Mining and Visualization: Case Studies Using Open-Source Tools
9.1.1
Python
Python [1] is an open-source high-level programming language developed under an OSI
approved open-source license. Although considered an excellent choice for the user who is
only starting to learn how to program, Python offers enough features and functionality
to meet the demands of the professional object-oriented programmer as well. It is highly
scalable, portable, embeddable, and suitable for development projects of all sizes. Another
great advantage of Python is a very active community that enthusiastically develops, releases, and supports commercial quality libraries, packages, tools, and platforms that enhance and extend the core Python capabilities. Two such packages, NLTK and scikit-learn,
are extensively used in this chapter.
Python may already be installed on your machine. The simplest way to determine this
to open a command prompt or terminal and type ’python’ or ’python3’. If Python is already
installed you should see something similar to the first two or three lines of Figure 9.1. If it is
not, it can be downloaded from https://www.python.org/downloads/. Installation
instructions will vary depending on your Operating System (OS), but there are many excellent references and tutorials available on the web. For this chapter Python version 3.4.1,
32 bit, was used. This is the version of Python recommended for use with NLTK version 3
at the time of writing.
9.1.2
The Natural Language Toolkit
The Natural Language Toolkit (NLTK) [2] is a leading platform for building Python programs to work with human language data. The NLTK provides a suite of libraries to handle
text processing tasks such as corpora parsing, tokenisation, stemming, and classification.
The NLTK also provides libraries that support part-of-speech tagging, which converts a
sentence into a list of word / tag tuples, and chunk extraction that can be used to extract
a short phrase from a sentence that has previously been part-of-speech tagged. The NLTK
provides easy access to WordNet, a large lexical database of English nouns, verbs, adjectives,
and adverbs and over 50 text corpora.
Like Python, the NLTK is a free, open-source project. It is community driven and is
available for Linux, Windows, and Mac OS machines. Instructions for installing NLTK on
different operating systems is given at http://www.nltk.org/install.html. On this
webpage an optional installation of the NumPy 1.8.1 package is suggested. Although NumPy
is listed as optional, it, and SciPy, are both required by scikit-learn. It is suggested that
both are installed at this point. NumPy and SciPy can be downloaded from the following
locations:
• http://sourceforge.net/projects/numpy/files/NumPy/
• http://sourceforge.net/projects/scipy/files/scipy/
NLTK 3.0.0, NumPy 1.8.1 [3] and SciPy version 0.14.0 [4] were used in the Python
scripts developed for this chapter.
9.1.3
scikit-learn
scikit-learn [5]is a simple and efficient tool for data mining, data analysis, and machine
learning in Python. It utilises other Python libraries such as NumPy, SciPy, and Matplotlib.
Like Python and NLTK, it is open-source available under a BSD license. scikit-learn includes
many supervised learning algorithms including Naı̈ve Bayes, Support Vector Machines, Random Forests and Decision Trees, as well as many unsupervised learning techniques, including
clustering and hidden Markov models.
Text Classification Using Python
201
As well as supporting cross-validation and providing comprehensive grid search capabilities for parameter optimization, scikit-learn also provides extensive model evaluation
metrics and scoring options including, for example, precision, recall, and accuracy as well
as an easily accessible confusion matrix. Feature extraction including TF-IDF is also supported.
Like the NLTK, if you are using Mac or a flavour of Linux, scikit-learn can be installed
using the Pip package install utility. If you are installing on a Windows machine an installer can be downloaded from http://www.lfd.uci.edu/˜gohlke/pythonlibs/
#scikit-learn
9.1.4
Verifying Your Environment
Python and the required third-party libraries and packages should now be installed and
ready to use. To verify your environment, open the Python interpreter by typing ’python’
at your command or terminal prompt. Information about the version of Python installed on
your machine will be displayed followed, on the next line, by three greater than signs >>>.
Enter the following commands to verify the each required package is correctly installed.
If each import command completes without throwing an error then you have successfully
installed all required packages.
>>>
>>>
>>>
>>>
import
import
import
import
nltk
sklearn
numpy
scipy
The version of each package installed can also be displayed using the following commands. The expected output from running all commands is shown in Figure 9.1.
>>>
>>>
>>>
>>>
nltk.__version__
sklearn.__version__
numpy.__version__
scipy.__version__
FIGURE 9.1: Verifying your Python environment on a Windows machine.
9.1.5
Movie Review Data Set
Now that the Python environment is correctly configured and ready to use, attention
can be turned to getting the movie review data set. These data were first used in [6]. The
202
Text Mining and Visualization: Case Studies Using Open-Source Tools
movie review data set is one of the many corpora available with the NLTK. To install it,
run the Python interpreter, and enter the following commands:
>>> import nltk
>>> nltk.download()
A new window should open, showing the “NLTK Downloader”. Click on the File
menu, select Change Download Directory, and specify a central location, for example,
c:\nltk_data. Select the Corpora tab, and then scroll down and highlight movie review
data set. Click the download button to retrieve the corpus. Figure 9.2 shows the NLTK
Downloader tool.
FIGURE 9.2: Using the NLTK built-in NLTK Downloader tool to retrieve the movie review
corpus.
The data set contains 1,000 positive movie reviews and 1,000 negative movie reviews and
will be examined in more detail later. Further background information about the corpus
can be found in the read me file in the data set root folder.
9.1.6
Precision, Recall, and Accuracy
When evaluating the performance of a model precision, recall and accuracy will be used.
When predicting the class of a review there are four possible outcomes:
1. True Positive (TP)
A true positive is where a review is predicted as positive and was classified as positive
2. False Positive (FP)
A false positive is where a review is predicted as positive but was classified as negative
3. False Negative (FN)
A false negative is where a review is predicted as negative but was classified as positive
Text Classification Using Python
203
4. True Negative (TN)
A true negative is where a review is predicted as negative and was classified as negative
The overall accuracy of a model is calculated as
N umberof T rueP ositives + N umberof T rueN egatives
T otalN umberof Examples
(9.1)
Positive Class Precision is calculated as
N umberof T rueP ositives
N umberof T rueP ositives + N umberof F alseP ositives
(9.2)
Positive Class Recall is calculated as
N umberof T rueP ositives
N umberof T rueP ositives + N umberof F alseN egatives
(9.3)
Negative Class Precision is calculated as
N umberof T rueN egatives
N umberof T rueN egatives + N umberof F alseN egatives
(9.4)
Negative Class Recall is calculated as
N umberof T rueN egatives
N umberof T rueN egatives + N umberof F alseP ositives
(9.5)
Precision and recall are inversely related, which mean that as precision increases recall
falls and inversely where recall rises precision decreases. The critical decision then is to
develop a model that gives high precision and low recall or a model the delivers a low
precision value but has high recall. A high precision value with a lower recall value implies
that a high percentage of reviews that are predicted as positive will actually be positive.
However, the downside of such a model is that a lot of positive reviews will not be correctly
identified. High recall and low precision values obtained from a model would imply that a
high percentage of positive reviews are correctly identified as such, but, as a consequence, a
large number of negative reviews would be incorrectly identified as positive yielding a lower
precision value. Usually a trade-off has to be made between precision and recall depending
on the situation and the preferred outcomes.
9.1.7
G-Performance
In addition to standard classifier performance measures (e.g., accuracy, precision, recall),
Kubat and Matwin [7] describe another measure that uses the geometric mean of the accuracies measured separately on each class. This measure they call the g-performance. The
goal of this measure is to maximise the recall of both class but at the same time keeping
them balanced such that a poor value for either the positive or negative class will give an
overall poor performance for the classifier.
G-performance is calculated as
s
T rueP ositives
T rueN egatives
g=
×
(9.6)
T rueP ositives + F alseP ositives T rueN egatives + F alseP ositives
The g-performance of each model developed will be calculated and analysed to determine
which models have performed the best.
204
9.2
Text Mining and Visualization: Case Studies Using Open-Source Tools
Modelling with the Natural Language Toolkit
In this section you will develop several classifier models using the movie review corpus
and the Natural Language Toolkit. Models will be developed with and without stops words
and with unigrams, bigrams, and trigrams. However, before developing your first model it
is important to understand the data set that you will be using.
9.2.1
Movie Review Corpus Data Review
Before starting any text mining exercise it is very important to first understand the data
set that will be used. As you can see in Figure 9.3, beneath the top-level folder the movie
reviews are separated into two folders: one for positive reviews, the ’pos’ folder; and one
for negative reviews, the ’neg’ folder. As mentioned earlier a read me text file in the root
folder gives the reader some background information on the history and use of the data set.
There are 1,000 reviews in each folder.
FIGURE 9.3: The structure of the movie review corpus.
Figure 9.4 shows a sample positive review. Examining the text it can be seen that
all characters are lowercase. It can also be observed that punctuation marks such as full
stops, commas, and brackets, are preceded and followed by a space. It also appears that
there are no non-ASCII characters present in the files. This shows that a certain degree
of preprocessing has already been performed on the text of the reviews. Standardising the
case of all characters, removing non-ASCII characters, and separating the punctuation in
this manner are all steps aimed at producing a standard set of tokens to be used in the
development of the classification model. When seen by the learning algorithm, ’acting,’
and ’acting’ are two completely different tokens. So rather than having two occurrences
of a single “acting” token, there are two different tokens that each occur only once.
Not all non-alphanumeric characters have been completely separated from words as highlighted in Figure 9.4. The author of the review has used a double dash, --, to isolate some
of their thoughts from the rest of the review. These dashes could lead to the generation of
unique tokens not seen elsewhere in the corpus. These tokens could then play an overemphasised role in the prediction of a positive or negative class, which is not what is wanted.
On the other hand, the second highlighted token, father-and-son, is a concatenation of
three words where it makes sense that they are treated as a single token.
Text Classification Using Python
205
FIGURE 9.4: A sample positive review from the movie corpus.
9.2.2
Developing a NLTK Naı̈ve Bayes Classifier
The first model we develop will use the Natural Language Toolkit utilising a Naı̈ve
Bayes learner and a feature-based bag-of-words approach. As seen, the data are stored in
the NLTK corpus format where each review is written to a file in a folder that represents
whether it was classified as a positive or negative review. The complete listing is available
for download from the book website.
The development of a NLTK Naı̈ve Bayes classifier is a relatively straight forward task.
The first step is to create a categorised plain text corpus reader. The reader takes three
parameters. The first parameter is the root folder for the corpus, the second is a regular
expression used to identify the files in the corpus, the third is a regular expression used to
determine the category types which are derived from the folder names within the corpus.
In the code snippet shown in Listing 1, the corpus to be read is movie_reviews and all
files with a .txt extension are included. Now that the reader is initialised the next step is
to identify the file ids of each category type.
# Set the working folder to the nltk_data corpora folder
import os
from
nltk.corpus.reader import CategorizedPlaintextCorpusReader
os.chdir (’C:/nltk_data/corpora’)
reader = CategorizedPlaintextCorpusReader(’./movie_reviews’,
r’.*\.txt’,
cat_pattern=r’(\w+)/*’)
Listing 1: Creating a categorised plain text corpus reader
Identifying the file ids of each category is an important step. The words in each file
must be correctly labelled as either being positive or negative in nature. Achieving this
is a simple matter of calling the fileids method of the reader object just created and
manually passing the category name of the required file ids. The code snippet in Listing 2
returns two file lists. The first, pos_ids is a list of all the files, including each files relative
path, in the positive review category while the neg_ids is the list of all files in the negative
review category.
It’s very simple to review a sample of the contents of each of these lists. Once you have
successfully run the entire python script typing the following command into the interpreter
will show you the first three elements of each list. You can also get the length, or number
of elements, of each list using len().
206
Text Mining and Visualization: Case Studies Using Open-Source Tools
# The positive reviews
pos_ids = reader.fileids(’pos’)
# The negative reviews
neg_ids = reader.fileids(’neg’)
Listing 2: Identifying the file ids of each category
>>> pos_ids [:3]
[’pos/cv000_29590.txt’, ’pos/cv001_18431.txt’, ’pos/cv002_15918.txt’]
>>> neg_ids [:3]
[’neg/cv000_29416.txt’, ’neg/cv001_19502.txt’, ’neg/cv002_17424.txt’]
>>> len(pos_ids)
1000
>>> len(neg_ids)
1000
The NLTK uses the concept of a feature, in this case any word or token, and the presence
of that token in a sample of text in order to predict which class the sample might belong
to. The code in Listing 3 shows how a positive and negative feature dictionary is generated
for each positive review file and for each negative review file. The word_feats function
returns a dictionary, or hash, of key value pairs where each key is the individual word, or
token, and the value is always true. The pos_feat and neg_feat objects are, in fact, a
list of lists. Each element in the features list contains two elements. The first element is the
dictionary returned by the word_feats function. The second element represents whether
the category of the review is either positive or negative.
def word_feats(words):
return dict([(word, True) for word in words])
pos_feat = [(word_feats(reader.words(fileids=[f])),
’pos’)
for f in pos_ids]
neg_feat = [(word_feats(reader.words(fileids=[f])),
’neg’)
for f in neg_ids]
Listing 3: Generation of the positive and negative feature dictionaries
It should be noted that all sentence word order has been lost when generating the feature
dictionaries as would be expected when using a bag-of-words approach. A similar process is
performed to create the negative features. To view the first element of the positive features
list you can enter pos_feat[0] into the interpreter.
>>> pos_feat[0]
({’alan’: True, ’crazy’: True, ’won’: True, ’menace’: True,
’manson’: True, ’see’: True, ’isn’: True, ’carving’: True
...
’a’: True, ’identity’: True}, ’pos’)
Next, the Naı̈ve Bayes learner is trained. Unfortunately, NLTK does not natively support cross-validation so the samples were simply divided into a training set consisting of
the first 800 samples and a testing set of the remaining 200. A Naı̈ve Bayes classifier
was then created using the training data set as shown in Listing 4. Also shown is the
generation of a reference set and test set to be used later to calculate the performance
Text Classification Using Python
207
of the classifier. To achieve this the actual classification of each of the test samples is
loaded into the reference set and the value for the sample returned from the classifier,
observed = classifier.classify(feats), is loaded into the test set.
# Create the test and train data sets
train = pos_feat[:800] + \
neg_feat[:800]
test = pos_feat[800:] + \
neg_feat[800:]
# Create the classifier using the train data set
classifier = NaiveBayesClassifier.train(train)
# The NTLK metrics package uses sets to
# calculate performace metrics
# Create empty reference and test sets
refsets = collections.defaultdict(set)
testsets = collections.defaultdict(set)
#Populate the metrics sets with values
for i, (feats, label) in enumerate(test):
refsets[label].add(i)
observed = classifier.classify(feats)
testsets[observed].add(i)
Listing 4: Generation of the positive and negative feature dictionaries
The performance of the model can then be calculated using the NLTK Metrics package
and the most informative features can also be displayed. Listing 5 shows how the performance of the model is calculated and the top five most informative features are displayed.
The calculated performance values are then output to the screen as shown in Figure 9.5.
FIGURE 9.5: Performance values of the first NLTK model developed.
Considering the top five most informative features, it is clear to see how these tokens,
real words in this example, could easily be included in a positive or negative review. Of the
top five most informative features two are for postive reviews and three are for negative
reviews. Taking outstanding as an example, where this token appears in a review it is
13.9 times more likely to appear in a positive review than in a negative review. The next
token, insulting, is 13.7 times more likely to appear in a negative review than in a
positive review.
208
Text Mining and Visualization: Case Studies Using Open-Source Tools
# Calculate precision. recall, f-measure and accuracy
pos_pre = nltk.metrics.precision(refsets[’pos’], testsets[’pos’])
pos_rec = nltk.metrics.recall(refsets[’pos’],
testsets[’pos’])
neg_pre = nltk.metrics.precision(refsets[’neg’], testsets[’neg’])
neg_rec = nltk.metrics.recall(refsets[’neg’],
testsets[’neg’])
accuracy = nltk.classify.accuracy(classifier, test)
gperf
= math.sqrt(pos_rec * neg_rec)
# Show the top 5 most informative features
classifier.show_most_informative_features(5)
# Print out the overall performace measures
print (’pos precision:\t’, pos_pre)
print (’pos recall:\t’,
pos_rec)
print (’neg precision:\t’, neg_pre)
print (’neg recall:\t’,
neg_rec)
print (’model accuracy:\t’, accuracy)
print (’g-performance:\t’, gperf)
Listing 5: Calculating the performance of the model and displaying output to screen
The overall accuracy of the model, 73.5%, could suggest that this was a very successful
first attempt at developing a model to predict the class of these movie reviews. However,
looking closely at the precision and the recall for the different classes, this is clearly not
the case. The negative class precision, 95.1%, shows that nearly every sample classified as
negative was an actual negative review sample. However, a negative class recall value of
49.5% shows that overall only half of all the negative samples, approximately 100 of the 200
samples, were identified as such. Examining the positive class next, a recall value of 97.5%,
shows that nearly all of the positive reviews were correctly identified. However, a precision
value of just 65.8% tells us that only approximately two-thirds of all the reviews predicted
as positive were actually positive, that is approximately 300 of the test set samples were
predicted as positive.. This tells us that this model is overpredicting test samples as positive
and underpredicting negative samples.
9.2.3
N-Grams
Having developed the basic NLTK model, it is a simple matter to augment and hopefully
enhance the performance of the model by using n-grams and stop word removal.
N-Grams are a contiguous sequence of n items from an extract of text or speech and
in this occurrence the n-grams refer to words. Typical n-grams used in text mining are
unigrams, bigrams, and trigrams. For example, given a sequence of words “The quick brown
fox” the following trigrams, or sequences of three words, can be formed “The quick brown”
and “quick brown fox”. Modifying the script from Section 9.2.2 to use bigrams, for example,
is a simple matter of importing the NLTK ngrams package and, instead of passing single
words or tokens to the positive and negative feature generator, pass bigrams instead. This
is shown for positive features in Listing 6.
Examining the performance metrics from the model using bigrams, Figure 9.6 shows
that the introduction of bigrams has, in fact, positively impacted the overall accuracy of the
model. Although the positive precision has fallen slightly all other measures have increased
leading to a better model. Considering the top most informative feature, we understand
Text Classification Using Python
209
from nltk.util import ngrams
pos_feat = [(word_feats(ngrams(reader.words(fileids=[f]), 2)),
’pos’)
for f in pos_ids]
Listing 6: Generation of the positive and negative feature dictionaries
that when the two words not and funny appear together we are 14.3 times more like to
be looking at a negative movie review than a positive one.
FIGURE 9.6: Performance of NLTK model using bigrams.
Now that bigrams features have been modelled, using trigrams is as simple as changing
the parameter passed to the ngrams function from 2 to 3 as shown in Listing 7.
pos_feat = [(word_feats(ngrams(reader.words(fileids=[f]), 3)),
’pos’)
for f in pos_ids]
Listing 7: Generation of the positive and negative feature dictionaries
Looking at the performance of the model where trigrams were used, Figure 9.7, we can
again see that there has been overall improvement when compared against the unigram
and bigram models. Although there have been another very slight decrease in the positive
class precision the values for positive recall and for negative precision and recall have all
increased. A word of warning though. A look at the most informative features shows two
trigrams which may suggest that the model may be over trained to the data. The trigrams
’work with ,’ and ’well , and’ are classified as being indicative of negative and positive reviews respectively. It is clear that neither n-gram is obviously positive or negative
and it could be suggested that this model may not generalise well to the wider population
of all movie reviews.
9.2.4
Stop Words
When processing natural language, stop words are words that are filtered out of the
corpus, typically early on in the process. Stop words are usually what are called short
function words or words with little or ambiguous meaning. Examples of English stop words
include “the”, “and”, “at”, and “which”.
As we require that the stop words are removed before the n-grams are generated the only
210
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 9.7: Performance of NLTK model using trigrams.
part of the code that needs to change to achieve this is in the call made to the word_feats
function as shown in Listing 8.
pos_feat = [(word_feats(
ngrams([word for word in reader.words(fileids=[f])
if not word in stopwords.words(’english’)], 3)),
’pos’)
Listing 8: Generation of the positive and negative feature dictionaries
Analysing this code further, it is saying that for every word found in file f, if a word
is not an English stop word pass it to the ngrams functions. The ngrams function then
generates the required size n-gram, in this case three, and subsequently passes each n-gram
to the word_feats function, which was previously described.
Running this model for unigram, bigrams, and trigrams with stop words removed returns the performance results summarised in Table 9.2.4. Comparing these results to those
achieved earlier, when stop words were not excluded, shows that the performance of unigrams and trigrams has actually fallen. Only the performance of bigrams has improved.
These performance results, coupled with a factor of 10 increase in the execution time of
the script, would suggest that the removal of stop words was not beneficial. On closer inspection it was observed that the increase in execution time was solely caused by the stop
word removal. The time taken to generate the model, and the classification of the unseen
samples, was mostly unchanged. It was also seen that punctuation characters appear in the
top-five most informative features for both bigrams and trigrams. For example:
• (’.’, ’cameron’)
= True
• (’.’, ’basically’, ’,’) = True
pos : neg = 11.7 : 1.0
neg : pos =
9.0 : 1.0
This could indicate that the removal of punctuation is another potential line of enquiry
to improve the performance of the model.
9.2.5
Other Things to Consider
In the development of the initial model we considered n-grams and stop word removal as
potential approaches to increase the performance of the model. Some other transformations
we can apply to the text include:
• Removing punctuation
Text Classification Using Python
211
TABLE 9.1: Performance comparison of unigrams, bigrams, and trigrams when stop words
have been removed.
Measurement Uni-grams Bi-grams Tri-grams
pos precision
pos recall
neg precision
neg recall
model accuracy
g-performance
execution time
0.650
0.975
0.950
0.475
0.725
0.681
802.1
0.750
0.930
0.908
0.690
0.810
0.801
799.8
0.783
0.795
0.792
0.780
0.788
0.787
651.5
• Removing single characters or smaller tokens of a specific length
• Stemming
As previously seen, the format of the given review text has separated most punctuation
characters from the rest of the text, for example "sverak also wrote the script ,
while his son , jan , directed". These punctuation characters also appeared in
some of the most informative bigram and trigram features. In this section we will examine
how these characters can easily be removed. We will also develop a simple routine to remove
tokens that are shorter than a specified characters length.
Stemming describes the process for reducing words to their word root, or stem. An
English word stemmer might, for example, identify “stemmer”, “stemming”, and “stemmed”
as all being based on ”stem”. Julie Beth Lovins first published a paper on stemming in
1968 [8]. Two stemmers provided by NLTK are called Porter and Snowball. In the sample
provided the Snowball stemmer is used.
In the Python scripts developed to this point, the generation of the positive and negative
features for use in the classifier was done in a single complex statement. To make it easier
to understand the different steps taken during the processing of the text, a function was
developed that separates each step out. The signature of this function is shown in Listing
9. The function takes 1 mandatory parameter, file_ids which holds the list of files to
process, and six optional parameters. The optional parameters are:
• removepunctuation
The remove punctuation parameter is used to signal if punctuation characters should
be left in the text, 0, or removed from the text 1
• minwordsize
This parameter defines the minimum size token allowed. For example, if set to 3 then
only token of three characters are more are included. The default is 1 character which
essentially includes all text.
• removestopwords
The remove stop words parameter is used to signal if stop words should be left in the
text, 0, or removed from the text 1
• stemwords
The stem words parameter is used to signal if stemming should be applied to the
words in the corpus. The default value, 0, means words will not be stemmed, whereas
1 will cause all words to be stemmed.
212
Text Mining and Visualization: Case Studies Using Open-Source Tools
• ngramsize
The n-grams parameter is used to specify the size of n-grams to use. The default value
is 1, which corresponds to using unigrams.
• reviewtype
Specified whether the review is positive in nature, pos, or negative neg. The reviews
are tagged as positive by default.
These parameters can be modified as required to turn on or off options and set the size
for n-grams and minimum word size as required.
def prepare_review( file_ids,
removepunctuation=0,
minwordsize=1,
removestopwords=0,
stemwords=0,
ngramsize=1,
reviewtype=’pos’):
Listing 9: The signature for the prepare review function
Listing 10 shows some of the details of the logic of how each of the text processing steps
are implemented. Most of the processing steps simply treat the corpus text as a string.
Regular expressions are used to remove punctuation characters. Minimum word size, stop
words and word stemming is performed by testing each word in a review and, if it meets the
specified criteria, it is re-included in the string for further processing. N-gram and feature
generation is handled the same as before.
The new function can then be called to generate the positive and negative review feature
as shown in Listing 11. This invocation of the script, using the default parameters, is the
equivalent of the initial script described in Section 9.2.2. Unsurprisingly the performance
results achieved and the most informative features obtained with this new script, and those
obtained with the original script, are exactly the same.
An extreme example of how the prepare_review function could be called is shown
in Listing 12. In this example punctuation characters are removed, the minimum allowed
word size is four characters, stop words are removed, words are stemmed and an n-gram
size 2 is used.
The results of this invocation of the script is actually quite promising with an over all
g-performance measure of 0.801, see Figure 9.8. Although better results were achieved by
some of the models explored earlier it is worth noting that the negative recall value of 0.73
achieved by this model was the best obtained using bigrams. There are endless potential
combinations to explore using this simple function.
The complete listing for the prepare review function, including additional looping and
variable definitions, is available for download from the book website.
9.3
Modelling with scikit-learn
The Natural Language Toolkit produced some promising models and was quite easy to
use. It did not, however, offer much choice or scope for different lines of investigation during
model development. Although additional functionality is made available to the developer
Text Classification Using Python
213
# Remove punctuation characters from the string
if removepunctuation:
review = re.sub(r’[ˆ\w\s]’,’’,review)
# Remove words less than minwordsize length
if (minwordsize > 1):
review = " ".join(
word for word in review.split()
if len(word) >= minwordsize
)
# Remove stopwords if required
if removestopwords:
review = " ".join(
[word for word in review.split()
if not word in stopwords.words(’english’)])
# Stemm words if required
if stemwords:
review = " ".join(
[stemmer.stem(word)
for word in review.split()])
# Prepare ngrams and Generate Features
if (ngramsize > 1):
# Bi-grams or Tri-grams are quired
review = [word for word in ngrams(
review.split(), ngramsize)]
# Generate Features
review = (word_feats(review), reviewtype)
else:
# Just Generate Features
review = (word_feats(review.split()), reviewtype)
reviews.append(list(review))
Listing 10: The details of the prepare review function
using NLTK through a wrapper for scikit-learn, in this section the NLTK models are left
behind and focus, instead, turns to the development of models using scikit-learn.
9.3.1
Developing a Naı̈ve Bayes scikit-learn Classifier
Although the development of the scikit-learn model follows a very similar flow to that
utilised for the NLTK model, scikit-learn could be considered as a more advanced toolkit.
As well as extensive data manipulation functionality, scikit-learn also provides a better
selection of learner algorithms though for the investigation, a Naı̈ve Bayes algorithm will
be used and then a Support Vector Machine learner is considered.
The first step in the process is to access the data. scikit-learn uses the load_file
method to import text files. Like NLTK, the subfolder names are used as category labels.
An object of datatype “bunch” is returned, which is a dictionary-type object. The Python
is shown in Listing 13.
Once created, the bunch object can be examined as shown. The data element holds the
text of each review.
214
Text Mining and Visualization: Case Studies Using Open-Source Tools
pos_feat = prepare_review(pos_ids)
neg_feat = prepare_review(neg_ids,
reviewtype=’neg’)
Listing 11: Calling the prepare review function with default parameter values
pos_feat = prepare_review(pos_ids,
removepunctuation=1,
minwordsize=4,
removestopwords=1,
stemwords=1,
ngramsize=2)
neg_feat = prepare_review(neg_ids,
removepunctuation=1,
minwordsize=4,
removestopwords=1,
stemwords=1,
ngramsize=2,
reviewtype=’neg’)
Listing 12: An extreme example of calling the prepare review function
>>> bunch.keys()
dict_keys([’target’, ’target_names’, ’filenames’, ’DESCR’, ’data’])
>>> bunch.target[0]
0
>>> bunch.target_names[0]
’neg’
>>> bunch.filenames[0]
’C:/nltk_data/corpora/movie_reviews/neg\\cv405_21868.txt’
The next step in the development of the model is to create the training and testing data
sets. scikit-learn provides a cross_validation.train_test_split operator that will
automatically split the data set into random training and testing subsets. The data and
target arrays, obtained from the bunch object created earlier, are the first two parameters
passed. The third parameter passed is the proportion of the data set to include in the test
set. In this case 20% of the data will be held back for the test data set. The final parameter
is the pseudo-random number generator state used for random sampling. A random number
in the range 1 to 100 is used for each invocation. By convention the returned arrays are
FIGURE 9.8: Performance of NLTK model using the prepare review function.
Text Classification Using Python
215
from sklearn.data sets import load_files
bunch = load_files(’C:/nltk_data/corpora/movie_reviews/’
Listing 13: Loading the movie review corpus using scikit-learn
named X train, X test, y train, y test where X train, X test are the data sets representing
the split data array and y train, y test are the data sets representing the split target array.
Listing 14 shows how this cross_validation.train_test_split function is used.
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn import cross_validation
import random
X_train, X_test, y_train, y_test \
= cross_validation.train_test_split(bunch.data,
bunch.target,
test_size=0.20,
random_state =
(random.randrange(1 ,100)))
Listing 14: Splitting the data into training and testing data sets
The training and test data sets produced by the cross-validation split function are passed
to a TfidfVectorizer function, which performs two main tasks. The first is to convert a
collection of text documents, in this case the training data, into a matrix that is a sparse
representation of token counts. The second task transforms the count matrix into a term
frequency inverse document frequency (TF-IDF) representation.
TF-IDF is a statistical measure of how important a given word is in a collection of
documents. Each document is represented as a vector of words, and each word is represented
in the vector by a value that is indicative of its importance in predicting the class of the
document. The TF-IDF weight for a word i in document j is given as
T F IDFij = T Fij · IDFi
(9.7)
where TF is a measure of a words importance in a document and is calculated as
T Fi̇j =
nij
Σk nkj
(9.8)
where the number of times a word i appears in a document j is represented by nij , and Σk nkj
is a count of all words in document j. This means that the more times a word appears in a
document the larger its value for TF will get. The TF weighting of a word in a document
shows its importance within that single document. IDF then shows the importance of a
word within the entire collection of documents or corpus and is calculated as
IDFi = log
|P |
| {pj : ti pj } |
(9.9)
where | P | is the total number of documents in the corpus or collection, and | {pj : ti pj } |
represents the number of documents in which the word, or term, ti appears. The nature of
the IDF value is such that terms appearing in a lot of documents will have a lower score or
216
Text Mining and Visualization: Case Studies Using Open-Source Tools
weight. This means terms that appear in only a single document, or in a small percentage
of the documents, will receive a higher score. This higher score makes that word a good
discriminator between reviews.
In all, there are over 20 parameters that can be used to customise and fine tune the
performance of this operator, but for our purposes here only the following will be examined:
1. ngram range: tuple (min n, max n)
The lower and upper boundary of the range of n-values for different n-grams to be
extracted. All values of n such that minn <= n <= maxn are used.
2. stop words: string ’english’, list, or None (default)
english is currently the only supported string value and this has the affect of removing English stop words.
3. tokenizer: callable or None (default)
Override the string tokenization step while preserving the preprocessing and n-grams
generation steps.
Other parameters allow preprocessing of the corpus and more advanced control of the
TF-IDF calculation and are beyond the scope of this text.
In order to closely manage the tokenisation of the review text we will develop our own
tokenize along similar lines of the prepare_review function seen earlier. The tokenizer
performs four steps. It first separates each review into a list of individual tokens before
removing punctuation, removing purely numerical tokens before finally removing tokens
less than the specified length. It would be a simple matter to add an additional steps,
stemming for example, if required. The new step to remove purely numerical tokens, 123
for example, was added as it was seen that a lot of the most informative negative features
contained numerical values. More on this later.
In Listing 15, the TF-IDF Vectorizer object is created. In this example unigrams, bigrams, and trigrams are considered, and English stop words are removed. The tokenizer
just created is utilised. Whereas the training data are transformed into a document/token
matrix and the TF-IDF of each token is calculated, the test data are transformed into the
sparse matrix format only for use later with the generated model.
tfidf = TfidfVectorizer(tokenizer=tokenize,
stop_words=’english’,
ngram_range=(1, 3))
X_train = tfidf.fit_transform(X_train)
X_test = tfidf.transform(X_test)
Listing 15: Create TF/IDF vectorizer object
The final step is to fit the model using the training data and then to predict the classes
of the testing data. Listing 16 shows the creation and fitting of a Multinomial Naı̈ve Bayes
model followed by the class prediction for the training data set. The predicted class for
each of the test samples is stored in the y_predicted array and can easily be manually
compared to the actual class values.
>>> y_predicted[:20]
array([1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 1])
>>> y_test[:20]
array([1, 1, 0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 0, 1, 0, 0, 0, 1, 1, 1])
Text Classification Using Python
217
from
sklearn.naive_bayes import MultinomialNB
classifier = MultinomialNB(alpha=.01).fit(X_train, y_train)
y_predicted = classifier.predict(X_test)
Listing 16: Fit the model and classify the training data
The scikit-learn metrics package makes it very easy to access the performance measurements of the generated model. Listing 17 shows how the classification_report
function is used to show the precision, recall, and F1 Measurement values achieved for each
class and Figure 9.9 shows the results displayed. The confusion matrix is also calculated. A
confusion matrix is a table that describes the performance of a classification model on a set
of test data for which the true values are known.
from sklearn.metrics import classification_report
print (’\nHere is the classification report:’)
print (classification_report(y_test, y_predicted))
print (’\nHere is the confusion matrix:’)
print (metrics.confusion_matrix(y_test, y_predicted))
Listing 17: Using scikit-learn classification report
FIGURE 9.9: Performance of the first Naı̈ve Bayes scikit-learn model.
From the classification report the g-performance of this model can be calculated. Using
the recall results for the negative review class, 0, and
√ the positive class, 1 the g-peformance
for this invocation of the model is calculated as 0.78 ∗ 0.80 = 0.7899. Examining the
confusion matrix further, if we view it as a table of dimensions 2x2, we can name each of
the rows, columns and cells as follows. To top row can be named “actual positive”, and the
bottom row “actual negative”. The left column can be named “predicted positive”, and the
right column “predicted negative”.
Getting the most informative features for this scikit-learn model is a bit more involved
than the simple method provided by NLTK. A show_most_informative_features
function is provided in the complete script available for download from the book website.
In the sample shown the most negative informative features are shown on the left hand side
and the most informative positive features are shown on the right hand side.
218
Text Mining and Visualization: Case Studies Using Open-Source Tools
TABLE 9.2: Confusion matrix for a Naı̈ve Bayes scikit-learn classifier.
predicted negative predicted positive
actual negative
actual positive
163
38
46
153
FIGURE 9.10: Naı̈ve Bayes scikit-learn most informative features.
9.3.2
Developing a Support Vector Machine scikit-learn Classifier
Developing a SVM model requires only a simple two line change to the script just
described. Instead in instantiating a Naı̈ve Bayes classifier object, a Linear Support Vector
classifier is used instead.
# Create the Support Vector Machine model
from sklearn.svm import LinearSVC
classifier = LinearSVC().fit(X_train, y_train)
Listing 18: Fit the model and classify the training data
The rest of the script can remain unchanged. The listing available for download from the
book website actually contains both options, but the LinearSVC lines are commented out.
Figures 9.11 and 9.12 show the performance and most informative features for this model.
The g-performance value this SVM model, 0.840, was the best achieved by any model.
Reviewing the top most informative features for the negative class it is very conceivable
that worst, boring, and stupid are words you would not be surprised to see in a
negative movie review. Likewise, great and excellent are words that you would expect
to see in a positive movie review.
FIGURE 9.11: Performance of the SVM scikit-learn model.
Text Classification Using Python
219
FIGURE 9.12: SVM scikit-learn model most informative features.
9.4
Conclusions
This chapter was a whistlestop tour through binary text classification using Python,
NLTK, and scikit-learn. Although testing was not exhaustive, it was found that the scikitlearn Linear SVC classifier provided results that were slightly better than NLTK and scikitlearn Naı̈ve Bayes.
Only the surface of these tools has been scratched and hopefully this chapter encourages
you to explore each tool in detail.
Bibliography
[1] The official home of the python programming language.
[2] The natural language toolkit.
[3] The official home of numpy: The fundamental package for scientific computing with
python.
[4] The official home of scipy: Scientific computing tools for python.
[5] The home of scikit-learn: Machine learning in python.
[6] Bo Pang and Lillian Lee. A sentimental education: Sentiment analysis using subjectivity
summarization based on minimum cuts. In Proceedings of the ACL, 2004.
[7] Miroslav Kubat and Stan Matwin. Addressing the curse of imbalanced training sets:
One-sided selection. In In Proceedings of the Fourteenth International Conference on
Machine Learning, page 179186. Morgan Kaufmann, 1997.
[8] Julie B. Lovins. Development of a Stemming Algorithm. Mechanical Translation and
Computational Linguistics, 11, June 1968.
This page intentionally left blank
Part IV
R
221
This page intentionally left blank
Chapter 10
Sentiment Analysis of Stock Market
Behavior from Twitter Using the R Tool
Nuno Oliveira
ALGORITMI Research Centre, University of Minho, Guimarães, Portugal
Paulo Cortez
ALGORITMI Research Centre/Department of Information Systems, University of Minho,
Guimarães, Portugal
Nelson Areal
Department of Management, University of Minho, Braga, Portugal
10.1
10.2
10.3
10.4
10.5
10.6
10.7
10.8
Abstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Preprocessing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Inclusion of Emoticons and Hashtags Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
223
223
225
225
227
230
233
236
238
238
Abstract
Sentiment analysis is a hot research topic with widespread applications in organizations.
Recently, it has been applied to create investor sentiment indicators that can be utilized in
the prediction of stock market variables such as returns, volatility and trading volume.
This chapter presents a use case about sentiment analysis in tweets related to stock market using the open-source R tool. We provide and explain code for the execution of diverse
tasks, such as tweets collection, preprocessing tasks, utilization of opinion lexicons, application of machine learning (ML) methods, and respective evaluation. A data set composed
by stock market tweets is also provided.
This use case shows that the utilization of ML techniques using manually classified
training data outperforms some standard sentiment word lists when applied to test data.
223
224
Text Mining and Visualization: Case Studies Using Open-Source Tools
10.1
Introduction
Sentiment analysis (SA) is the computational treatment of opinion, sentiment, and subjectivity in text [1]. In recent years, social media platforms, and in particular microblogging,
have enabled an explosion of unstructured content with opinions regarding a huge variety
of topics and have provided a valuable and inexpensive source for opinion analysis about
organizations and individuals. The analysis of all these data is not feasible without the use
of automated computation. The identification and summarization of important information in vast amounts of data is very challenging for the average person [2] and the current
amount of social media data makes human processing impracticable. These limitations can
be overcome by SA systems that mine large amounts of opinionated content and automatically extract, summarize and classify the opinions about an entity (e.g., product, event,
organization) [1]. SA may assign diverse sentiment categories (e.g., positive, negative, or
neutral) or attribute a sentiment score. These systems have widespread applications, such
as reputation management, sales prediction, stock management, ad placements, products,
and services benchmarking.
Diverse machine learning (ML) methods can be applied to learn the correct sentiment
value from a set of input features such as the term frequency of unigrams or bigrams,
emoticons, or hashtags. Whenever there is insufficient labeled data, the utilization of existing opinion lexicons may allow unsupervised SA to be performed. Opinion lexicons are
composed of a list of opinion words and the respective sentiment value (positive, negative,
neutral, or a sentiment score). The presence of opinion words in text permits the discernment of the sentiment orientation. For example, the sentence “the computer is good” can
be easily classified as positive if the opinion lexicon contains the word “good” as positive. A
common text representation is the bag-of-words , which represents the text as a collection
of lexical items and ignoring the syntactic structure of sentences. However, particularly in
short documents, a more rigorous linguistic approach (e.g., sentiment composition) that
considers the interaction between sentence constituents (e.g., negation) may permit better
results [3].
SA can be performed at diverse levels such as document, sentence, and aspect levels.
Document-level classification is a common approach. It classifies the sentiment of the whole
document assuming that the document expresses opinions on a single entity (e.g., customer
reviews) [4]. However, it is too coarse for some applications because some documents may
include opinions regarding diverse entities or the applications may require the classification
regarding various aspects of the entity.
The extraction of investor sentiment indicators from social media data in order to forecast stock market behavior is a very recent research topic (e.g., [5, 6, 7]). These studies
suggest that a model that accounts for investor sentiment can potentially be used to predict
key stock market variables, such as returns, volatility, and volume. Moreover, it has been
demonstrated that individuals’ financial decisions are affected by their emotions and mood
[8]. Microblogging is a social medium particularly suitable for the creation of investors’ sentiment indicators. There are several arguments recommending the usage of microblogging
data:
• The analysis of investors microblogging interactions can provide reliable indicators of
their sentiment in a faster and less expensive manner than traditional sources (e.g.,
large-scale surveys)
• The community of investors that utilizes these services has grown and is potentially
more representative
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
225
• Users react promptly to events allowing a near real time sentiment assessment
• The character constraints require greater objectivity from the author
Diverse studies confirm the informative value of microblogging data for forecasting some
stock market variables. Sentiment indicators extracted from microblogging data seems to
have predictive value for future market directions [6, 9] and returns [10] and posting volume
can improve the forecasts of trading volume [10, 11, 12].
In this chapter, we present a use case about SA in microblogging conversations related to stock market using the open-source R tool. We have significant experience in this
topic because we are working in a related project. In the PhD work “Mining Social Media
Sentiment to Forecast Stock Market Behavior”, we are collecting social media data (e.g.,
Twitter, StockTwits, message boards) to create sentiment and attention indicators that
can be utilized in the forecasting of stock market behavior (e.g., volatility, trading volume, returns). We present and explain some code for diverse steps comprising this process,
such as tweets collection, preprocessing tasks, utilization of opinion lexicons, application of
ML methods, and respective evaluation. A data set composed by stock market messages
from the most popular microblogging service (i.e., Twitter) is provided for this use case.
This data set and the R files containing all code utilized in this example are available at
http://www.text-mining-book.com.
10.2
Methods
R is a free language and environment for statistical computing and graphics [13] that
allows the execution of the various tasks required by this topic. It provides a wide diversity
of functionalities, and it is easily extensible by applying the diverse available packages. For
instance, the package twitteR provides an interface to the Twitter web API enabling the
collection of a large number of messages filtered by search queries. Additionally, R provides a
rich set of Natural Language Processing (NLP) packages that are very useful to perform SA.
Packages such as tm [14], RTextTools [15], SnowballC ([16]), wordnet [17] or openNLP
(http://opennlp.apache.org/) allows diverse NLP operations to be performed such
as stemming, Part of Speech tagging, tokenization, access to lexical databases, easy creation
of Document Term matrices, and application of ML methods.
The Comprehensive R Archive Network (http://cran.r-project.org/) is a large
repository containing up-to-date code and documentation for R. For example, an exhaustive
list of packages is presented and described in this platform, and the installation files of
current Windows, Linux, or Mac OS X versions of R can be downloaded from its webpage.
There are diverse forms to install packages, but an usual and simple manner is to execute
the function install.packages. For instance, the RTextTools package can be easily
installed by executing the command install.packages("RTextTools") directly from
the R console. This function receives a vector of packages names, downloads these packages
from the repositories, and installs them. The library function permits the utilization of
installed packages (e.g., library("RTextTools")). The R online documentation can be
accessed by using the command help.start(). Further details about the R language can
be found in [18].
226
10.3
Text Mining and Visualization: Case Studies Using Open-Source Tools
Data Collection
To be able to collect Twitter data, it is necessary to sign up on Twitter (https://
twitter.com/signup) and create a Twitter application (https://apps.twitter.
com/app/new). This account will provide the required information (e.g., consumer
key, consumer secret) to connect to the Twitter API. The package twitteR [19] provides the R interface to Twitter API. A data collection example is coded in file
tweets collection.R.
### tweets_collection.R file ###
# load package twitteR: provides interface to Twitter API
library (twitteR)
# download cacert.pem
download.file(url=’http://curl.haxx.se/ca/cacert.pem’, destfile=’cacert.pem’)
# assign your Twitter application data to R variables
# request token URL
reqURL <- "https://api.twitter.com/oauth/request_token"
# access token URL
accessURL <- "https://api.twitter.com/oauth/access_token"
# authorize URL
authURL <- "https://api.twitter.com/oauth/authorize"
# consumer key
consumerKey <- "your_consumer_key"
# consumer secret
consumerSecret <- "your_consumer_secret"
# create a OAuth object to access to Twitter API
twitter_cred <- OAuthFactory$new(consumerKey=consumerKey,
,→
consumerSecret=consumerSecret, requestURL=reqURL, accessURL=accessURL,
,→
authURL=authURL)
# execute handshake method to enable connection to Twitter API
# (you must open the provided URL in a web browser and write the code in the
,→
Console)
twitter_cred$handshake(cainfo = system.file(’CurlSSL’, ’cacert.pem’, package =
,→
’RCurl’))
# save the OAuth object for future sessions
save(twitter_cred, file="twitter_auth.Rdata")
# load the OAuth object
load("twitter_auth.Rdata")
# register OAuth credentials to Twitter R session
registerTwitterOAuth(twitter_cred)
# tweets collection example: last 15 tweets containing the word GOOGLE
tweets_list <- searchTwitter("GOOGLE", cainfo=’cacert.pem’, n=15)
# each tweets_list element contains several attributes such as:
print(tweets_list[[1]]$id) # tweet id
print(tweets_list[[1]]$text) # tweet text content
print(tweets_list[[1]]$screenName) # user name
print(tweets_list[[1]]$created) # creation time
The access to Twitter API requires OAuth authorization. This authentication demands
valid Twitter application data (i.e., consumer key, consumer secret, request URL, access
token URL, and authorize URL) and "cacert.pem" file, a collection of trusted root
certification authorities. Therefore, Twitter application data is assigned to R variables
for posterior operations and download.file command downloads "cacert.pem" file.
OAuth authorization is obtained by the creation of an OAuth object (OAuthFactory$new
command) and the execution of the respective handshake method. Nevertheless, a manual task is needed to enable Twitter API connection. It is necessary to open an
URL that appears in the console after running the handshake method, then authorize the application and write the provided code in the R console. This inconve-
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
227
nient last step can be avoided in subsequent automated use by loading a R object
with a completed handshake. Therefore, the OAuth object is saved after the successful handshake execution. After loading the OAuth object, the registerTwitterOAuth
function registers the OAuth credentials to the Twitter R session and permits
the start of data collection. Tweet gathering is done with the searchTwitter
function. Diverse search criteria can be included in this function such as search
query (e.g., searchString="GOOGLE"), time interval (e.g., since=2011-03-01,
until=2011-03-02), geocode (e.g., geocode=41.365,-51.1151121,12mi), and
number of tweets (n=15). In this example, the last 15 tweets containing the word
“GOOGLE” are retrieved. The result is a list of status objects composed by diverse attributes such as tweet id, text, creation date, and user name. Tweets text shall be converted
to ASCII format because other encodings may compromise code execution.
10.4
Data Preprocessing
This use case provides a data set with 5,000 tweets containing cashtags of a comprehensive set of U.S. stocks (e.g., $AAPL, $GOOG, $GS). Cashtags are composed by the
stock ticker preceded by the “$” symbol. These symbols are commonly used by the investor
community in discussions related to the respective security. Concentrating on only these
messages results in a less noisy data set. Tweets were collected using the Twitter API between January 1, 2013, and March 31, 2013, and were manually classified as positive or
negative. A alternative to this manual classification would be the use of lexicons (as explained in the Introduction). However, we opted for a manual classification given that we
also want to test in this work the quality of two lexicons within the financial domain (as
describe in Section 10.5). These tweets were manually classified by MSc students from the
area of Finance, under the supervision of co-author Nelson Areal (expert in finance). For
privacy reasons, all mentions and URL addresses were normalized to “@user” and “URL”,
respectively. There are three attributes for each message: creation date, text content, and
the assigned sentiment label.
Preprocessing is an essential part of text mining because it prepares input data for
subsequent analysis. For instance, raw text might be formatted inadequately (e.g., HTML
format) or contain many irrelevant terms (e.g., “the”, “a”). Thus, preprocessing tasks such
as data import, cleaning, and structuring provide an improved input for later Text Mining
steps (e.g., application of ML methods) [14]. The file preprocessing.R contains the code
for data preprocessing.
### preprocessing.R file ###
# load R package tm: framework for text mining applications
library(tm)
# load R package RTextTools: Automatic Text Classification via Supervised Learning
library(RTextTools)
# read a CSV file containing tweets about stock market
twt<-read.csv("twt_sample.csv",stringsAsFactors=FALSE)
# convert "created_at" column from character to time
twt$created_at<-strptime(twt$created_at,"%Y-%m-%d %H:%M:%S")
# order messages by creation time
twt<-twt[order(twt$created_at),]
# preprocess some usual expressions in stock market microblog conversations
,→
(punctuation removal will eliminate these expressions)
228
Text Mining and Visualization: Case Studies Using Open-Source Tools
# substitute the isolated "@" by "at"
twt$text<-gsub("(ˆ|[[:space:]])(@)([[:space:]])","\\1at\\3",twt$text)
# transform @<number> into "at <number>":
twt$text<-gsub("(ˆ|[[:space:]])(@)([$]?[0-9.]+)([[:space:]]|[[:punct:]]|$)","\\1at
,→
\\3\\4",twt$text)
# substitute "+<number>" by "more <number>"
twt$text<-gsub("(ˆ|[[:space:]])([+])([$]?[.]?[[:digit:]]+)","\\1more
,→
\\3",twt$text)
# substitute "-<number>" by "less <number>"
twt$text<-gsub("(ˆ|[[:space:]])([-])([$]?[.]?[[:digit:]]+)","\\1less
,→
\\3",twt$text)
# substitute "˜<number>" by "nearly <number>"
twt$text<-gsub("(ˆ|[[:space:]])([˜])([[:space:]]?)([$]?[.]?[[:digit:]]+)","\\1nearly
,→
\\4",twt$text)
# substitute "#<number>" by "number <number>"
twt$text<-gsub("(ˆ|[[:space:]])([#])([[:space:]]?)(([[:digit:]])+)","\\1number
,→
\\4",twt$text)
# substitute "<number>%" by "<number> percent"
twt$text<-gsub("([[:digit:]])([%])","\\1 percent",twt$text)
# substitute all numbers by the word "num"
twt$text<-gsub("[$]?[[:digit:]]*[.]?[[:digit:]]+","num",twt$text)
# substitute cashtags by the word "cshtg"
twt$text<-gsub("[$][[:alpha:]]+","cshtg",twt$text)
# create a list containing the emoticons of each tweet
m<-gregexpr("(ˆ|[[:space:]])(([<>]?[:;=][-o*’]?([)]|[(]|[]]|[[]|[dDpP}]|[/:{|\\]))|
,→
(([)]|[]]|[(]|[[]|[dDpP}]|[/:{|\\])[-o*’]?[:;=][<>]?))($|[[:space:]])",twt$text)
l_emot<-regmatches(twt$text,m)
# remove these emoticons from text
twt$text<-gsub("(ˆ|[[:space:]])(([<>]?[:;=][-o*’]?([)]|[(]|[]]|[[]|[dDpP}]|[/:{|\\]))|
,→
(([)]|[]]|[(]|[[]|[dDpP}]|[/:{|\\])[-o*’]?[:;=][<>]?))($|[[:space:]])","\\1
,→
",twt$text)
# create a list containing the hashtags of each tweet
m<-gregexpr("#\\w+",twt$text)
l_hstg<-regmatches(twt$text,m)
# remove hashtags from text
twt$text<-gsub("#\\w+","",twt$text)
# creation of a Document Term Matrix with stemming, conversion to lowercase,
,→
removal of punctuation and sparse terms and utilization of "Term frequency"
,→
weighting for unigrams
DTM <- create_matrix(twt$text, language="english", removeStopwords=FALSE,
,→
removePunctuation=TRUE, removeSparseTerms=0.99, toLower=TRUE, stemWords=TRUE,
,→
weighting=weightTf)
# save R objects required for future tasks
save(list=c("DTM", "twt", "l_emot", "l_hstg"), file="preproc.RData")
The tm [14] and RTextTools [15] packages are loaded by the library function. tm is
a framework for text mining applications and RTextTools is a ML package for automatic
text classification. The function read.csv reads a CSV file and creates a data frame (R
matrix object with column names) from it. In this situation, read.csv imports the Twitter
data set.
Diverse operations are executed to reduce dimensionality, such as:
• Convert all words to lowercase
• Perform text stemming , which reduces each word to its stem (e.g., the stem of words
“waiting” and “waited” is “wait”)
• Remove punctuation and very infrequent terms (i.e., sparse terms)
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
229
• Normalize all numbers and cashtags to “num” and “cshtg”
Punctuation and sparse terms removal, stemming, and lowercase conversion are accomplished in the Document Term matrix creation. These matrices contain the frequency of
terms in documents and are commonly used in NLP. Rows correspond to documents (e.g.,
tweets), columns correspond to terms (e.g., unigrams), and each entry contains the frequency
of the corresponding term in the respective document.
However, punctuation removal may eliminate informative items such as emoticons and
hashtags. Moreover, there are usual expressions in stock market microblogging conversations containing punctuation symbols (e.g., "@<number>" means “at <number>”, and
"+<number>" signifies “more <number>”). Thus, some operations are executed in this
demonstration before the creation of the Document Term matrix. Two lists holding emoticons and hashtags are created and some expressions (e.g., @<number>, +<number>, <number>, #<number>) are substituted by identical expressions (e.g., at <number>,
more <number>, less <number>, number <number>). For instance, these operations will
convert the original message “STO Mar 425 put @ +1.98, $AAPL will fill gap to $500.”
into “STO Mar num put at more num, cshtg will fill gap to num.”.
Regular expressions are an important concept for strings matching, manipulation,
and extraction. They are composed by regular characters and metacharacters with special meaning that permit the description of a set of strings. For instance, the regular expression "a[[:digit:]]" will match all patterns having the letter “a” immediately followed by a digit (e.g., "a1", "a9"). All functions utilized in this example
to substitute, eliminate, and extract expressions receive regular expressions as patterns
to match. The gsub function is applied in all expression modifications, and its inputs
are the regular expressions of items to modify (e.g., "[[:digit:]]+" for numbers, "
[$][[:alpha:]]+" for cashtags), the replacement strings (e.g., "num", "cshtg"), and
the character vector where these alterations take place (i.e., tweets text). The creation
of lists of emoticons and hashtags is done with gregexpr and regmatches functions.
The gregexpr command receives a character string containing regular expressions to
be matched (i.e., "#\\w+" for hashtags) and a character vector where these expressions
are browsed (i.e., tweets text) and returns a list containing the positions of every match
in text. The regmatches command extracts these matches from text, taking the list
produced by the gregexpr command and respective tweets text as inputs. The regular
expressions to match emoticons were adopted from Christopher Potts’s tokenizing script
(http://sentiment.christopherpotts.net/tokenizing.html#emoticons).
The create.matrix function creates a Document Term matrix from a collection of
documents, including the following arguments:
• language - the language used for stemming
• removePunctuation - the logical parameter specifying whether to remove punctuation
• removeSparseTerms - the maximal allowed sparsity
• removeStopwords - logical parameter defining whether to remove stopwords
• stemWords - logical parameter indicating whether to stem words
• toLower - logical parameter specifying whether to convert all text to lowercase
• weighting - weighting function to calculate each entry value (e.g., weightTf for term
frequency)
230
Text Mining and Visualization: Case Studies Using Open-Source Tools
The first Document Term matrix of this demonstration performs text stemming for
English language, converts all text to lowercase, removes punctuation and terms absent in
more than 99% of all tweets, maintains stopwords, and contains the raw term frequencies
in each entry. For example, after performing all preprocessing tasks, the original tweet
“$SBUX buy when it clears resistance at 54.5” will have nonzero term frequency in the
columns corresponding to terms “buy”, “cshtg”, “num”, “resist” and “when”. The save
command saves R objects required for future tasks.
10.5
Sentiment Analysis
We explore two types of SA: supervised classification performed by ML methods and
unsupervised classification by applying opinion lexicons. The ML methods utilized in this
demonstration are Support Vector Machines (SVM) and Naı̈ve Bayes (NB). They represent
two common and distinct methods applied for text classification. SVM is a discriminative
classifier, whereas NB is a generative classifier. A discriminative classifier supplies a model
for the target variable conditional on the observed variables, whereas a generative model
is a full probabilistic model of all variables, specifying the joint probability distribution of
target and observed variables [20].
Lexicon-based SA is supported by existing opinion lexicons and permits the analysis
to be performed in an unsupervised manner, thus avoiding the laborious task of classifying data manually. In this approach, each tweet is classified by considering the “positive”
and “negative” words that it contains, according to each lexicon. This demonstration uses
Harvard General Inquirer (GI) [21], a popular domain independent lexicon, and a financial
lexical resource (FIN) created by Loughran and McDonald [22]. GI lexicon comprises more
than 11,000 words classified in 182 categories. These categories come from four sources: the
Harvard IV-4 dictionary; the Lasswell value dictionary; categories recently constructed, and
“marker” categories containing syntactic and semantic markers for disambiguation. This example applies all words of the “positiv” and “negativ” categories. FIN resources contain
6 word lists commonly applied in financial text documents: negative, positive, uncertainty,
litigious, modal strong and modal weak words. This demonstration utilize the negative and
positive word lists because the other lists do not have a clear sentiment polarity . The file
proc lexicons.R codes the download and preparation of these lexical resources to be
applied in SA.
### proc_lexicons.R file ###
library(tm) # load R package tm: framework for text mining applications
library(SnowballC) # load SnowballC: An R interface to the C libstemmer library
,→
that implements Porter’s word stemming algorithm
library(XLConnect) # load XLConnect package: Manipulation of Excel files
# process financial lexicon (FIN) created by Loughran and McDonald
FIN_n<-read.csv("http://www3.nd.edu/˜mcdonald/Data/Finance_Word_Lists/LoughranMcDonald_
,→
Negative.csv", stringsAsFactors = FALSE)[,1] # read negative word list
FIN_p<-read.csv("http://www3.nd.edu/˜mcdonald/Data/Finance_Word_Lists/LoughranMcDonald_
,→
Positive.csv", stringsAsFactors = FALSE)[,1] # read positive word list
# create a data frame with all words and respective sentiment label (1 for
,→
positive and -1 for negative)
lex_FIN<-as.data.frame(rbind(cbind(FIN_p,rep(1,length(FIN_p))),
,→
cbind(FIN_n,rep(-1,length(FIN_n)))),stringsAsFactors=FALSE)
colnames(lex_FIN)<-c("stem","sentiment")
lex_FIN$stem<-tolower(lex_FIN$stem) # convert to lowercase
lex_FIN$stem<-stemDocument(lex_FIN$stem) # stem all words
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
231
lex_FIN<-unique(lex_FIN) # remove duplicated elements
# write the processed FIN lexicon to a CSV file
write.csv(lex_FIN,"FIN.csv",row.names=FALSE)
# process Harvard General Inquirer (GI) lexicon
download.file(url="http://www.wjh.harvard.edu/˜inquirer/inquirerbasic.xls",
,→
destfile="inquirerbasic.xls", mode="wb") # download GI lexicon
GI_book = loadWorkbook("inquirerbasic.xls") # load XLS file containing GI lexicon
GI_sheet = readWorksheet(GI_book,sheet=1,header=TRUE) # read lexicon sheet
# select all words from "Negativ" category
GI_n<-GI_sheet$Entry[!is.na(GI_sheet$Negativ)]
# select all words from "Positiv" category
GI_p<-GI_sheet$Entry[!is.na(GI_sheet$Positiv)]
# create a data frame with all words and respective sentiment label (1 for
,→
positive and -1 for negative)
lex_GI<-as.data.frame(rbind(cbind(GI_p,rep(1,length(GI_p))),
,→
cbind(GI_n,rep(-1,length(GI_n)))),stringsAsFactors=FALSE)
colnames(lex_GI)<-c("stem","sentiment")
lex_GI$stem<-tolower(lex_GI$stem) # convert to lowercase
# process repeated words (e.g., ABOUT#1, ABOUT#2)
lex_GI$stem<-gsub("#(.*)$","",lex_GI$stem)
lex_GI$stem<-stemDocument(lex_GI$stem) # stem all words
lex_GI<-unique(lex_GI) # remove duplicated elements
# exclude stems appearing in both lists
lex_GI<-lex_GI[!(lex_GI$stem %in% lex_GI$stem[duplicated(lex_GI$stem)]),]
# write the processed GI lexicon to a CSV file
write.csv(lex_GI,"GI.csv",row.names=FALSE)
FIN word lists are available in CSV format at http://www3.nd.edu/˜mcdonald/
Word_Lists.html. The read.csv function reads the positive and negative word
lists from their URL. GI lexicon can be downloaded in different formats at http:
//www.wjh.harvard.edu/˜inquirer/spreadsheet_guide.htm. In this example,
the download.file function retrieves the XLS version. XLConnect [23] package permits
the production of XLS files. Its loadWorkbook and readWorksheet functions allow the
loading of the GI workbook file and reading of the corresponding sheet. For each lexicon, a data frame containing all words and respective sentiment value (i.e., 1 for positive
and −1 for negative) is created. The cbind and rbind functions combines by columns
or rows, a series of vectors, matrices or data frames. In this case, the utilization of these
functions permits the production of a matrix containing lexical items and corresponding sentiment as columns. The as.data.frame function converts this matrix into a data frame.
Both lexicons have their words stemmed (stemDocument function), converted to lowercase (tolower function) and repeated items removed (unique function). SnowballC [16]
package implements Porter’s word stemming algorithm for collapsing words to a common
root. GI lexicon has repeated words represented by the word followed by a “#” character
and a sequential number (e.g., ABOUT#1, ABOUT#2). The gsub function removes the “#”
character and sequential numbers and maintains the word for every GI item. After this
operation and text stemming, GI lexicon has 39 items having both positive and negative
labels. These item are excluded from GI lexicon in this demonstration. The write.csv
function writes the processed lexicons to CSV files.
SA in Twitter messages about stock market is coded in file sentim analysis.R.
### sentim_analysis.R file ###
library(tm) # load R package tm: framework for text mining applications
# load R package RTextTools: Automatic Text Classification via Supervised Learning
library(RTextTools)
library(rminer) # load R package rminer: Application of Data Mining Methods
# load R package slam: Data structures and algorithms for sparse arrays and
,→
matrices
library(slam)
232
Text Mining and Visualization: Case Studies Using Open-Source Tools
load("preproc.RData") # load preprocessing objects
total<-nrow(twt) # total number of tweets
n_training<-round(0.75*total) # number of messages of the training set (first 75%)
### machine learning methods for sentiment analysis
# application of Support Vector Machines (SVM) method
# creation of a container using the Document Term Matrix, the sentiment labels and
,→
the definition of the training and test set size
container <- create_container(DTM,twt$sentiment,trainSize=1:n_training,
,→
testSize=(n_training+1):total, virgin=FALSE)
svm_model <- train_model(container,"SVM") # train SVM model
# predict test set using SVM model
svm_results <- classify_model(container,svm_model)
# application of Na\"{i}ve Bayes (NB) method
# convert Document Term matrix to data.frame
df<-as.data.frame(as.matrix(DTM),stringsAsFactors=FALSE)
rownames(df)<-NULL
df<-cbind(df,as.factor(twt$sentiment)) # add a sentiment column
colnames(df)[ncol(df)]<-"sentiment"
nb_model=fit(sentiment˜.,df[1:n_training,],model="naivebayes",task="class") #
,→
train NB model
# predict test set using NB model
nb_results<-predict(nb_model,df[(n_training+1):total,])
### unsupervised sentiment classification
# read CSV file containing the processed GI lexicon
lex_GI<-read.csv("GI.csv",stringsAsFactors=FALSE)
# read CSV file containing the processed FIN lexicon
lex_FIN<-read.csv("FIN.csv",stringsAsFactors=FALSE)
# creation of a Document-Text matrix with:
#
- minimum length of letters = 2 (minimum length of GI and FIN words)
#
- without the removal of sparse words to match all GI and FIN items
DTM_lex <- create_matrix(twt$text, language="english", removeStopwords=FALSE,
,→
minWordLength=2, removePunctuation=TRUE, toLower=TRUE, stemWords=TRUE,
,→
weighting=weightTf)
# creation of a Document-Text matrix containing only GI terms
DTM_GI<-DTM_lex[,which(colnames(DTM_lex) %in% lex_GI$stem)]
# creation of a Document-Text matrix containing only FIN elements
DTM_FIN<-DTM_lex[,which(colnames(DTM_lex) %in% lex_FIN$stem)]
# exclude GI elements absent from all tweets
l_GI<-lex_GI[lex_GI$stem %in% colnames(DTM_GI),]
# exclude FIN elements absent from all tweets
l_FIN<-lex_FIN[lex_FIN$stem %in% colnames(DTM_FIN),]
# order GI lexicon (Document-Text matrix columns are ordered alphabetically)
l_GI<-l_GI[order(l_GI$stem),]
# order FIN lexicon (Document-Text matrix columns are ordered alphabetically)
l_FIN<-l_FIN[order(l_FIN$stem),]
# score each message by summing each word sentiment in that tweet
#
word sentiment = number of occurrences of the word * word
,→
sentiment (1 for positive,
#
-1 for negative, 0 for neutral ou not present in
,→
lexicon)
total_GI<-tcrossprod_simple_triplet_matrix(DTM_GI, t(l_GI$sentiment))
total_FIN<-tcrossprod_simple_triplet_matrix(DTM_FIN, t(l_FIN$sentiment))
# tweets with positive score are labeled positive; tweets with negative score are
,→
labeled negative
total_GI[total_GI>0]<-"positive"
total_GI[total_GI<0]<-"negative"
total_FIN[total_FIN>0]<-"positive"
total_FIN[total_FIN<0]<-"negative"
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
233
# save R objects required for future tasks
save(list=c("total", "n_training", "svm_results", "nb_results", "total_GI",
,→
"total_FIN", "DTM_GI", "DTM_FIN"), file="sent_analysis.RData")
The load function reads preprocessed data. In this example, ML models are created in
the training set composed by the first 75% messages, whereas the evaluation of all methods is performed in the test set containing the remaining 25% tweets. The 75%–25% is a
commonly adopted holdout split for evaluating the quality of the ML predictions. Also,
only one run is executed for each ML model in order to reduce the computational effort
required by the example. However, a more robust validation could be achieved by executing
multiple runs and averaging the individual performances and by testing different holdout
splits. RTextTools package permits easy training (train model function) and classification of (classify model function) SVM models. Both functions receive a container of
class matrix container. The create container function creates a container from a
Document Term matrix , the corresponding sentiment labels and the size of training and
test set. However, this package does not include the NB algorithm. Therefore, the fit
function from rminer package [24] is applied to train the NB classifier. This function does
not accept a matrix container or a DocumentTermMatrix object as input, so the
Document Term matrix is converted to a data frame and the respective sentiment labels
are added to this object. The classification is performed by the predict function, taking
the created NB model and the test set as inputs. To simplify the R code, we adopted the
default parameters of the SVM and NB models, which are more likely to be used by non
ML experts and that achieved a good classification result in this example data set. Nevertheless, the best parameters of the ML models could be searched using the training data
and R functions such as tune from e1071 package [25].
In the lexicon-based models, a message is considered: positive, if the number of “positive” words is higher than the number of “negative” words; negative, if the number of
“negative” words is higher than the number of “positive” words; and neutral, if the number
of both word polarity types is equal. In order to match all GI and FIN items, a Document
Term matrix is created with a lower minimum number of letters and without the elimination of sparse terms. From this matrix a distinct matrix is produced for each lexicon
composed only by its terms. Sentiment scores are calculated by the multiplication of the
Document Term matrix and the transpose of a list containing the sentiment values of the
respective terms (i.e., 1 for positive and −1 for negative) according to the corresponding
lexicon. This operation multiplies the term frequency by the respective sentiment value
and sums these results for each tweet, producing a sentiment score for each message. The
tcrossprod simple triplet matrix function from slam package [26] computes this
matrix cross-product. The applied sentiment lists are created from the lexicon data frames.
In order to have a perfect match with the Document Term Matrix, these lists are ordered
alphabetically and items absent from all tweets are excluded. Tweets having positive sentiment score are considered positive and those with negative score are labeled as negative.
Table 10.1 shows some examples of tweets classification according to each model and the
respective sentiment metadata (Sent column).
Lexicon-based models present a significant number of neutral classifications indicating
some lack of coverage of these lexical resources. The save function saves the R objects
necessary for future operations.
234
Text Mining and Visualization: Case Studies Using Open-Source Tools
TABLE 10.1:
Examples of sentiment analysis according to each model (pos =
positive, neg = negative, neu = neutral).
Tweet
SVM NB GI FIN Sent
$OCZ trying to find a bottom here, like it over 2.20 pos pos neu neg pos
$OPXA today +12% after my call
pos pos pos neg pos
Stopped on rest of $GOOG +$1.25
pos pos neg neg neg
$ADBE Expecting a correction anytime now
pos neg neg neu neg
$ANAD on 2.14 Launch Pad
pos neg pos pos pos
$LIFE is a short if it can’t hold this 62.30 level
neg neg neu pos neg
There goes $LIFE short....
neg neg neu neu neg
$MELA When will it will go UP ?? mmmmm
pos neg neu neu pos
10.6
Evaluation
Model evaluation is important to assess the relevance of all models. This example compares the results of SA using all models performed on the test set. The evaluation metrics
are as follows:
• Classification accuracy (Acc): percentage of messages correctly classified
• For each label, positive (Pos) and negative (Neg), the following are calculated:
– Precision (PrePos , PreNeg ) – measures the proportion of true positives relative to
the number of true and false positives
– Recall (RecPos , RecNeg ) – measures the proportion of true positives relative to the
number of true positives and false negatives (also known as Sensitivity)
recision∗Recall
– F1 = 2 PPrecision+Recall
score (F1Pos , F1Neg ) – which considers both Precision and
Recall under a single measure
• Macro-averaged F-score (FAvg ) – averages all F-scores (i.e., F1Pos and F1Neg ).
The evaluation.R file codes the model evaluation.
### evaluation.R file ###
library(tm) # load R package tm: framework for text mining applications
library(rminer) # load R package rminer: Application of Data Mining Methods
load("preproc.RData") # load preprocessing objects
load("sent_analysis.RData") # load sentiment analysis objects
# evaluation metrics
svm_eval=mmetric(as.factor(twt$sentiment[(n_training+1):total]),svm_results[,1],
,→
metric=c("ACC","TPR","PRECISION","F1")) # SVM model evaluation
nb_eval=mmetric(as.factor(twt$sentiment[(n_training+1):total]),nb_results,
,→
metric=c("ACC","TPR","PRECISION","F1")) # NB model evaluation
GI_eval=mmetric(factor(twt$sentiment[(n_training+1):total],levels=c("0","negative",
,→
"positive")),as.factor(total_GI[(n_training+1):total]),metric=c("ACC","TPR",
,→
"PRECISION","F1")) # GI model evaluation
FIN_eval=mmetric(factor(twt$sentiment[(n_training+1):total],levels=c("0","negative",
,→
"positive")),as.factor(total_FIN[(n_training+1):total]),metric=c("ACC","TPR",
,→
"PRECISION","F1")) # FIN model evaluation
# create matrix with evaluation results
tab_eval<-rbind(c(svm_eval[c("ACC","PRECISION2","TPR2","F12","PRECISION1","TPR1","F11")],
,→
mean(svm_eval[c("F11","F12")])),
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
235
c(nb_eval[c("ACC","PRECISION2","TPR2","F12","PRECISION1","TPR1","F11")],
,→
mean(nb_eval[c("F11","F12")])),
c(GI_eval[c("ACC","PRECISION3","TPR3","F13","PRECISION2","TPR2","F12")],
,→
mean(GI_eval[c("F12","F13")])),
c(FIN_eval[c("ACC","PRECISION3","TPR3","F13","PRECISION2","TPR2","F12")],
,→
mean(FIN_eval[c("F12","F13")])))
colnames(tab_eval)<-c("Accuracy","Prec_Pos","Recall_Pos","F1_Pos","Prec_Neg",
,→
"Recall_Neg","F1_Neg","F_avg")
rownames(tab_eval)<-c("SVM","NB","GI","FIN")
tab_eval<-round(tab_eval)
print(tab_eval)
# exclude tweets without lexicon items in evaluation metrics calculation
# sum term frequency of GI items for each tweet of test set
num_GI<-rowSums(as.matrix(DTM_GI)[(n_training+1):total,])
# test set tweets without GI items (% values)
print(length(num_GI[num_GI==0])/length(num_GI)*100)
# GI model evaluation
GI_eval_exc=mmetric(factor(twt$sentiment[(n_training+1):total][num_GI>0],
,→
levels=c("0","negative","positive")),as.factor(total_GI[(n_training+1):
,→
total][num_GI>0]),metric=c("ACC","TPR","PRECISION","F1"))
# sum term frequency of FIN items for each tweet of test set
num_FIN<-rowSums(as.matrix(DTM_FIN)[(n_training+1):total,])
# test set tweets without FIN items (% values)
print(length(num_FIN[num_FIN==0])/length(num_FIN)*100)
FIN_eval_exc=mmetric(factor(twt$sentiment[(n_training+1):total][num_FIN>0], # FIN
,→
model evaluation
levels=c("0","negative","positive")),as.factor(total_FIN[(n_training+1):
,→
total][num_FIN>0]),metric=c("ACC","TPR","PRECISION","F1"))
# create matrix with evaluation results
tab_eval_exc<-rbind(c(svm_eval[c("ACC","PRECISION2","TPR2","F12","PRECISION1",
,→
"TPR1","F11")], mean(svm_eval[c("F11","F12")])),
c(nb_eval[c("ACC","PRECISION2","TPR2","F12","PRECISION1","TPR1","F11")],
,→
mean(nb_eval[c("F11","F12")])),
c(GI_eval_exc[c("ACC","PRECISION3","TPR3","F13","PRECISION2","TPR2","F12")],
,→
mean(GI_eval_exc[c("F12","F13")])),
c(FIN_eval_exc[c("ACC","PRECISION3","TPR3","F13","PRECISION2","TPR2","F12")],
,→
mean(FIN_eval_exc[c("F12","F13")])))
colnames(tab_eval_exc)<-c("Accuracy","Prec_Pos","Recall_Pos","F1_Pos","Prec_Neg",
,→
"Recall_Neg","F1_Neg","F_avg")
rownames(tab_eval_exc)<-c("SVM","NB","GI","FIN")
tab_eval_exc<-round(tab_eval_exc)
print(tab_eval_exc)
Preprocessed and SA data is required for this task. The calculation of most evaluation metrics is made by the mmetric function of rminer package. This function receives
vectors of target factors, predicted factors and metrics characters. In this situation, the
metrics characters are: "ACC" for global accuracy, "TPR" for recall (or true positive rate),
"PRECISION" for precision, and "F1" for each class F-score. The macro-averaged F-score
is computed by the mean function. Functions factor and as.factor encodes vectors of
data as factors. Since GI and FIN models assign an extra class (i.e., 0 for neutral), it is
important to include this class in the conversion of target sentiment labels to factors via
the levels argument. The evaluation metrics of all models are aggregated in a matrix using the rbind command. Table 10.2 presents the evaluation measures for all experimented
models.
Despite a low recall value for negative messages, the discriminative SVM model obtains
the best overall results. For example, it presents the highest values for accuracy and macroaveraged F-score. Moreover, ML models clearly outperform the lexicon-based models. The
latter models do not obtain the best value in any evaluation metric. However, many messages
do not contain any lexicon entry. Indeed, 58% tweets of test set do not hold any FIN item
236
Text Mining and Visualization: Case Studies Using Open-Source Tools
TABLE 10.2:
Sentiment analysis results (in %, best values in bold).
Model Acc PrePos RecPos F1Pos PreNeg RecNeg F1Neg FAvg
SVM
76
76
92
83
76
45
57
70
NB
67
82
62
71
52
75
61
66
GI
43
76
45
56
53
39
45
51
FIN
21
78
15
26
45
32
38
32
and 22% do not contain any GI item. This example also evaluates GI and FIN models after
excluding tweets without GI or FIN items, respectively. The rowSums function permits the
summation of each row of a matrix. Therefore, when applied to a Document Term Matrix,
it returns the total number of items that each document has. The resulting objects of this
operation on the GI and FIN test set matrices are used to select tweets containing lexicon
items. The utilization of a condition demanding values greater than zero for these objects
(e.g., num GI>0, num FI>0) on vectors of target or predicted labels permits to exclude
tweets without lexicon items. Table 10.3 shows these evaluation results.
TABLE 10.3:
Sentiment analysis results excluding messages without lexicon items
(in %, best values in bold).
Model Acc PrePos RecPos F1Pos PreNeg RecNeg F1Neg FAvg
GI
55
76
59
66
53
48
50
58
FIN
51
78
38
51
45
74
56
53
The evaluation results of ML models remain better than lexicon-based models. The main
reason for this behaviour is that there is a lack of opinion lexicons adapted to stock market
microblogging conversations. Domain-independent lexicons are not adjusted to the specific
terminology and semantics of stock market contents. Many lexical items holding significative
sentiment value in stock market conversations are not included in generic lexicons or have
opposite sentiment polarities, as shown in [27]. Moreover, the few public domain financial
lexicons (e.g., [22]) are not properly adapted to microblogging text. It should be noted that,
as previously explained, ML models have the drawback of requiring the use of labeled data,
which can be difficult or costly to get (e.g., manual classification).
10.7
Inclusion of Emoticons and Hashtags Features
This demonstration also tests the inclusion of emoticon and hashtags information in
the best performing classifier (i.e., SVM classifier). The following features are added to the
original Document Term matrix:
• An emoticon score corresponding to the difference between the number of positive
and negative emoticons existing in each tweet.
• The presence or absence of a set of hashtags. A matrix, containing all hashtags in
columns and all tweets in rows, specifies the presence of each hashtag in each message.
SA applying these features is coded in file emot hstg.R.
### emot_hstg.R file ###
library(tm) # load R package tm: framework for text mining applications
# load R package RTextTools: Automatic Text Classification via Supervised Learning
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
237
library(RTextTools)
library(rminer) # load R package rminer: Application of Data Mining Methods
load("preproc.RData") # load preprocessing objects
load("sent_analysis.RData") # load sentiment analysis objects
# function that calculates emoticon score (number of positive emoticons - number
,→
of negative emoticons)
scr_emot <- function (vec, tab_emot) {
if (length(vec)>0) {
vec<-gsub("[[:space:]]","",vec) # remove spaces
return(sum(tab_emot$sentiment[tab_emot$emoticon %in% vec]))
} else return (0)
}
# read a CSV file containing emoticons and respective sentiment polarity
t_emot<-read.csv("emoticons.csv",stringsAsFactors=FALSE)
emot_score<-unlist(lapply(l_emot, scr_emot, t_emot))
DTM<-cbind(DTM,emot_score) # add the emoticon score to Document Term matrix
colnames(DTM)[ncol(DTM)]<-"emot_score"
# function that verifies the presence or absence of each element of a list of
,→
hashtags
prs_hstg <- function (vec, n_hstg) {
return(as.numeric(n_hstg %in% vec))
}
hstg<-unique(unlist(l_hstg)) # hashtags vector
DTM_hstg <- unlist(lapply(l_hstg,prs_hstg,hstg))
DTM_hstg <- matrix(DTM_hstg,ncol=length(hstg),byrow=TRUE)
colnames(DTM_hstg)<-hstg
DTM<-cbind(DTM,DTM_hstg) # add hashtag matrix to Document Term matrix
# utilization of SVM method
container <- create_container(DTM,twt$sentiment,trainSize=1:n_training,
,→
testSize=(n_training+1):total, virgin=FALSE) # creation of a container
svm_model <- train_model(container,"SVM")
svm_results <- classify_model(container,svm_model)
# evaluation metrics
svm_eval=mmetric(as.factor(twt$sentiment[(n_training+1):total]),svm_results[,1],
,→
metric=c("ACC","TPR","PRECISION","F1"))
svm_eval<-c(svm_eval[c("ACC","PRECISION2","TPR2","F12","PRECISION1","TPR1","F11")],
,→
mean(svm_eval[c("F11","F12")]))
names(svm_eval)<-c("Accuracy","Prec_Pos","Recall_Pos","F1_Pos","Prec_Neg",
,→
"Recall_Neg","F1_Neg","F_avg")
print(svm_eval)
The lists of emoticons and hashtags (i.e., l emot and l hstg) extracted in data preprocessing are necessary in this step. This use case provides a file (emoticons.csv) containing a comprehensive list of emoticons and the respective sentiment (1 for positive, -1
for negative) in order to facilitate the identification of positive and negative emoticons. The
function scr emot is created to compute the emoticon score of a tweet. It receives the corresponding l emot entry and the emoticons data frame and returns the difference between
the number of positive and negative emoticons. The function lapply applies scr emot
function to each l emot element, producing a list of emoticon scores. Unlist converts this
list into a vector of emoticon scores and cbind function adds this vector as a column to the
Document Term matrix. The function prs hstg verifies the presence or absence of each
element of a list of hashtags in a tweet. The inputs are two vectors containing the hashtags
of a tweet and the hashtags to check the presence. The result is a vector of the same length
of the latter input vector indicating the presence (i.e., 1) or absence (i.e., 0) of each hashtag
in the former input vector. The combined utilization of unlist and unique functions on
l hstg list produces a vector of all hashtags appearing in the data set without duplicated
238
Text Mining and Visualization: Case Studies Using Open-Source Tools
elements. These are the hashtags that are checked in this example. The lapply function
applies prs hstg function to each l hstg element and passes the vector of hashtags to
check as an argument. The matrix function transforms the lapply result into a matrix
having all tweets in rows and all hashtags in columns. This matrix is attached to the Document Term matrix using the function cbind. The resulting matrix feeds a SVM classifier.
Table 10.4 presents SA results for the SVM classifier with emotions and hashtags features
(SVMEmtHsht ) and without these features (SVMOriginal ). As shown in Table 10.4, the addition
of emoticons and hashtags features improves the majority of evaluation results. Therefore,
emoticons and hashtags demonstrate some informative value.
TABLE 10.4:
Sentiment analysis results with emoticons and hashtags features (in %,
best values in bold).
Model
Acc PrePos RecPos F1Pos PreNeg RecNeg F1Neg FAvg
SVMOriginal
76
76
92
83
76
45
57
70
SVMEmtHsht 77
76
93
84
78
47
58
71
10.8
Conclusion
This chapter presented a use case about sentiment analysis (SA) in Twitter data related
to stock market using the R tool. Mining microblogging data to forecast stock market behavior is a recent and interesting research topic. Some studies showed that sentiment indicators
created from microblogging data appears to have predictive value for stock market variables
(e.g., [6, 9, 10]). Moreover, the production of these indicators is faster and less expensive
than traditional sources (e.g., large-scale surveys). R is a free language and environment
that allows the diverse steps within this topic to be performed. There are several packages
(e.g., twitterR, RTextTools, rminer) that provide the required functionalities for
this tool.
This use case provided and described R code for diverse tasks forming this process
and supplied a Twitter data set with stock market messages for the execution of those
operations. The addressed tasks were tweets collection, preprocessing tasks, utilization of
opinion lexicons, application of supervised ML methods, and respective evaluation.
Overall, the supervised ML models outperformed the unsupervised lexicon-based models. These latter models have insufficient coverage and inferior accuracy than former models.
This behavior is due to the current lack of lexicons properly adapted to stock market microblogging conversations [27]. However, it should be noted that ML models require labeled
data for the learning procedure (which can be costly to get), whereas lexicon-based models do not. Considering ML models, the discriminative representative (i.e., Support Vector
Machine) obtained higher results that the generative representative (i.e., Naı̈ve Bayes). In
addition, emoticons and hashtags features have shown to add informative content to SA
models, since the inclusion of these features allowed to improve SA results.
Sentiment Analysis of Stock Market Behavior from Twitter Using the R Tool
239
Bibliography
[1] B. Pang and L. Lee. Opinion Mining and Sentiment Analysis. Foundations and Trends
in Information Retrieval, 2(2):1–135, 2008.
[2] M. Hu and B. Liu. Mining and summarizing customer reviews. Proceedings of the 2004
ACM SIGKDD international conference on Knowledge discovery and data mining KDD
04, 04(2):168, 2004.
[3] Y. Choi and C. Cardie. Learning with Compositional Semantics as Structural Inference
for Subsentential Sentiment Analysis. Proceedings of the Conference on Empirical
Methods in Natural Language Processing EMNLP 08, (October):793–801, 2008.
[4] Bing Liu and Lei Zhang. A Survey of Opinion Mining and Sentiment Analysis. In
Charu C. Aggarwal and ChengXiang Zhai, editors, Mining Text Data, pages 415–464.
Springer, 2012.
[5] W. Antweiler and Murray F. Is All That Talk Just Noise? The Information Content
of Interest Stock Message Boards. Journal of Finance, 59(3):1259, 2004.
[6] J. Bollen, H. Mao, and X. Zeng. Twitter mood predicts the stock market. Journal of
Computational Science, 2(1):1–8, 2011.
[7] P. C Tetlock. Giving Content to Investor Sentiment : The Role of Media in the Stock
Market. Journal of Finance, 62(3):1139–1168, 2007.
[8] J. R. Nofsinger. Social Mood and Financial Economics Social Mood and Financial
Economics. Journal of Behavioral Finance, 6(3):144–160, 2005.
[9] C. Oh and O. R Liu Sheng. Investigating Predictive Power of Stock Micro Blog Sentiment in Forecasting Future Stock Price Directional Movement. In ICIS 2011 Proceedings, Shanghai, China, 2011.
[10] T. O Sprenger and I. M Welpe. Tweets and Trades: The Information Content of Stock
Microblogs. Social Science Research Network Working Paper Series, pages 1–89, 2010.
[11] N. Oliveira, P. Cortez, and N. Areal. Some experiments on modeling stock market behavior using investor sentiment analysis and posting volume from Twitter. In Proceedings of the 3rd International Conference on Web Intelligence, Mining and Semantics
(WIMS ’13), Madrid, Spain, 2013. ACM.
[12] N. Oliveira, P. Cortez, and N. Areal. On the Predictability of Stock Market Behavior using StockTwits Sentiment and Posting Volume. In 16th Portuguese Conference
on Artificial Intelligence (EPIA 2013), Angra do Heroı́smo, Açores, Portugal, 2013.
Springer Berlin Heidelberg.
[13] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2013.
[14] I. Feinerer, K. Hornik, and D. Meyer. Text Mining Infrastructure in R. Journal of
Statistical Software, 25(5):1–54, 2008.
[15] T. Jurka, L. Collingwood, A. Boydstun, E. Grossman, and W. van Atteveldt. RTextTools: Automatic Text Classification via Supervised Learning, 2012. R package version
1.3.9.
240
Text Mining and Visualization: Case Studies Using Open-Source Tools
[16] M. Bouchet-Valat. SnowballC: Snowball stemmers based on the C libstemmer UTF-8
library, 2014. R package version 0.5.1.
[17] I. Feinerer and K. Hornik. wordnet: WordNet Interface, 2014. R package version 0.1-10.
[18] A.F. Zuur, E.N. Ieno, and E. Meesters. A Beginner’s Guide to R. Springer, 2009.
[19] J. Gentry. twitteR: R based Twitter client, 2013. R package version 1.1.7.
[20] A Jordan. On discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes. Advances in Neural Information Processing Systems, 14:841,
2002.
[21] P. J. Stone, D. C. Dunphy, M. S. Smith, and D. M. Ogilvie. The General Inquirer: A
Computer Approach to Content Analysis, volume 08. MIT Press, 1966.
[22] T. Loughran and B. McDonald. When is a liability not a liability? Textual analysis,
dictionaries, and 10-Ks. Journal of Finance, 66(1):35–65, 2011.
[23] Mirai Solutions GmbH. XLConnect: Excel Connector for R, 2014. R package version
0.2-9.
[24] Paulo Cortez. Data mining with neural networks and support vector machines using
the R/rminer tool. In Petra Perner, editor, Advances in Data Mining - Applications
and Theoretical Aspects 10th Industrial Conference on Data Mining (ICDM 2010),
volume 6171 LNAI of 10th Industrial Conference on Advances in Data Mining, ICDM
2010, pages 572–583, Berlin, Germany, 2010. Springer.
[25] D. Meyer, E. Dimitriadou, K. Hornik, A. Weingessel, F. Leisch, C-C Chang, and C-C
Lin. e1071: Misc Functions of the Department of Statistics (e1071), TU Wien, 2014.
R package version 1.6-4.
[26] K. Hornik, D. Meyer, and C. Buchta. SLAM: Sparse Lightweight Arrays and Matrices,
2014. R package version 0.1-32.
[27] N. Oliveira, P. Cortez, and N. Areal. Automatic Creation of Stock Market Lexicons for
Sentiment Analysis Using StockTwits Data. In Proceedings of the 18th International
Database Engineering & Applications Symposium, IDEAS ’14, pages 115–123, New
York, NY, USA, 2014. ACM.
Chapter 11
Topic Modeling
Patrick Buckley
Institute of Technology, Blanchardstown, Ireland
11.1
11.2
11.3
11.4
11.5
11.6
11.7
11.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.1.1 Latent Dirichlet Allocation (LDA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Aims of This Chapter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2.1 The Data Set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.2.2 R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Loading the Corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3.1 Import from Local Directory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.3.2 Database Connection (RODBC) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Preprocessing the Corpus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Document Term Matrix (DTM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Creating LDA Topic Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6.1 Topicmodels Package . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6.2 Determining the Optimum Number of Topics . . . . . . . . . . . . . . . . . . . . . . . .
11.6.3 LDA Models with Gibbs Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6.4 Applying New Data to the Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6.5 Variational Expectations Maximization (VEM) Inference Technique
11.6.6 Comparison of Inference Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
11.6.7 Removing Names and Creating an LDA Gibbs Model . . . . . . . . . . . . . . . .
Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
241
242
243
243
243
244
244
245
246
247
249
250
251
255
259
259
261
263
263
264
Introduction
Topic models can be described as the idea that documents naturally contain a mix
of topics where a topic can be described as a distribution over the terms in a corpus of
documents [1]. They are desirable where there is a massive amount of textual data stored
in large collections and where short descriptions of those documents could be created and
used in a variety of ways; to identify similarities across the corpus, to create visualizations
of where documents fit relative to others, and also in searching and indexing methods [2].
The main goal is to find short descriptions of the topics in a corpus while still maintaining
the statistical relationships in the corpus [3]. In this way, popular tasks like classification,
similarity, and summation can be applied.
Topic models are applicable to a variety differing sectors like scientific, political, fraud,
software, and Web usage to name a few, all of which benefit from the topic modeling technique of identifying hidden structures in otherwise unstructured text collections [4]. While
241
242
Text Mining and Visualization: Case Studies Using Open-Source Tools
traditionally accepted to be a text based methodology, topic models have been extended
into other applications including images, music and video analysis.
There are a variety of topic modeling techniques but arguably the most popular technique is known as Latent Dirichlet Allocation (LDA), which was developed by Blei et al.
[3].
11.1.1
Latent Dirichlet Allocation (LDA)
The LDA topic model can be described as a probabilistic model that makes the assumption that given a corpus of documents, each document is a distribution over the topics and
in turn every topic is a distribution over the words in the corpus [3]. It uses a generative
model to mimic how documents are created. Simply put, the idea behind the LDA model
is that documents can be represented as random mixtures of hidden or latent topics in a
corpus. It allows for the reduction in the complexity of the corpus of text, the capturing of
key information and their latent underlying statistical relationships.
If you take an example of a collection of scientific documents, there will be a variety
of fields of study within the collection so therefore there will be a variety of different topics. Documents sampled from the collection might contain information regarding human
anatomy and genetics, another document could contain information on cloning and technology; and another might contain information on genetics and cloning [4]. Figure 11.1
shows a news article extract taken from The New York Times (12/04/2014) illustrating
this idea at a high level. Since documents when created are formed from a series of topics,
a number of potential topics are highlighted in the figure to express this idea. For example,
words highlighted in red could collectively be named a topic about “energy”, similarly words
highlighted in blue could form a topic on “geographic locations” [5].
FIGURE 11.1 (See color insert.): News extract topics.
Topic Modeling
243
LDA uses probabilistic sampling methods to create a generative model for documents
by describing how the words in documents could be created with the use of hidden or latent
variables. The main aim of generative models is to find the group of variables that best
describe/explain the words observed in documents [1]. In a generative model, the concept
of mutual exclusivity doesn’t apply, meaning words can be part of one or more topics. In
this way, topic models are enabled to capture words that exhibit polysemy.
As the generative model is hypothetical, given that all that is actually observed are
the terms in the set of documents, LDA uses a variety of inference techniques to estimate
the posterior model distribution. One of these methods is Gibbs sampling [1]. It works by
assigning a value to each variable in the model and then taking each word token in a corpus
and for each it works on estimating the probability of assigning the current word to each
topic. It is from the posterior distribution that we can derive the topics for a corpus.
11.2
Aims of This Chapter
The aim of this chapter is to guide the reader through the process of creating a corpus,
right through to the creation of a Latent Dirichlet Allocation (LDA) topic model using the
statistical package R. The following are some of the main areas that will be discussed and
explained:
1. Creating a corpus in R
2. Pre-processing the corpus
3. Creating a Document Term Matrix
4. Finding the optimum number of topics for the data
5. Creating the topic model
6. Interpreting the models
7. Applying to unseen documents
11.2.1
The Data Set
The data used throughout this chapter is from the UCI machine learning repository and
is based upon 20,000 messages taken from 20 newsgroups (http://kdd.ics.uci.edu/
databases/20newsgroups/20newsgroups.html). Given the size of the full data set,
there is a smaller version of this available which will be used, so depending on the computational power available to you it may make more sense to use that. The mini newsgroup data
set contains 2,000 messages from the 20 newsgroups and can be downloaded as a tar.gz file
(mini newsgroups.tar.gz).
11.2.2
R
All of the processing and topic modeling will be performed using the R statistical package. The version of R used for this chapter is 3.1.1 (64 bit). There are a number of packages
available in R for topic modeling such as LDA, topicmodels, and others. For the purposes
of this chapter the topicmodels package is explained and used throughout.
244
11.3
Text Mining and Visualization: Case Studies Using Open-Source Tools
Loading the Corpus
There are multiple ways to load a corpus depending on where your data source is stored.
Two of the more common methods are importing a corpus from a local file repository and
importing files from a database.
11.3.1
Import from Local Directory
The first step is to install and load the text mining (tm) package library into R, this contains the input methods used to create the corpus from the folder of documents. The package
can be installed by either using the “Load Packages” option in the “Packages” menu in the R
console or by entering the following directly into the R terminal install.packages(“tm”).
R should then ask you to specify a CRAN mirror from which the download can begin.
tm: This package in R allows for a wide range of tasks for text mining including data
importation, creation of a corpus, preprocessing of the text data and creation of matrices for
the terms and documents that can then be used by other packages to perform topic modeling
amongst multiple other things. (http://cran.r-project.org/web/packages/tm/
index.html)
After loading the package, R by default sets a working directory location where it
looks for input files and equally where it will store any outputs you may save. If you
are unsure where the working directory is, simply entering getwd() into the console
will give you the file path. If you would prefer to change the default location for the
working directory to another location, you can use the set working directly command
setwd(“c:/your preferred location”). Note that the forward slash character should be
used in the file path and not backslash. After this has been changed, you can use getwd()
to confirm the directory changes have taken place.
Now that you have set the preferred working directory (or are using the default), unzip
the mini newsgroup data and add the newsgroup folder into the directory. The following R
code can be used to import the files and convert them into a corpus.
We can use getSources() to list the sources and use getReaders() to view the document formats that are available to use in the tm package. There are different approaches
based on the input files you are working with, but since this exercise is dealing with text
based files, the following commands point R to a direct source (DirSource). By entering
the file path to the mini newsgroup folder, R can begin to import the files. Since our data
set contains multiple folders with text files, setting the parameter recursive=TRUE tells
R to look at all the folders within the directory. The Corpus command is used to load the
documents and create a corpus given the directory source.
#LOADING A CORPUS FROM LOCAL FOLDERS:
#Load the text mining package(s)
library("tm")
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
#Here the file path is input using a Direct Source command
#which is in turn transformed into a corpus. Also telling R
#to look in multiple folders (recursive)
corpus <- corpus(DirSource(
Topic Modeling
245
directory = "C:/Users/Documents/R/DATASET/mini_newsgroups/",
encoding = "UTF-8",
recursive = TRUE,
mode = "text"
))
After running these commands in the R console, by entering corpus, this shows the
results of the newly created corpus as follows:
corpus
> A corpus with 376 text documents
Similar to importing files from a local directory, the corpus can be imported through a
database connection.
11.3.2
Database Connection (RODBC)
There are also multiple other database connection options including RMySQL, ROracle, and RJDBC, each of which are available from http://cran.r-project.org. The
RODBC package in R was used to perform this connection. In this example, RODBC
works by creating an open database connection (ODBC) or channel from the client packages (RODBC). In order to establish the connection, the setup of an ODBC (open database
connectivity) was required. This involved downloading the required driver for ODBC and
mySQL and creating the connection to the database required.
RODBC: is one package that allows a connection to be made to an ODBC database
(http://cran.r-project.org/web/packages/RODBC/index.html)
After the initial setup, the package was installed in R and the following code can be
used to create a connection to where the data in the mySQL database is stored. Following
on from that the sqlFetch command was used to pull the required table from the database
into R.
#LOADING A CORPUS FROM DATABASE:
#Load the text mining package(s)
library("tm")
library("RODBC")
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
#Create connection to your database and then retrieve data
database <- odbcConnect("your_database_connection", "root",
"password")
datatable <- sqlFetch(database, "database_name")
In addition to importing the data from the database, some further processing is required to transform the data from its database table format into a structure that can
be transformed into a corpus. This first step is to create an R object (e.g., myReader)
instructing how R should read in the data. Assuming that the data has been stored in
the database with the main body of text in one column and the name of the file in another column (amongst other columns of data), the variable m is set so that it reads in
and sets the “ID” as the “file name” and the contents as the “text” (note: the names
should reflect the column names in your database tables). m then becomes an input into
the next command, which specifies that R reads the database table as tabular and to
246
Text Mining and Visualization: Case Studies Using Open-Source Tools
map to the requirements stated in m. More information on this can be found at http:
//cran.r-project.org/web/packages/tm/vignettes/extensions.pdf.
#Converting database tabular structure for text documents
m <- list(ID = "file_name", Content = "text")
myReader <- readTabular(mapping = m)
The next stage then is to create the corpus. Whereas in the previous example a
DirSource command was used, this time the DataframeSource command is used as
the data in the database is stored as a data frame. This in combination with the previously
created myReader ensures that the corpus is created.
#Creating the corpus from a data frame
corpus <- corpus(DataframeSource(datatable),
readerControl = list(reader = myReader, language = "en"))
Now that we have examine some methods for importing data and creating a corpus, the
next stage we will look at is preprocessing techniques.
11.4
Preprocessing the Corpus
After importing the data and converting the documents into a corpus, a number of various methods for preprocessing can be applied. These include transformations to lower-case,
removal of numbers and punctuation removal of stopwords and white space The package
SnowballC is one example of a stemmer that can be used and is the only package that is
required in addition to the tm package that is already installed and loaded.
SnowballC: Snowball provides a stemmer for the words in the corpus as part of preprocessing (http://cran.r-project.org/web/packages/SnowballC/index.html)
In the R code shown below, each of the commands use the transformation function
tm map to transform the corpus text to a more useful format.
The first line of code sets up the stopwords that you would like to remove from the
corpus. In addition to the standard English stopwords dictionary that is supplied with
the tm package, you can further remove custom words as shown by the “my”, “custom”,
“words”. The text mining package also contains other stopwords dictionaries in a variety of
languages while also providing a second English words dictionary called SMART dictionary
which is much more extensive.
Following on from setting up your stopwords, the corpus can then be transformed removing; numbers, stopwords, white space, and punctuation. Then the stemmer is applied
using the SnowballC package.
#PRE-PROCESSING THE CORPUS:
#Load the text mining package(s)
library("SnowballC")
#Setting the stopwords dictionaries and custom words
myStopwords <- c(stopwords("english"), "my", "custom", "words")
#Pre-processing and transforming the corpus
myStopwords <- c(stopwords("english"), stopwords("SMART"))
Topic Modeling
your_corpus
your_corpus
your_corpus
your_corpus
your_corpus
your_corpus
<<<<<<-
247
tm_map(corpus, content_transformer(tolower))
tm_map(your_corpus, removeWords, myStopwords)
tm_map(your_corpus, removeNumbers)
tm_map(your_corpus, removePunctuation)
tm_map(your_corpus, stripWhitespace)
tm_map(your_corpus, stemDocument)
After the corpus has been successfully preprocessed, the next stage is to convert the
corpus into a document term matrix, which will allow for further inspection.
11.5
Document Term Matrix (DTM)
This is simply a matrix where the documents are the rows and the words the columns,
the matrix cells then contain a frequency count of the words (weightTf), which is the default weighting function in the tm package. DTMs can also contain different weighting
functions such as weightTfIdf (term frequency - inverse document frequency) weightBin
(Binary weight a term-document matrix), and weightSMART (weights according to a combination of weights specified in SMART notation). This is a necessary input into the topic
modeling package topicmodels so that modeling can be performed. The command DocumentTermMatrix is used to create the matrix and is part of the tm package. More
information on this can be found by entering the following TermDocumentMatrix. The
control list as shown gives you an option to remove terms of less than and greater than a
specified number. In this case, the document terms matrix keeps terms that are 3 characters
minimum and can be infinite in length.
#CREATING A DOCUMENT TERM MATRIX:
#Create a document term matrix (containing between 3 &
#Infinite characters)
myDtm <- DocumentTermMatrix(
your_corpus, control=list(
wordLengths=c(3,Inf)
)
)
At this stage, some examination of the DTM can be conducted which gives you a chance
to access the quality of the data and to get a feel for the corpus. To simply examine the
output of the topic model, entering myDtm into the terminal will display information
about the matrix created. Examining the output shows useful statistics like the sparsity of
the DTM (meaning the level of emptiness or zeros), this shows a very sparse matrix where
the majority of entries are zero. In the case of the mini newspaper corpus, we have a corpus
with 2,000 documents containing 34,943 terms, with a huge level of sparsity.
myDtm
<<DocumentTermMatrix
Non-/sparse entries:
Sparsity
:
Maximal term length:
Weighting
:
(documents: 2000, terms: 34943)>>
194343/69691657
100%
232
term frequency (tf)
Another method to inspect the DTM is to use the Find Frequent Terms function findFreqTerms(myDtm, 100), this looks to find words in the matrix that occur 100 times
248
Text Mining and Visualization: Case Studies Using Open-Source Tools
or more. This displays a list of the terms that occur over 100 times. After examining these
frequent terms, there are some terms that are very long in length that don’t actually make
much sense and haven’t been removed as part of preprocessing.
findFreqTerms(myDtm, 100)
[55] "cantaloupesrvcscmu"
[83] "composmswindowsmisc"
[85] "compsysmachardwar"
"car"
"compsysibmpchardwar"
"comput"
In order to remove theses words, we can create the document term matrix again
only this time we can set the maximum length for words to keep to be 15 or under
(wordLengths=c(3,15)).
<<DocumentTermMatrix
Non-/sparse entries:
Sparsity
:
Maximal term length:
Weighting
:
(documents: 2000, terms: 28039)>>
183484/55894516
100%
15
term frequency (tf)
This has reduced the DTM significantly indicating that there were many terms of greater
than 15 characters in length. However, the quantity of terms in the corpus is still relatively
large and the sparsity level is still 100%, which is understandable given that the matrix
would mostly be filled by zeros. Words that occur frequently in documents and across
other documents are not good differentiators in topic modeling, words that occur in many
documents and many times in a single document are given a lower weighting using TF-IDF
(term-frequency - inverse document frequency). This then allows terms with a low frequency
to be omitted as they are insignificant to the model and likewise for terms occurring in many
documents.
In order to calculate the term frequency-inverse document frequency, the package
“slam” is required.
slam: provides data structures and algorithms for sparse arrays and matrices (http:
//cran.r-project.org/web/packages/slam/index.html)
The following code is used to calculate the TF-IDF from the DTM that was created. It
will also display a summary of the results as shown:
#CALCULATING THE TERM FREQUENCY - INVERSE DOCUMENT FREQUENCY:
library("slam")
#Calculate TF-IDF
term_tfidf <tapply(myDtm$v/row_sums(myDtm)[myDtm$i], myDtm$j, mean)
* log2(nDocs(myDtm)/col_sums(myDtm > 0))
#Display results of TF-IDF
> summary(term_tfidf)
Min. 1st Qu. Median
Mean 3rd Qu.
Max.
0.00000 0.03821 0.07409 0.09763 0.12750 1.64100
Examining the results above, we can determine that we should only include terms with
a value of greater than 0.074 as that is slightly below the median number. The median is
used as it is not affected by large TF-IDF values in the data whereas the mean would be
more influenced. Applying this in the following code and also removing documents where
no features a present results in a much smaller DTM with 14, 081 terms.
Topic Modeling
249
#Retaining features that have a high TF-IDF value
myDtm <- myDtm[,term_tfidf >= 0.074]
myDtm <- myDtm[row_sums(myDtm) > 0,]
summary(col_sums(myDtm))
#Reports the dimensions of the DTM
dim(myDtm)
[1] 2000 14081
#Saving the DTM to local folder
save(myDtm, file = "my_Dtm.Rdata")
Another method that can be applied is the removal of sparse terms command, removeSparseTerms(your corpus, 0.3), this will remove the sparse terms to beyond a specified
sparsity level set by the user. This method can sometimes result in most of the DTM being
removed leaving only a small percentage of the words so care must be taken not to remove too much. More information on this can be found by typing ?removeSparseTerms
directly into the R terminal.
Based on the above TF-IDF example, a word-cloud can be used to visualize the DTM
as shown in Figure 11.2. If when examining the output from the word-cloud shows anomalies or words that you know should be removed (e.g., a person’s name or something that
a stop-words dictionary has not picked up), adding these to the list of custom stop-words
and repeating the preprocessing is advised. The package “wordcloud” is used to create the
output.
wordcloud: used in the creation of word-clouds from text corpus. (http://cran.
r-project.org/web/packages/wordcloud/index.html)
#CREATING A WORD CLOUD:
library("wordcloud")
#convert document term matrix to a term document matrix
myTdm <- t(myDtm)
#define tdm as matrix
m = as.matrix(myTdm)
# get word counts in decreasing order
word_freqs = sort(rowSums(m), decreasing=TRUE)
# create a data frame with words and their frequencies
dm = data.frame(word=names(word_freqs), freq=word_freqs)
# plot word-cloud with maximum of 200 words
wordcloud(dm$word, dm$freq, max.words=200, random.order=FALSE,
rot.per=.2, colors=brewer.pal(9, "Dark2"))
11.6
Creating LDA Topic Models
The topicmodels package is used throughout this section in performing topic modeling of the data set. The first stage is to install the package which can be completed by
entering install.packages(“topicmodels”) into R. Once it has been installed, use the
250
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 11.2: Word-cloud of the DTM.
library(“topicmodels”) command to load the package.
topicmodels: was developed to provide an interface to R for the C code developed by
[3].The package contains capability to create topics models based on Latent Dirichlet Allocations (LDA) but also for the Correlated Topic Model (CTM). It uses Variational EM for
estimating the topics but also provides access to Gibbs sampling methods created by XuanHieu Phan and co-authors in C++. (http://cran.r-project.org/web/packages/
topicmodels/index.html)
A Latent Dirichlet Allocation (LDA) will be built using Gibbs sampling, and the
following represents the default values for this Gibbs sampling. These are a list of
all the possible parameters or controls in fitting an LDA model using Gibbs sampling [6] from the topicmodels package. More information on each can be found in
the topicmodels CRAN document [http://cran.r-project.org/web/packages/
topicmodels/topicmodels.pdf]:
R> control_LDA_Gibbs <- list(
alpha = 50/k, estimate.beta = TRUE, verbose = 0, prefix =
tempfile(), save = 0, keep = 0, seed = as.integer(Sys.time()),
nstart = 1, best = TRUE, delta = 0.1, iter = 2000, burnin = 0,
thin = 2000)
In the next section, some of these parameters for creating an LDA topic model using
Gibbs sampling will be explained.
11.6.1
Topicmodels Package
To give an overview of the parameters in the package for topic modeling , the following
definitions are from the ‘topicmodels’ paper by [6] and from the topicmodels CRAN package
pdf [http://cran.r-project.org/web/packages/topicmodels/index.html]:
• estimate.alpha, alpha, estimate.beta - The default value for α is estimated (esti-
Topic Modeling
251
mate.alpha = TRUE) and as Griffiths and Steyvers suggested [1], this value should be
set at 50/k, where k is the number of topics. Where α is not set as estimated, it will
stay at the initial value. β, the term distribution over topics, is set at estimate.beta
= TRUE. If however, the term distribution have been fitted by a previous model,
setting estimate.beta = FALSE allows hold-out data or testing data to be applied to
the fitted model.
• verbose: verbose is set to 0 as default which means during the algorithm, nothing
will be printed, if this value changes to a positive integer all iterations during the
algorithm will be printed.
• save & prefix: save is default to zero meaning no intermediate results are saved to
file with prefix
• keep: when keep is set to a positive integer, the values for the log-likelihood will be
store for every keep iteration.
• seed: allows for reproducibility
• nstart: shows how many repeated runs there are with random initializations
• delta: Parameter of the prior distributions for terms distribution over topics.
• iter, burnin, thin: determine how many draws are made over the corpus during
Gibbs sampling. The number of burnin iterations that are set indicates the number
that will be discarded before each thin iteration after that point is stored for iter
iterations.
• best: default value is set to TRUE which means only the best posterior likelihood is
returned, if value is set to FALSE all draws are returned.
Note that not every parameter requires input from the user when running an LDA
model. The main parameters that are used will be discussed in the next section.
11.6.2
Determining the Optimum Number of Topics
One of the main issues with finding topic models in a corpus is determining the optimum
number of topics. This is something that is set by the modeler. Perplexity is often used as a
method of evaluation for the optimum number of topics and is used on the held-out or test
portion of a data set. Perplexity analysis characterizes the quality of the generative model.
This method acts as a rough guide to help narrow down the area of focus and does not
mean that you will get the precise number of actual topics, this is where the knowledge of
the domain expert is required.
A cross-validation is applied to the data in order to gain an insight into the performance
of the model. The cross-validation is based on the methodology used in [6]. It makes 10
test and training sets, each set is based on different partitions of the whole data. Each set
contains 90% for training and 10% for testing, the result is 10 performances which are then
averaged.
The following controls were put in place after installing the ‘topicmodels’ package, loading the library and setting up the working directory. Firstly the number of topics was set
to run from 5 to 35 using the stages set out in the following code, for example 5 times the
sequence from 1 to 3, giving 5, 10, 15 and then 25 and 35 [Note: Topics 2 to 5 are considered
later in this section].
252
Text Mining and Visualization: Case Studies Using Open-Source Tools
#SETTING LDA PARAMETERS FOR TOPIC MODELLING:
#Load the text mining package(s)
library("topicmodels")
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
#Setting up parameters for LDA topic modeling
topics <- 5 * c(1:3, 5, 7)
SEED <- 20080809
BURNIN = 1000
THIN = 100
ITER = 1000
BEST = FALSE
#Loading your document term matrix
load("my_Dtm.Rdata")
data <- myDtm
Following on from this, for each iteration of the Gibbs sampling, the following controls
were used. A “burnin” rate of 1,000 was set meaning the model would run 1,000 times before
starting to record any results. Then it was set to run for another 1,000 iterations and to
record the results for every 100th run. Setting best to FALSE allowed for each log-likelihood
value to be returned and not just the lowest value. In this way, plots can be developed to
examine how the perplexity of the model looks for the test data. The parameter α was left
at its default value, which is 50/k with k as the number of topics to be generated. Once
this was set, the document term matrix created and saved earlier was then loaded into the
model and set as “data”.
All of the controls mentioned are used in both the training and the testing models.
However, the exception is the addition of the ‘estimate.beta’ control in the testing models.
This is set to FALSE (by default it is set to TRUE) so that the term distributions that were
created in the training models can now be used to apply the test data to and a perplexity
score can be calculated.
The following code (source [6]) represents the cross-validation portion where the split
between training and testing is created and the models are run for both. Each of the
models are created using LDA and Gibbs sampling and the results are then saved to an
output folder. Creating a folder in your working directory (prior to running the code) called
“results” is necessary for the outputs to be saved.
#X-VALIDATION TO DETERMINE OPTIMUM NUMBER OF TOPICS (1):
# X-Validation - splitting data into 10 Test data sets
# with remainder used in Training
D <- nrow(data)
folding <- sample(rep(seq_len(10), ceiling(D))[seq_len(D)])
for (k in topics)
{
for (chain in seq_len(10))
{
FILE <- paste("Gibbs_", k, "_", chain, ".rda", sep = "")
training <- LDA(data[folding != chain,], k = k,
control = list(seed = SEED,
burnin = BURNIN, thin = THIN, iter = ITER, best = BEST),
method = "Gibbs")
best_training <- [email protected][[which.max(logLik(training))]]
Topic Modeling
253
testing <- LDA(data[folding == chain,], model = best_training,
control = list(estimate.beta = FALSE, seed = SEED,
burnin = BURNIN,
thin = THIN, iter = ITER, best = BEST))
save(training, testing, file = file.path("results", FILE))
}
}
Once the previous process has run, you should end up with a number of files in a
folder called “results” in your working directory. This contains all the results from each
run through the X-validation. It is from this that comparisons can be made as to which
model was generated with the optimum number of topics. Note that the process can take
a considerable amount of time (45minutes in my case) as there will be 50 LDA models
produced using Gibbs sampling.
The following code (source [6]) then compiles the results and outputs a file displaying
the log likelihood for the test data. From this you can see which number produced the lowest
perplexity.
#X-VALIDATION TO DETERMINE OPTIMUM NUMBER OF TOPICS (2):
topics <- 5 * c(1:3, 5, 7)
SEED <- 20080809
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
load("my_Dtm.Rdata")
data <- myDtm
D <- nrow(data)
folding <sample(rep(seq_len(10), ceiling(D))[seq_len(D)])
News_test <News_alpha <- list()
for (method in c("Gibbs"))
{
News_alpha[[method]] <- News_test[[method]] <- matrix(NA,
nrow = length(topics), ncol = 10,
dimnames = list(topics, seq_len(10)))
for (fold in seq_len(10))
{
for (i in seq_along(topics))
{
T <- topics[i]
FILE <- paste(method, "_", T, "_", fold, ".rda", sep = "")
load(file.path("results", FILE))
News_alpha[[method]][paste(T),fold] <if (is(training, "Gibbs_list")) [email protected][[1]]@alpha
else [email protected]
News_test[[method]][paste(T),fold] <- perplexity(testing,
data[folding == fold,], use_theta = FALSE)
}
}
}
save(News_alpha, News_test, file = "News.rda")
The output from this is a file named “News.rda” containing the alpha values used during
254
Text Mining and Visualization: Case Studies Using Open-Source Tools
each model and also the perplexity values calculated during each model build for the given
number of topics. The model that minimizes the perplexity value should be chosen as shown:
#Find the lowest average log-likelihood
p <- News_test
q <- as.data.frame(p)
rowMeans(q)
5
10
15
25
35
4616.543 4679.822 4782.607 5045.052 5334.478
For the mini-newsgroup corpus, the optimum number of topics seems to be minimized at
5 topics as this produced the lowest perplexity from the X-validation. This can be seen in
Figure 11.3, it also could be possible that the number of topics is below 5.
FIGURE 11.3 (See color insert.): Cross-validation — optimum number of topics.
Running the analysis again only this time for topics 2 to 5 shows that there is very
little difference found by the cross validation possibly indicating that the optimum number
of topics is somewhere in this region, this can be seen in Figure 11.4. In order to run this
analysis for topics 2 to 5, the parameters that were set earlier in this section [topics ¡- 5 *
c(1:3, 5, 7)] should be changed to [topics ¡- 2 * c(1, 1.5, 2, 2.5)].
Examining the averages for each shows that 5 topics has the actual lowest perplexity
score.
\begin{minted}[breaklines, fontsize=\footnotesize]{r}
#Find the lowest average log-likelihood
p <- News_test
q <- as.data.frame(p)
rowMeans(q)
2
3
4
5
4647.33 4638.55 4644.35 4639.25
In conclusion, we can assume that the actual number of topics in the corpus is low and
lies somewhere in the range of 2 to 10.
Topic Modeling
255
FIGURE 11.4 (See color insert.): Cross-validation — optimum number of topics (2 to 5).
11.6.3
LDA Models with Gibbs Sampling
Once you know the range for the optimum number of topics in your corpus, you can then
proceed to build and test the model parameters and start interpreting the results. Since
the lowest perplexity score calculated above was for 5 topics, the model is set to produce
that number of topics. The following creates an LDA model using Gibbs sampling for topic
inference:
#CREATE AN LDA MODEL USING GIBBS SAMPLING:
library("topicmodels")
load("my_Dtm.Rdata")
#Set parameters
SEED <- 20080809
BURNIN = 1000
ITER = 1000
k = 5
#number of topics to create
#Create an LDA model using GIBBS Sampling & save the output
model_lda <- LDA(myDtm, k = k, method = "Gibbs",
control = list(seed = SEED, burnin = BURNIN, iter = ITER))
print(model_lda)
save(model_lda, file = "LDA_model.RData")
Now that we have created our LDA Gibbs model with 5 topics, we can start to examine
the output created by finding the top 10 terms in each topic as follows. Similarly we can
find the what topics are associated with what topic by using topics(model lda, 5), the
result is quite long so I have not added the results here.
#DISPLAY TERMS FROM POSTERIOR DISTRIBUTION:
#Display the top 10 terms in each topic
256
Text Mining and Visualization: Case Studies Using Open-Source Tools
terms(model_lda, 10)
Topic 1
Topic 2
Topic 3
Topic 4
Topic 5
[1,] "wire"
"god"
"imag"
"israel" "game"
[2,] "car"
"key"
"file"
"drug"
"gun"
[3,] "jim"
"encrypt" "window" "launch" "team"
[4,] "sale"
"moral"
"drive"
"station" "stephanopoulo"
[5,] "circuit" "church"
"card"
"mission" "armenian"
[6,] "cabl"
"jesus"
"color"
"orbit"
"player"
[7,] "pin"
"homosexu" "format" "arab"
"fan"
[8,] "insur"
"jew"
"graphic" "isra"
"steve"
[9,] "electr" "clipper" "jpeg"
"shuttl" "tax"
[10,] "bike"
"fbi"
"display" "food"
"turkish"
#Display the top 5 topics in each document
topics(model_lda, 5)
Looking at the first 10 terms in each of the 5 topics created, you can inspect whether
the topics make sense. A domain specialist may be required to interpret the outputs as
sometimes it is not always clear which topic has been produced. In some cases, the model
will produce topics that don’t make any sense to the human eye. In Topic 5 the word
“stephanopoulo” is associated with the topic but unless someone is familiar with the set
of documents provided, it is difficult to determine if this is an appropriate word to include
or not. If it was deemed to be not relevant, it could be worthwhile removing it from the
analysis and the DTM created earlier. In this case, there seems to be a mix of topics within
the 5 topics created so it may be worth investigating whether a smaller or larger number
of topics would produce a better output. This is where some trial and error is required and
where expertise of a domain expert would help to identify the best model fit.
After testing and creating models with topics less than and greater than 5 topics, reproducing the model using the aforementioned techniques, but increasing the number of topics
to 7 seems to give better results as shown in the following:
#DISPLAY TERMS FROM POSTERIOR DISTRIBUTION:
terms(model_lda, 10)
Topic 1
Topic 2
Topic 3
Topic 4
[1,] "window" "god"
"game"
"gun"
[2,] "drive"
"moral"
"key"
"car"
[3,] "card"
"church"
"team"
"jim"
[4,] "wire"
"jesus"
"encrypt" "bike"
[5,] "dos"
"homosexu" "player" "batf"
[6,] "driver" "fbi"
"fan"
"ride"
[7,] "disk"
"frank"
"clipper" "firearm"
[8,] "monitor" "koresh"
"insur"
"illinoi"
[9,] "mac"
"sin"
"canada" "dave"
[10,] "video"
"jon"
"score"
"bmw"
[1,]
[2,]
[3,]
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Topic 5
"imag"
"file"
"format"
"color"
"jpeg"
"display"
"ftp"
"launch"
"gif"
"graphic"
Topic 6
"stephanopoulo"
"drug"
"steve"
"tax"
"signal"
"food"
"patient"
"candida"
"pin"
"input"
Topic 7
"armenian"
"israel"
"turkish"
"muslim"
"arab"
"greek"
"isra"
"islam"
"detector"
"jew"
In this case, the output of the topics is much clearer, and we can easily distinguish
Topic Modeling
257
between the majority of the topics produced. For example, Topic 1 is “computer” related,
Topic 2 is related to “religion”, Topic 5 is possibly related to “images/photographs”, Topic
7 is “country/nationality” related. However like in the previous example, Topic 6 shows
“stephanopoulo” and “steve” as the 1st and 3rd most important word in that topic but to
the reader the meaning of the topic is still unknown.
In order to further examine the outputs, the model outputs can be saved into CSV files
which can be examined further as follows:
#SAVING MODEL OUTPUTS:
#Save the top 100 terms in each topic to csv.
write.csv(terms(model_lda, 100), file = "model_mini_news.csv")
#Save the topic terms distributions to csv
lda_terms <- posterior(model_lda)$terms
write.csv(lda_terms, file = "LDA_TERMS_mini_news.csv")
#Save the topic document distributions to csv
lda_topics <- posterior(model_lda)$topics
write.csv(lda_topics, file = "LDA_TOPICS_mini_news.csv")
To examine the terms distributions visually, the following can be used to create a quick
plot of the terms distributions across the topics Figure 11.5 shows the output for the term
distributions (code is based on example used in https://github.com/OhLookCake/
xkcd-Topics/blob/master/scripts/topicExtraction.R). As can be seen, the
first 5-10 words account for the majority of the probability that make up each topic, meaning
that only a small number of words define the topic and therefore whether a document fits into
it or not based on the words in that document. (Note that each topic assigns a probability
to every word in the corpus).
#PLOTTING THE TOPIC TERM DISTRIBUTIONS:
require(reshape2)
require(ggplot2)
require(RColorBrewer)
termgenerator <- posterior(model_lda)$terms
###1: relative probabilities of words in each topic ###
termimportance <- apply(termgenerator,1,
,→
function(x)
x[order(x,decreasing=T)[1:20]])
termimportance.longform <- melt(termimportance, value.name="Probability",
,→
varnames=c("TermNumber","Topic"))
#creating the plot
ggplot(data=termimportance.longform,
aes(
x=TermNumber,
y=Probability,
color=factor(Topic),
group=Topic)) +
geom_line()
Similarly, the topic distributions can also be examined and used as a physical check
against the documents allocated to each topic. Table 11.1 is showing the first two topics
and their document distributions for the 1st 10 documents.
Examining the actual documents shows that for Topic 1, the first document subject is
on “Electrical wiring FAQ” and is from the sci.electronics collection, the second document
is called ‘Looking for X windows on a PC’ and is from the comp.windows collection.
258
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 11.5 (See color insert.): Term distribution.
Topic Modeling
259
The first document in Topic 2 is called “Americans and Evolution” and is from
alt.atheism collection of news, the second document is named “Scroggs” and again is from
a religious related group (soc.religion.christian). This is a good way to verify that the
documents do actually belong to that topic and acts as a physical check on the accuracy of
the topics that were generated.
TABLE 11.1: Topic document distributions.
Document
53,569
68,012
52,090
52,113
76,505
76,117
52,190
51,846
76,851
51,595
11.6.4
Topic 1
0.94
0.80
0.80
0.67
0.64
0.62
0.58
0.57
0.57
0.56
Document
53,564
21,754
20,952
21,597
21,648
21,396
53,351
21,585
20,540
53,623
Topic 2
0.75
0.72
0.69
0.69
0.66
0.62
0.61
0.57
0.57
0.56
Applying New Data to the Model
The topicmodels package also allows new data to be applied to the newly created
model. This can be useful in the example where you have new documents but would like to
categorize them into the predefined topics. The following R code can be used to complete
this. In order to apply the new data, the same methods as described earlier are required
to turn the new documents into a corpus, preprocess them and finally turn them into
a document term matrix (DTM). The DTM then becomes the input into the model as
shown. Note that if the new documents contain terms that were not present in the model
training, these will not be picked up. To help overcome this, the model should not stay
static and should be retrained periodically as new documents become available.
#APPLYING NEW DATA TO THE CREATED MODEL:
#Find the posterior for model, option to add in new unseen data
# at this point.
post <- posterior(model_lda, newdata = myDtm)
#PRINT the first 10 topics from the posterior with 4 decimal places
round(post$topics[1:10,], digits = 4)
write.csv(round(post$topics, digit=4), "posterior_newdata.csv")
11.6.5
Variational Expectations Maximization (VEM) Inference
Technique
The model that was created before used Gibbs sampling as the inference technique
for creating the posterior distribution. the topicmodels package also allows for another
inference technique to be used. By using another inference method, we can then compare
and contrast the models. VEM is actually the default inference technique used in creating
260
Text Mining and Visualization: Case Studies Using Open-Source Tools
topics models in the topicmodels package. The value for alpha is estimated in the default
version (starting at 50/k).
CREATE AN LDA MODEL USING VEM INFERENCE:
library("topicmodels")
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
#Loading your document term matrix
load("my_Dtm.Rdata")
data <- myDtm
#Setting up parameters for LDA topic modeling
SEED <- 20080809
k = 7
#LDA using VEM inference
model_vem <- LDA(myDtm, k=k, control=list(seed=SEED))
print(model_vem)
terms(model_vem, 10)
save(model_vem, file = "VEM_model.RData")
Examining the outputs of the LDA model using VEM inference shows that similar topics
are created when looking at the outputs from the Gibbs inference LDA model. Topic 1 seems
related to “image/photograph”, Topic 4 looks “country/nationality” related, Topic 5 has a
“religious” theme and Topic 7 is “computer” related. However there does seem to be less
clarity when compared to the Gibbs sampling LDA model outputs. For example, Topic 4 has
the terms “car” and “bike” mixed into the “country/nationality” related theme, similarly
the term “stephanopoulo” appears in the “image/photograph” topic, again possibly pointing
out that this terms should be removed from the analysis.
Topic 1
Topic 2
Topic 3
Topic 4
[1,] "imag"
"key"
"game"
"armenian"
[2,] "file"
"encrypt"
"wire"
"car"
[3,] "jpeg"
"moral"
"team"
"turkish"
[4,] "format"
"clipper"
"player" "bike"
[5,] "color"
"jim"
"fan"
"muslim"
[6,] "stephanopoulo" "algorithm" "circuit" "turkey"
[7,] "display"
"escrow"
"electr" "armenia"
[8,] "window"
"des"
"outlet" "greek"
[9,] "gif"
"jon"
"venus"
"modem"
[10,] "ftp"
"msg"
"score"
"ride"
[1,]
[2,]
[3,]
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Topic 5
"god"
"gun"
"jesus"
"church"
"drive"
"sin"
"launch"
"pin"
"firearm"
"car"
Topic 6
"homosexu"
"israel"
"fbi"
"arab"
"jew"
"isra"
"shuttl"
"abort"
"insur"
"orbit"
Topic 7
"window"
"card"
"drive"
"dos"
"disk"
"monitor"
"driver"
"graphic"
"video"
"scsi"
Topic Modeling
11.6.6
261
Comparison of Inference Techniques
In order to examine the similarity between the two model outputs, we can use a distance
measure using Hellinger distance to calculate how similar topics are to each other. In addition to the topicmodels package, the package clue is required to compute the distance
similarities.
clue: used for creating and analyzing cluster ensembles and includes methods for measuring proximity (http://cran.r-project.org/web/packages/clue/index.html)
The following code is used to calculate the similarity between the two models (LDA
Gibbs and LDA VEM):
#COMPARISON USING HELLINGER DISTANCE:
library("topicmodels")
library("clue")
#Set the working directory
setwd("C:/Users/Documents/R/DATASET/")
#Load the models saved
load("LDA_model.Rdata")
load("VEM_model.Rdata")
#get the posterior distributions for both models
lda_terms <- posterior(model_lda)$terms
vem_terms <- posterior(model_vem)$terms
#calculate the distance between the term distributions
dist_models <- distHellinger(lda_terms, vem_terms)
matching <- solve_LSAP(dist_models)
dist_models <- dist_models[, matching]
d <- mean(diag(dist_models))
#Order by the best matching
best_match <- order(diag(dist_models))
terms(model_lda, 10)[,best_match[1:7]]
terms(model_vem, 10)[,matching[best_match[1:7]]]
#Order by worst matching
worst_match <- order(diag(dist_models), decreasing = TRUE)
terms(model_lda, 10)[,worst_match[1:3]]
terms(model_vem, 10)[,matching[worst_match[1:3]]]
After running this code, the outputs as shown below should give the 3 best matching
topic between the LDA Gibbs and LDA VEM models. As you can see, the topics are quite
well matched with the themes of “image/photograph”, “computer”, and “religion” all quite
clear. However, as mentioned earlier the Gibbs model seems to have produced better results
with more terms in the top 10 seeming to fit the theme, whereas there are more “outlier”
terms mixed into the VEM model output (e.g., stephanopoulo in Topic 1)
BEST AND WORST MATCHING TOPIC COMPARISONS:
terms(model_lda, 10)[,best_match[1:3]]
Topic 5
Topic 1
Topic 2
[1,] "imag"
"window" "god"
[2,] "file"
"drive"
"moral"
[3,] "format" "card"
"church"
262
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Text Mining and Visualization: Case Studies Using Open-Source Tools
"color"
"jpeg"
"display"
"ftp"
"launch"
"gif"
"graphic"
"wire"
"dos"
"driver"
"disk"
"monitor"
"mac"
"video"
"jesus"
"homosexu"
"fbi"
"frank"
"koresh"
"sin"
"jon"
terms(model_vem, 10)[,matching[best_match[1:3]]]
Topic 1
Topic 7
Topic 5
[1,] "imag"
"window" "god"
[2,] "file"
"card"
"gun"
[3,] "jpeg"
"drive"
"jesus"
[4,] "format"
"dos"
"church"
[5,] "color"
"disk"
"drive"
[6,] "stephanopoulo" "monitor" "sin"
[7,] "display"
"driver" "launch"
[8,] "window"
"graphic" "pin"
[9,] "gif"
"video"
"firearm"
[10,] "ftp"
"scsi"
"car"
Looking at the output for the three least or worst matching topics shows that to the
observer there is nothing similar at all between the topics. In fact, it is quite difficult to
determine the theme for the majority of the topics shown. In the LDA Gibbs, Topic 7 has
a clear enough theme and related terms, Topic 6 and 4 may have a theme that is quite
specific so a domain expert might be required to help identify it.
terms(model_lda, 10)[,worst_match[1:3]]
[1,]
[2,]
[3,]
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Topic 6
"stephanopoulo"
"drug"
"steve"
"tax"
"signal"
"food"
"patient"
"candida"
"pin"
"input"
Topic 7
"armenian"
"israel"
"turkish"
"muslim"
"arab"
"greek"
"isra"
"islam"
"detector"
"jew"
Topic 4
"gun"
"car"
"jim"
"bike"
"batf"
"ride"
"firearm"
"illinoi"
"dave"
"bmw"
terms(model_vem, 10)[,matching[worst_match[1:3]]]
[1,]
[2,]
[3,]
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Topic 2
"key"
"encrypt"
"moral"
"clipper"
"jim"
"algorithm"
"escrow"
"des"
"jon"
"msg"
Topic 6
"homosexu"
"israel"
"fbi"
"arab"
"jew"
"isra"
"shuttl"
"abort"
"insur"
"orbit"
Topic 4
"armenian"
"car"
"turkish"
"bike"
"muslim"
"turkey"
"armenia"
"greek"
"modem"
"ride"
For the VEM models, identifying a theme for the topics is more difficult as there doesn’t
seem to be anything evident and the topics look like a mismatch of terms. Again the inclusion
of names seems to be hindering the underlying theme. Although the names can indeed be
important to a topic, for someone who is not familiar with the data in the documents, it
can be difficult to extract the theme.
Topic Modeling
11.6.7
263
Removing Names and Creating an LDA Gibbs Model
As was noted in the previous section, the models produced contain names that can
obscure the topic. In this section, some further preprocessing is applied to the data set
to remove them from the corpus. In addition to the stopwords dictionaries that were used
earlier in preprocessing, names that occurred frequently are added as follows. Note that in
the topic models produced, some of the words will have been stemmed so in the case of
“stephanopoulo”, the full name “stephanopoulos” would need to be added.
myStopwords <- c(stopwords("english"), stopwords("SMART"),
"stephanopoulos","frank", "jon", "steve", "jim", "dave", "des",
"brian", "tom", "andi")
After preprocessing the data and removing the names, an LDA model using Gibbs
sampling was produced for 7 topics and gave the following results:
Topic 1
Topic 2
Topic 3 Topic 4
[1,] "imag"
"window" "game"
"key"
[2,] "file"
"drive"
"team"
"encrypt"
[3,] "format" "card"
"player" "armenian"
[4,] "jpeg"
"dos"
"fan"
"clipper"
[5,] "color"
"disk"
"tax"
"fbi"
[6,] "display" "driver" "score" "turkish"
[7,] "ftp"
"monitor" "trade" "muslim"
[8,] "israel" "video"
"drug"
"koresh"
[9,] "gif"
"graphic" "abort" "islam"
[10,] "arab"
"sale"
"insur" "algorithm"
[1,]
[2,]
[3,]
[4,]
[5,]
[6,]
[7,]
[8,]
[9,]
[10,]
Topic 5
"wire"
"gun"
"car"
"circuit"
"cabl"
"pin"
"bike"
"electr"
"switch"
"clock"
Topic 6
"launch"
"station"
"mission"
"orbit"
"shuttl"
"venus"
"detector"
"bbs"
"pat"
"cub"
Topic 7
"god"
"moral"
"church"
"jesus"
"homosexu"
"jew"
"sin"
"atheist"
"mormon"
"pit"
Examining the topics show two new topics that up to now had not been picked up. Topic
3 looks to be “sports” related and Topic 6 seems to be related to “space”. The “image”,
“computer” and “religion” topics have remained. Topics 4 and 5 are slightly less obvious
and could possibly be a mix of topics. Overall this model looks to be a better fit than the
earlier LDA Gibbs and LDA VEM models.
11.7
Summary
In this chapter, the popular topic modeling technique LDA was examined and modeled
using the R statistical tool. Using the newsgroup data, the process of loading and transforming a set of document into a corpus was shown. Various preprocessing techniques were
applied to the corpus prior to converting it into a document term matrix so that it could
be used in topic modeling.
Topic modeling was then applied using the topicsmodels package. The analysis showed
264
Text Mining and Visualization: Case Studies Using Open-Source Tools
that small numbers of topics reflected more accurately the structure of the data set. Overall, it was found that the model produced quite good topics that were for the most part
coherent. The process for creating a model can involve some trial and error, in that as you
develop models you learn more about the corpus, which then helps to inform changes and
improvements that enhance the final model.
Bibliography
[1] M. Steyvers and T. Griffiths. Probabilistic topic models. Handbook of Latent Semantic
Analysis, 427(7):424–440, 2007.
[2] D. J. Hu. Latent dirichlet allocation for text, images, and music. 2009.
[3] D. M. Blei, A. Y. Ng, and M.I. Jordan. Latent dirichlet allocation. The Journal of
Machine Learning Research, 3:993–1022, 2003.
[4] D.M. Blei and J.D. Lafferty. Topic models. Text Mining - Classification, Clustering,
and Applications, 10:71, 2009.
[5] D.M.
Blei.
Probabilistic
topic
models,
http://www.cs.princeton.edu/ blei/blei-mlss-2012.pdf, 2012.
available
from
[6] B. Grun and K. Hornik. topicmodels — an r package for fitting topic models. Journal
of Statistical Software, 40, 2011.
Chapter 12
Empirical Analysis of the Stack Overflow
Tags Network
Christos Iraklis Tsatsoulis
Nodalpoint Systems, Athens, Greece
12.1
12.2
12.3
12.4
12.5
12.6
12.7
12.8
12.9
12.1
Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Acquisition and Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Full Graph — Construction and Limited Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Clusters and Connected Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Node Degrees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Clustering Coefficient, Path Length, and Small-world Characteristics . . . . . . . .
Reduced Graph — Construction and Macroscopic Analysis . . . . . . . . . . . . . . . . . .
General Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Node Importance: Centrality Measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Betweenness Centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Closeness Centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Eigenvector-based Centrality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Community Detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Fast Greedy Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Infomap Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Visualization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Visualizing the Communities Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Egocentric Visualizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Appendix: Data Acquisition & Parsing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Data Acquisition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing the Tag List & Frequencies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Parsing the Tag Co-occurences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Building the Adjacency Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
265
267
268
269
270
271
272
272
273
273
275
276
277
277
279
282
283
284
285
290
290
290
291
291
292
Introduction
Stack Exchange is self-described as “a fast-growing network of 129 question and answer
sites on diverse topics from software programming to cooking to photography and gaming”,
with a current base of more than 5.5 million users and over 82 million monthly unique
visitors1 . By far, the greatest among these 129 sites is Stack Overflow, devoted to questions
about computer programming and related topics.
1 http://stackexchange.com/.
265
266
Text Mining and Visualization: Case Studies Using Open-Source Tools
Question tags are an essential part of the submitted questions in Stack Overflow, allowing answering users to monitor questions relevant to their field of expertise and to answer
promptly to submitted questions. Only users holding an “advanced” status in Stack Overflow are allowed to generate new tags, with the rest of the community limiting themselves
in using the existing tags for their questions. Tagging the submitted questions is mandatory
(i.e., there is a minimum of one tag per question), and each question can contain up to five
tags.
In this chapter, we depart slightly from text analysis proper in order to undertake a task
inspired by the recent advances in what is usually termed as “network science” : hence, we
carry out an analysis of the Stack Overflow tags viewed as a network, or a graph. Specifically,
we aim to get some insight about the user communities by representing tags and their cooccurrences as a graph, where the graph nodes are the tags themselves, and an edge between
two nodes exists if the corresponding tags are found together in the same Stack Overflow
question. The resulting network edges are weighted, i.e., the more questions exist with two
tags co-occurring, the higher the weight of the relevant edge between them will be; for
example, if the tag python is found to co-exist with the tag android in 500 questions,
the weight of the edge between the nodes python and android in our graph will be 500.
In graph terminology, the resulting graph is a weighted undirected one.
What is the motivation behind our approach? The recent development of network science
[1], largely triggered by the publication of two seminal papers in the late ’90s [2, 3], has since
made clear that the graph representation of data and processes is certainly of benefit, and
can lead to insights not directly or easily achievable otherwise. Our scope is to demonstrate
the merit of such a representation in a text-related area and to possibly alert the interested
reader for the availability of the relevant principles, tools, and techniques.
There are several introductory textbooks available on network science: Easley and Kleinberg [4] aim for an advanced undergraduate audience and have made their book freely
available online;2 the books by Jackson [5] and Newman [6] are standard graduate-level
introductions to the subject, from the perspectives of economics and statistical physics respectively; Kolaczyk [7] provides another graduate-level introduction, with emphasis to the
statistical analysis of network data; and Barabási [1] is gradually making available an online
introductory book as a work in progress. We also highly recommend the popular book by
Duncan Watts [8] for a smooth, non-mathematical introduction to the relevant concepts,
notions, and ideas. We will not provide here a formal introduction to graph theory and
network science, something that would require a separate chapter by itself; instead, we will
introduce rather informally and intuitively the relevant concepts as we go along. For formal definitions, the interested reader can always consult some of the free online resources
mentioned above. For the purposes of our discussion, we will use the terms “network” and
“graph” loosely and interchangeably, as if they were synonyms (although they are not). The
other interchangeable pairs of terms we will use are “node” – “vertex” and “edge” – “link”.
We have chosen R [9] for our investigation, as it has become the de facto standard tool
for the statistical analysis of data. Among the several available packages for graph and
network analysis, we utilize the igraph library [10], a powerful and flexible tool which is
already used by at least a hundred other R packages [11].
The rest of this chapter is organized as follows: in the next section, we describe briefly the
data acquisition process and present some summary statistics of the raw data; in Section
3 we expose the construction of our first graph based on the whole data set, along with
some limited analysis; Section 4 describes the meaningful reduction of our raw data set,
which is used in the rest of the chapter; Section 5 deals with various measures of node
importance (centrality), and Section 6 we demonstrate the application of some community
2 http://www.cs.cornell.edu/home/kleinber/networks-book/.
Empirical Analysis of the Stack Overflow Tags Network
Number of records (questions)
Number of unique tags
No. of tag occurrences
Average no. of tags/question
267
7,214,697
36,942
21,294,348
2.95
TABLE 12.1: Statistics of raw data.
detection algorithms; some attempted visualizations are presented in Section 7, utilizing
the communities detected in Section 6; Section 8 concludes the chapter with some general
discussion.
12.2
Data Acquisition and Summary Statistics
Stack Exchange makes its updated “data dump” periodically available roughly twice a
year (January & September), under a Creative Commons BY-SA 3.0 License.3 The version
we are going to work with is of January 2014. The questions data are available as a large (˜
25GB) XML file, which was downloaded and parsed using BaseX, an open-source lightweight
XML database4 . A separate XML file with the tag frequencies is also provided. The details
of the raw data parsing and processing can be found in the Appendix; the results are:
• A list of the used tags, ordered by decreasing frequency
• An adjacency matrix keeping track of the tag co-occurrences, as described in the
Introduction.
These results are available as R workspace files (tag freq.RData and adj matrix.RData
respectively) from the author’s Github repository5 , and they constitute the departure point
for our analysis; as said, readers interested in the parsing procedure itself and the construction of these results from the raw data can consult the Appendix; the R scripts used for the
parsing are also available at Github, so the whole procedure is entirely reproducible, and
even extensible as new versions of the Stack Overflow data dump become available in the
future. The summary statistics of the raw data are shown in Table 12.1.
The tag freq dataframe contains two variables: the name of the tag and the corresponding frequency (counts) in the data set. We will load it and add a third variable, to
indicate the relative percentage frequency of each tag, and then view the 10 most frequent
tags:
> load("tag_freq.RData") # needs to be in the current working directory
> tag_freq$rel <- tag_freq$freq/sum(tag_freq$freq)*100
> head(tag_freq,10)
tag
freq
rel
1
c# 635338 2.983599
2
java 632575 2.970624
3 javascript 605552 2.843722
4
php 573279 2.692165
3 http://blog.stackexchange.com/category/cc-wiki-dump/.
4 http://basex.org/.
5 https://github.com/desertnaut.
268
5
6
7
8
9
10
Text Mining and Visualization: Case Studies Using Open-Source Tools
android
jquery
python
c++
html
mysql
503255
468723
297952
287483
284234
243701
2.363327
2.201162
1.399207
1.350044
1.334786
1.144440
We can thus see the 10 most frequent tags, along with their absolute and relevant
frequencies. Going a little further,
> sum(tag_freq$rel[1:10])
[1] 21.28307
we can easily see that the 10 most frequent tags account for about 21% of all tag occurrences
in the raw data.
We can similarly explore the other end of the list, that is how many tags appear very
infrequently in the data set:
> length(which(tag_freq$freq < 100))
[1] 26723
> length(which(tag_freq$freq < 10))
[1] 9614
We can thus see that the majority of our 36,942 tags appear fewer than 100 times,
whereas about 26% of the tags have a rather negligible presence (fewer than 10 times) in
our data.
These results suggest a highly skewed distribution, with relatively few frequent tags
dominating the data set. Plotting the tag frequencies for the first 100 tags (recall that tags
are indexed in decreasing frequency) confirms this intuition, as shown in Figure 12.1 :
> plot(tag_freq$freq, xlim=c(1,100),
+ xlab="Tag index", ylab="Frequency (counts)")
12.3
Full Graph — Construction and Limited Analysis
Having performed the initial exploration of our tag frequency distribution, we now proceed to construct the full graph using the adjacency matrix already built as shown in
Appendix A (file adj matrix.RData). The igraph library, from version 0.5.1 on, utilizes fully the sparse matrix infrastructure provided by the Matrix package [12], something
extremely convenient, since working with dense matrices of size 40,000 x 40,000 can be
really cumbersome for most common desktop computers.
> library(Matrix)
> library(igraph)
> load("adj_matrix.RData") # contains sparse matrix G
> g <- graph.adjacency(G, weighted=TRUE, mode=c("plus"), diag=FALSE)
> summary(g)
IGRAPH U-W- 36942 2699465 -attr: weight (e/n)
Empirical Analysis of the Stack Overflow Tags Network
269
FIGURE 12.1: Tag frequency distribution for the first 100 tags.
The summary() function provides some general information about the graph g, namely,
that it is undirected (’U’) and weighted (’W’), with 36,942 vertices (nodes) and about 2.7
million edges, and only one attribute (weight), which refers to the edges (’e’) and it is a
numeric one (’n’). We can decorate the graph with more attributes, such as a name for it
as well as names for each vertex (tag):
> # set the graph and vertices name attributes:
> g$name <- "FULL TAG GRAPH"
> V(g)$name <- tag_freq$tag # set tags as node names
> summary(g)
IGRAPH UNW- 36942 2699465 -- FULL TAG GRAPH
attr: name (g/c), name (v/c), weight (e/n)
We can see that our graph is now named (’N’), and it contains two more character
attributes (’c’), one for the whole graph (’g’) and one for the individual vertices (’v’).
Clusters and Connected Components
One of the first things in network diagnostics is to check if the network is globally
connected:
> is.connected(g)
[1] FALSE
Since our graph is not globally connected, the next step is to check for its connected
components (clusters):
> cl <- clusters(g)
> str(cl)
270
Text Mining and Visualization: Case Studies Using Open-Source Tools
List of 3
$ membership: num [1:36942] 1 1 1 1 1 1 1 1 1 1 ...
$ csize
: num [1:53] 36888 1 1 1 1 ...
$ no
: int 53
> cl$csize
[1] 36888
1
1
1
1
2
1
1
[9]
1
1
1
1
1
1
1
1
[17]
1
1
1
1
1
1
1
1
[25]
1
1
1
1
1
1
1
1
[33]
2
1
1
1
1
1
1
1
[41]
1
1
1
1
1
1
1
1
[49]
1
1
1
1
1
From the csize attribute of our clusters variable, it is apparent that our graph consists
of one giant connected component (36,888 nodes) and 52 clusters with only one to two
nodes each. It is trivial to write a short loop, in order to check the tag names in these small
isolated clusters (only the first 10 shown here, for brevity):
# print isolated tags ("graph periphery")
l <- cl$no # number of clusters
for (i in 2:l) {
label <- V(g)$name[which(cl$membership==i)]
print(label)
}
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
"untagged"
"beaker-testing"
"arden-syntax"
"zzt-oop"
"ontopia" "tolog"
"ebuild"
"nikeplus-api"
"infinity.js"
"jquery-ias"
"paperless"
Even without specific domain knowledge, it is apparent that all the tags in the “graph periphery” listed already are far from common terms expected to be found in a programmingrelated forum.
Node Degrees
The degree of a node is simply the number of its direct neighbors, or equivalently the
number of the links in which the node is involved [5, 6]. The degree is very frequently used
as a proxy for the importance or the significance of individual nodes in a graph. igraph
provides a relevant function degree(g), which can be used to compute the degrees of all
the nodes in the graph g; nevertheless, it is of no use in our case here, as it does not take into
account the weights of the graph edges. The correct igraph function for the computation
of the weighted degree [13] is graph.strength(g). Again, it is straightforward to see the
nodes with the highest weighted degrees in our graph:
> s <- graph.strength(g)
> V(g)$name[which.max(s)]
[1] "c#"
> s <- sort(s, decreasing=TRUE)
Empirical Analysis of the Stack Overflow Tags Network
271
> s <- as.data.frame(s) # for better display
> head(s, 10)
s
c#
1485191
java
1446078
javascript 1359147
php
1192764
jquery
1034841
android
965297
html
705233
c++
639678
ios
621107
python
608211
Not unsurprisingly, the above list is similar (but not identical) to the 10 most frequent
tags already shown. If we plot the weighted degree, the results are qualitatively very similar
with Figure 12.1 (not shown).
Clustering Coefficient, Path Length, and Small-world Characteristics
The local clustering coefficient of a node is the fraction of its direct neighbors that
are themselves direct neighbors (i.e., directly connected); in a social network framework,
for example, it would be the fraction of one’s friends that are also friends themselves [4].
The average clustering coefficient of the whole graph is simply the average of the nodes’
local clustering coefficients [5]. By definition, both coefficients lie between 0 and 1. igraph
provides the function transitivity() for the computation of these quantities, as well
as a simple function for the computation of the average path length in a graph. For these
calculations, we ignore the edge weights, treating the edges between nodes simply as binary
entities (present/absent):
# Clustering coef & average path length
# Use type="average" argument for average cluster coef
> cc <- transitivity(g, type="average", isolates="zero")
> cc
[1] 0.6099061
> path.length <- average.path.length(g, unconnected=TRUE)
> path.length
[1] 2.341709
Thus, on average, the “distance” between any two nodes in our graph is between 2 and 3
hops, making our graph a small-world network [4, 5, 6]. Before commenting on the clustering
coefficient, we take a moment to compute the same quantities for a random (Erdös-Renyi ER) graph with the same number of nodes and edges:
# Compare with an equivalent random (ER) graph
> set.seed(42) # for reproducibility
> g.ER <- erdos.renyi.game(length(V(g)), length(E(g)), type="gnm")
> cc.ER <- transitivity(g.ER, type="average", isolates="zero")
> cc.ER
[1] 0.00395176
> path.length.ER <- average.path.length(g.ER, unconnected=TRUE)
> path.length.ER
[1] 2.554753
We can see that, compared with a random graph with the same number of nodes and
272
Text Mining and Visualization: Case Studies Using Open-Source Tools
edges, our graph possesses roughly the same average path length between two nodes, with a
hugely higher (two orders of magnitude) clustering coefficient. This is a recurrent motif for
several real-world networks, including social ones [6], first captured in an abstract network
model by Watts and Strogatz in a seminal paper published in Nature [2], that partially
triggered the subsequent birth of the whole network science field.
12.4
Reduced Graph — Construction and Macroscopic Analysis
We have presented some general analysis for our full graph, but the information revealed
so far is rather trivial. To go deeper, we will need to revert to a reduced graph, due to the
prohibitively high computational power demanded for the analysis of the full tag graph; due
to the highly skewed distribution of the tag occurrences (see Section 12.2), we claim that
this reduction still captures almost all of the interesting characteristics of our data set.
We suggest that we proceed with a graph containing only those tags which occur a
minimum of 500 times in our data (recall that we have more than 7 million questions
and 21 million tag occurrences in total; see Table 12.1). We have taken special care when
building our adjacency matrix, so that its columns and rows extend in a decreasing tag
frequency; hence, it is a trivial operation to discard the following tags a given frequency in
the adjacency matrix:
> # select only tags with freq >= 500
> ind <- which(tag_freq$freq >= 500)
> max(ind)
[1] 3711
> tag_freq <- tag_freq[ind,]
> sum(tag_freq$freq)
[1] 19166670
> sum(tag_freq$rel)
[1] 90.00825
> # keep only these entries in the adjacency matrix
> G <- G[1:max(ind), 1:max(ind)]
There are thus 3711 tags in our reduced list, or about 10% of all tags used; the sum()
functions in the code above show that our reduced tag list contains more than 19 million
tag occurrences out of the 21 million in total, i.e., about 90%. Proceeding with a manner
absolutely similar to the one presented for the full graph in Section 12.3, we end up with
our reduced graph gr, containing about 1 million edges:
> summary(gr)
IGRAPH UNW- 3711 1033957 -- REDUCED TAG GRAPH
attr: name (g/c), name (v/c), weight (e/n)
We have thus now a much more “lightweight” graph, nevertheless capturing the vast
majority of the most frequently used tags. We are ready now to apply some more sophisticated techniques from the toolbox of graph theory and network science in order to reveal
possible unexpected information and insights, hidden in the network structure.
General Statistics
> is.connected(gr)
[1] TRUE
Empirical Analysis of the Stack Overflow Tags Network
273
> cc <- transitivity(gr, type="average", isolates="zero") #clustering coef.
> cc
[1] 0.4992685
> path.length <- average.path.length(gr)
> path.length
[1] 1.849807
> # Compare with the equivalent ER graph
> set.seed(42) # for reproducibility
> gr.ER <- erdos.renyi.game(length(V(gr)), length(E(gr)), type="gnm")
> is.connected(gr.ER)
[1] TRUE
> cc.ER <- transitivity(gr.ER, type="average", isolates="zero")
> cc.ER
[1] 0.1501704
> path.length.ER <- average.path.length(gr.ER)
> path.length.ER
[1] 1.849801
We can see that our reduced graph is now globally connected (i.e., there are no isolated
nodes); it has an even lower average path length between two nodes (1-2 hops), and its
clustering coefficient is still higher than the one of the equivalent random graph, although
this difference is not as marked as with our full graph.
In what follows, we will limit the discussion to the reduced tag network we have just
constructed.
12.5
Node Importance: Centrality Measures
We recall from the previous sections that the information revealed so far is rather trivial:
taking the weighted node degree as a rough importance measure, the most important nodes
essentially coincide with the most frequent tags; clearly, we did not need to go to a graph
representation in order to discover something like that.
The graph representation starts paying off when we consider other measures of node
importance, which in the networks literature usually go under the label of centrality measures (actually, the degree centrality, which we have already demonstrated, is just one of
them). Furthermore, these measures, in contrast with the macro measures considered so
far (connected components, average clustering coefficients and path lengths), are in essence
micro measures, since they do not refer to the graph as a whole, but rather to individual
nodes and how they relate to the overall network structure [5].
There are several measures of node centrality in the literature, and most of them are
implemented in igraph. We now proceed to have a look to some of them.
Betweenness Centrality
The betweenness centrality, first proposed in the context of social network analysis by
Freeman [14], measures the extent to which a node is located in the shortest paths between other pairs of nodes (i.e., ”between” them) [5, 11]. As with other centrality measures, igraph provides two ways of calculation: the straightforward betweenness()
function, which can be computationally demanding, especially for large graphs; and the
betweenness.estimate() function, with a cutoff argument for providing the maximum
274
Text Mining and Visualization: Case Studies Using Open-Source Tools
path length to consider in the calculation. Although the direct calculation of betweenness
is not prohibitive for our graph, we will hereby demonstrate the usage of the approximate
function, which will most probably be of use in a range of other situations, where the graphs
are prohibitively large for the direct application of the betweenness() function. In our
case, the reader can confirm that the approximate calculation gives results identical to the
exact one. Both functions take into account the edge weights [13].
A reasonable value to use as a cutoff value for the path length is the diameter of the
graph, i.e., the maximum (shortest) path length between two nodes [5]:
# calculate (approximate) betweenness
> diam <- diameter(gr)
> diam
[1] 3
> bc <- betweenness.estimate(gr, cutoff=diam)
> which.max(bc)
[1] 111
> V(gr)$name[which.max(bc)]
[1] "debugging"
Keeping always in mind that our node indices are sorted as per their occurrence frequency it the data set, the meaning of the which.max(bc) above should be apparent:
the node with the highest betweenness centrality in our graph corresponds to the tag
debugging, which is ranked only 111th in our tag list.
In hindsight, it seems appropriate that the tag linking most of the questions in a computer programming forum should be something that has to do with errors or bugs, but we
wonder how apparent this fact could be beforehand.
With a little data manipulation, it is easy to see more of the nodes with high betweenness centrality, and compare their betweenness ranking to the frequency ranking of the
corresponding tags:
>
>
>
>
# see betweenness ranking VS frequency ranking:
bc.df <- data.frame(bc=bc, freq.rank = 1:length(V(gr)))
bc.df <- bc.df[order(bc.df$bc, decreasing=TRUE),] #order in decreasing bc
head(bc.df, 20) # show the top-20
bc freq.rank
debugging
20363.39
111
design
19011.70
198
caching
17694.07
162
exception
17655.67
142
testing
16669.95
183
error-handling
16482.48
370
dynamic
15932.37
223
api
14840.20
78
data
14178.36
261
unit-testing
14144.10
101
user-interface
14009.76
185
parsing
13927.16
105
filter
13884.12
390
web-applications 13834.88
210
optimization
13730.19
201
memory-leaks
13653.39
305
logging
13646.06
207
object
13535.99
144
web
13496.93
254
text
13487.15
230
Empirical Analysis of the Stack Overflow Tags Network
275
Thus, not only the highest betweenness node is nowhere to be seen in the 100 most
frequent tags, but also none of the 10 most frequent tags scores high in betweenness centrality; we can see that, in fact, with the exception of the tag api, none of the 20 tags
with the highest betweenness centrality is in the top 100 frequency list. We can go a little
further, and check in particular how exactly the 10 most frequent tags score and rank in
betweenness centrality:
> bc.df$bc.rank <- 1:length(V(gr)) # add bc.rank field
> ind <- which(bc.df$freq.rank <= 10) # find the 10 most frequent tags
> bc.df[ind,] # display them
bc freq.rank bc.rank
mysql
7035.926
10
178
html
6295.522
9
226
jquery
5764.672
6
263
android
5645.326
5
269
javascript 4691.252
3
387
c++
4147.752
8
440
python
3967.835
7
465
c#
3665.712
1
511
php
3539.976
4
538
java
2007.629
2
914
Hence, we can explicitly see that all 10 most frequent tags rank lower than 100 in
betweenness centrality, with the 2nd most frequent one (java) barely making it to the
top-1000 of betweenness centrality.
Interestingly enough, all the high betweenness node-tags shown above are somewhat
“general” ones, in the sense that they are not related with specific software or hardware
platforms. This makes perfect sense intuitively, and it demonstrates how the network structure can capture relations and characteristics in the data that are not easily or directly
expressed in simple aggregate measures, such as the frequency of occurrence.
As said earlier, the reader can easily confirm that, for our graph, the approximate betweenness coincides with the exact one (which nevertheless would be computationally prohibitive for larger graphs):
> bc.exact <- betweenness(gr)
> length(bc==bc.exact)
[1] 3711
The last statement means that the condition bc==bc.exact is true for all our 3711
nodes.
Closeness Centrality
The closeness centrality of a node is a measure of how ‘close’ a node is to all the
other nodes in the network [5, 11]. Since ours is a graph with a rather low diameter
(3), we do not expect this measure to be of particular usefulness, but we include it here
for illustrative purposes. Following a procedure completely analogous with the one already demonstrated for the betweenness centrality, but using instead the igraph function
closeness.estimate(), we arrive at very similar results:
> head(clc.df, 20)
design
error-handling
clc freq.rank clc.rank
0.0001497230
198
1
0.0001496110
370
2
276
Text Mining and Visualization: Case Studies Using Open-Source Tools
debugging
0.0001485884
111
filter
0.0001484120
390
caching
0.0001483900
162
testing
0.0001483680
183
documentation
0.0001483239
593
exception
0.0001480604
142
logging
0.0001478415
207
dynamic
0.0001478197
223
web
0.0001477541
254
optimization
0.0001476887
201
runtime-error
0.0001474926
1021
parameters
0.0001474709
339
memory-leaks
0.0001474491
305
data
0.0001474274
261
user-interface 0.0001473188
185
frameworks
0.0001473188
429
crash
0.0001473188
506
design-patterns 0.0001472971
179
> clc.df[ind,] # the 10 most frequent tags:
clc freq.rank clc.rank
mysql
0.0001414027
10
778
html
0.0001408848
9
915
jquery
0.0001405284
6
1031
c++
0.0001394117
8
1488
android
0.0001385425
5
1944
python
0.0001382361
7
2128
php
0.0001372684
4
2810
javascript 0.0001366680
3
3222
c#
0.0001355381
1
3601
java
0.0001308729
2
3710
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Eigenvector-based Centrality
Eigenvector-based centrality measures express the importance of a node based on the
importance of its neighbors. Since this definition is somewhat self-referential, these measures
turn out to be the solutions of some appropriately defined eigenvalue problem (hence the
name) [5, 11]. The relevant igraph function is evcent(), which has a somewhat different
interface than the betweenness() and closeness() centrality functions already used:
> eigc <- evcent(gr) # returns a list, not an array
> eigc <- eigc$vector # coerce to array
> V(gr)$name[which.max(eigc)]
[1] "javascript"
Following a similar procedure to the one already described for the betweenness, we check
the 20 nodes with the highest eigenvector centrality (notice that this measure is normalized
to lie between 0 and 1):
> head(eigc.df,20)
eigc freq.rank
javascript 1.00000000
3
jquery
0.94831551
6
html
0.72171495
9
css
0.51204736
13
Empirical Analysis of the Stack Overflow Tags Network
php
ajax
mysql
c#
asp.net
html5
json
java
forms
jquery-ui
sql
arrays
.net
android
regex
css3
0.44673099
0.28340051
0.19656924
0.18316147
0.18129922
0.13735764
0.13219053
0.10853696
0.09822225
0.09133711
0.08945679
0.08822174
0.08759265
0.08142784
0.07787668
0.07391215
277
4
22
10
1
12
39
28
2
46
84
14
24
16
5
25
83
We notice that the results are very different compared to the betweenness and closeness
centralities: javascript and jquery are pronounced as the most important nodes, and
the leaderboard is populated by a mix of the most frequent and the less so tags.
12.6
Community Detection
Community detection is both a huge topic and the subject of intense current research: a
recent (2010) survey paper in a high-caliber physics journal runs for no less than 100 pages
[15], where we read that “[t]his problem is very hard and not yet satisfactorily solved, despite
the huge effort of a large interdisciplinary community of scientists working on it over the
past few years”. The subject also goes under some alternative labels, such as clustering and
graph partitioning [11, 15], although it is advised that the former term be avoided in order
to prevent confusion with the clustering coefficient measures mentioned already [6].
Intuitively speaking, the problem of community detection is to uncover a somehow “natural” division of a network into groups of nodes, such that there are many edges within
groups and relatively few edges between groups [6]. Like the measures introduced in the
first sections of this chapter, it is a macro one, aiming at discovering and understanding
the large-scale structure of graphs and networks. Although the first algorithms for community detection and graph partitioning in computer science were proposed in the early
1970’s, recent research on the topic was greatly triggered by the publication of a seminal
paper by Girvan and Newman in 2002 [16], considered as “historically important, because
it marked the beginning of a new era in the field of community detection and opened this
topic to physicists” [15]. Their method focuses on the concept of betweenness centrality
introduced already, and it remains highly popular today. It is implemented in igraph by
the function edge.betweenness.community(), but unfortunately it is computationally
very demanding, and we will not demonstrate it here. We will focus on two other popular
algorithms instead.
Fast Greedy Algorithm
The fast greedy algorithm was proposed in 2004 by Clauset, Newman, and Moore [17];
hence, it is sometimes referred to as the CNM algorithm. It tries to optimize a modularity
278
Text Mining and Visualization: Case Studies Using Open-Source Tools
measure for the graph [18] by applying, as its name implies, a fast greedy approach. It is
available in igraph via the fastgreedy.community() function:
> fc <- fastgreedy.community(gr)
All the community detection functions in igraph return their results as objects of the
dedicated class communities, which provides its own special functions and operations for
this kind of objects
> length(fc) # how many communities
[1] 10
> sizes(fc)
Community sizes
1
2
3
4
5
6
7
8
9 10
32 514 625 506 976 316 521 36 183
2
We can see that the fast greedy algorithm returns 10 communities, with the sizes shown
above. Before getting for a closer inspection, we can see that most of the detected communities consist of a rather high number of nodes. To get an idea of the quality of the returned
communities, we proceed to visually inspect some of the smaller ones:
> V(gr)$name[membership(fc)==1]
[1] "actionscript-3" "flash"
[4] "actionscript"
"air"
[7] "salesforce"
"adobe"
[10] "flex3"
"components"
[13] "actionscript-2" "textfield"
[16] "flex4.5"
"apex-code"
[19] "flash-player"
"flexbuilder"
[22] "visualforce"
"spark"
[25] "flv"
"builder"
[28] "flash-cs4"
"blazeds"
[31] "swfobject"
"flash-cs6"
"flex"
"flex4"
"flash-builder"
"swf"
"flash-cs5"
"mxml"
"movieclip"
"red5"
"loader"
"flashdevelop"
> V(gr)$name[membership(fc)==8]
[1] "excel"
"vba"
[3] "excel-vba"
"macros"
[5] "google-apps-script"
"ms-word"
[7] "automation"
"export"
[9] "google-drive-sdk"
"ms-office"
[11] "google-spreadsheet"
"range"
[13] "excel-formula"
"excel-2007"
[15] "word"
"excel-2010"
[17] "powerpoint"
"cell"
[19] "google-drive"
"spreadsheet"
[21] "formula"
"google-docs"
[23] "copy-paste"
"word-vba"
[25] "pivot-table"
"google-apps"
[27] "xls"
"google-docs-api"
[29] "openoffice.org"
"ole"
[31] "outlook-vba"
"excel-2003"
[33] "shapes"
"powerpoint-vba"
[35] "google-spreadsheet-api" "vlookup"
> V(gr)$name[membership(fc)==10]
[1] "tridion"
"tridion-2011"
Empirical Analysis of the Stack Overflow Tags Network
279
All three communities exposed above look meaningful: for example, community #1 seems
to be about Adobe Flash and related technologies, whereas community #8 contains mostly
tags related to “office” applications (by Microsoft, Google, and Apache). Nevertheless, we
find it difficult to imagine that the bigger communities, i.e., of size 300, 500, or even higher,
may contain such kind of coherent information that could possibly be labeled by a single
term (like “Adobe Flash”, or “office applications” for our two previous examples).
Could we possibly do better? The answer is not clear beforehand; recall from the discussion so far that, edge weights aside, our graph is densely connected, with a diameter of
only 3. It could very reasonably be the case that, given this “tightness” of our graph, we
cannot uncover any finer details regarding distinct coherent communities.
To see that this is not the case, we now turn to a powerful, state of the art algorithm,
that at the same time is not computationally prohibitive for a common desktop or laptop
computer, at least for the size of our graph.
Infomap Algorithm
The Infomap algorithm for community detection was suggested in 2008 by Rosvall and
Bergstrom [19, 20]; according to comparative tests, it “appears to be the best”, and it is “also
very fast, with a complexity which is essentially linear in the system size, so [it] can be applied
to large systems” [15]. It is available in igraph via the infomap.community() function.
Since it has a random element, we first set the random seed for exact reproducibility of the
results shown here:
> set.seed(1234) # for reproducibility
> imc <- infomap.community(gr)
> length(imc)
[1] 80
> plot(sizes(imc), xlab="Community no.", ylab="Community size")
Infomap reveals 80 communities, with the corresponding sizes as shown in Figure 12.2.
Despite the fact that here also we have some large communities detected, we have also
a rather high number of smaller ones, with sizes ranging from 2 to about 40 nodes each.
In our case, the only way for validating the communities is by “manual inspection”,
that is, by going through them and examining if the communities are meaningful. We claim
that the communities discovered by the Infomap algorithm are indeed highly meaningful;
we leave the complete set of this exercise to the reader, and we will illustrate here only some
typical cases.
Before examining the large communities, it makes sense to first examine the smaller
ones, as the possible (in)coherences will be much easier to spot. Let us begin with the
communities #17-19:
> V(gr)$name[membership(imc)==17]
[1] "facebook"
"facebook-graph-api"
[3] "comments"
"facebook-like"
[5] "facebook-javascript-sdk" "facebook-fql"
[7] "facebook-php-sdk"
"share"
[9] "facebook-opengraph"
"facebook-c#-sdk"
[11] "opengraph"
"social-networking"
[13] "like"
"photo"
[15] "facebook-apps"
"facebook-login"
[17] "facebook-android-sdk"
"facebook-ios-sdk"
[19] "access-token"
"facebook-comments"
[21] "sharing"
"social"
280
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 12.2: Communities revealed by Infomap, with corresponding sizes.
[23] "facebook-page"
[25] "facebook-oauth"
[27] "facebook-access-token"
"fbml"
"facebook-social-plugins"
"facebook-wall"
> V(gr)$name[membership(imc)==18]
[1] "r"
"plot"
[4] "statistics"
"latex"
[7] "gnuplot"
"time-series"
[10] "visualization"
"aggregate"
[13] "data.table"
"bar-chart"
[16] "regression"
"subset"
[19] "legend"
"knitr"
[22] "apply"
"rstudio"
[25] "heatmap"
"correlation"
[28] "reshape"
"finance"
[31] "labels"
"lattice"
> V(gr)$name[membership(imc)==19]
[1] "google-maps"
"google-maps-api-3"
[4] "maps"
"coordinates"
[7] "openlayers"
"gis"
[10] "kml"
"polygon"
[13] "distance"
"geospatial"
[16] "openstreetmap"
"postgis"
[19] "google-maps-api-2"
"spatial"
[22] "google-earth"
"infowindow"
"ggplot2"
"data.frame"
"markdown"
"histogram"
"data-visualization"
"plyr"
"stata"
"xts"
"shiny"
"igraph"
"geolocation"
"google-maps-markers"
"geocoding"
"latitude-longitude"
"google-places-api"
"google-fusion-tables"
"arcgis"
Community #18 could very easily be labeled as “R”: indeed, plyr, ggplot2, lattice,
xts, igraph, reshape, and data.table are R packages; rstudio is the dominant R
IDE, while statistics, regression, visualization, and correlation are per-
Empirical Analysis of the Stack Overflow Tags Network
281
haps the most typical tasks for which R is used; latex could be a possible outlier, but we
suspect that its presence is due to the increasing use of the LaTex output option, as facilitated by the knitr and shiny packages in a markdown framework; a similar argument
could be brought forward also for stata (questions regarding porting of Stata tasks from
and to R); apply is an R family of functions...
Similarly, it is rather straightforward to assign the labels “Facebook” and “GIS” to the
communities # 17 and #19, respectively.
How about discovering “new” knowledge from this community structure? We include an
illustrative case, as it occurred to us while writing. Community #33 clearly has to do with
install/setup tasks in a Windows environment:
> V(gr)$name[membership(imc)==33]
[1] "wiz"
"installation"
[4] "windows-installer" "setup"
[7] "msi"
"installshield"
[10] "uninstall"
"wix3.5"
"installer"
"inno-setup"
"nsis"
"custom-action"
Having never heard of wiz, we Googled it; it proved to be “The most powerful set of
tools available to create your Windows installation experience” (see wixtoolset.org).
The list can go on; communities #28 and #35 are devoted to “big data” and “machine
learning” respectively, with virtually no outliers at all:
> V(gr)$name[membership(imc)==28]
[1] "hadoop"
"mapreduce" "cassandra"
[6] "apache-pig" "hdfs"
"bigdata"
[11] "cloudera"
"zookeeper"
"hive"
"thrift"
"hbase"
"piglatin"
> V(gr)$name[membership(imc)==35]
[1] "machine-learning"
"nlp"
[3] "artificial-intelligence" "neural-network"
[5] "classification"
"cluster-analysis"
[7] "data-mining"
"weka"
[9] "scikit-learn"
"svm"
[11] "mahout"
"libsvm"
[13] "k-means"
We encourage the reader to google possible unknown terms (mahout, perhaps?), in
order to confirm that they are indeed relevant with the other members of their respective
community.
What about our very small communities? Let us try communities #65 and #69:
> V(gr)$name[membership(imc)==65]
[1] "sip"
"asterisk" "voip"
"skype"
> V(gr)$name[membership(imc)==69]
[1] "neo4j" "cypher"
In community #65, sip stands for the Session Initiation Protocol (SIP) in voip networks (of which skype is the most popular example), while asterisk is “an open-source
telephony switching and private branch exchange service” (see Asterisk.org). Regarding community #69, neo4j is perhaps the most widely used graph database, and cypher is its
dedicated query language [21].
As expected, the situation is not so clear-cut with the larger communities; nevertheless,
even in those cases, there is certainly some nontrivial coherence in the communities —
for example, there are different and distinct communities for iOS (#6) and Android (#7);
282
Text Mining and Visualization: Case Studies Using Open-Source Tools
almost each and every one of them are associated with a major programming language or
framework. We can easily take a compact glimpse at the first 4 node tags of each of the
first 10 communities with the following for loop:
for (i in 1:10) {
print(V(gr)$name[membership(imc)==i][1:4])
}
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
"c++" "c" "arrays" "linux"
"javascript" "jquery" "html" "css"
"php" "mysql" "sql" "sql-server"
"c#" "asp.net" ".net" "asp.net-mvc"
"java" "eclipse" "spring" "swing"
"ios" "iphone" "objective-c" "xcode"
"android" "sqlite" "android-layout" "listview"
"python" "django" "list" "google-app-engine"
"ruby-on-rails" "ruby" "ruby-on-rails-3" "activerecord"
"git" "svn" "version-control" "github"
We think the reader will agree that a certain coherence is already visible from the above
list. And as we gradually move towards smaller communities, the picture becomes even
more coherent. Here is a similar glimpse of the communities #11-20:
for (i in 11:20) {
print(V(gr)$name[membership(imc)==i][1:4])
}
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
[1]
"xml" "parsing" "xslt" "xpath"
"excel" "vba" "ms-access" "excel-vba"
"actionscript-3" "flash" "flex" "actionscript"
"opengl" "graphics" "opengl-es" "3d"
"node.js" "mongodb" "express" "websocket"
"unit-testing" "testing" "selenium" "automation"
"facebook" "facebook-graph-api" "comments"
"facebook-like"
"r" "plot" "ggplot2" "statistics"
"google-maps" "google-maps-api-3" "geolocation" "maps"
"unicode" "encoding" "utf-8" "character-encoding"
The specific setting of the random seed has no effect on the qualitative aspects of the
result: the reader is encouraged to perform the previous experiment with several different
random seeds, to confirm that the results are indeed robust and stable, although the exact
number of the uncovered communities may differ slightly.
12.7
Visualization
It is a common truth between graph researchers and practitioners that straightforward
visualizations for graphs with more than 100 nodes are of very little use, and when we move
to even a few thousand nodes, as with our reduced graph, visualization ceases to have any
meaningful value [11]. Nevertheless, there are still useful and meaningful ways to visually
explore such graphs. Two common approaches are [11]:
Empirical Analysis of the Stack Overflow Tags Network
283
• To “coarsen” the graph, by merging several nodes together, possibly exploiting the
results of an already existing graph partitioning (macro level)
• To highlight the structure local to one or more given nodes, resulting in the so called
egocentric visualizations, commonly used in social networks analysis (micro level)
We are now going to demonstrate both these approaches with our graph.
Visualizing the Communities Graph
For the first approach, we are going to further exploit here the partitioning into communities provided by the Infomap algorithm from the previous Section, effectively resulting
in a dimensionality reduction for our graph, in order to produce a useful visualization. As
we will see, igraph provides several convenient functions for such purposes. We demonstrate first the use of the contract.vertices() function, which will merge the nodes
according to the community to which they belong:
> # first, add "community" attribute to each node
> gr <- set.vertex.attribute(gr, "community",
+
value = imc$membership)
> grc <- contract.vertices(gr, V(gr)$community)
> grc <- remove.vertex.attribute(grc, "name")
> for (i in 1:length(V(grc))) {
+
grc <- set.vertex.attribute(grc, "name", index=i,
+
V(gr)$name[membership(imc)==i][1])
+ }
> grc <- set.graph.attribute(grc, "name",
+
value = "Contracted (communities) graph")
> summary(grc)
IGRAPH UNW- 80 1033957 -- Contracted (communities) graph
attr: name (g/c), name (v/c), weight (e/n)
The for loop aims to give to each node in our contracted (communities) graph the name
of the first member of the respective community (usually the community “label”), which
will be subsequently used for the graph visualization. From the summary, we can see that
we now have only 80 nodes (recall from the previous Section that this is the number of the
communities uncovered by the Infomap algorithm), but we still carry all the 1 million edges
from the gr graph. That is because, as its name may imply, the contract.vertices()
function does not affect the edges of the graph. We can further simplify the contracted
graph by merging also the edges, summing up the corresponding weights, utilizing the
simplify() function. This is an important point for visualizing, as one can discover by
trying to plot the graph as it is so far, with more than 1 million edges.
> grc <- simplify(grc, edge.attr.comb = "sum") # sum edge weights
> grc <- set.graph.attribute(grc, "name",
+
value = "Contracted & simplified (communities) graph")
> summary(grc)
IGRAPH UNW- 80 2391 -- Contracted & simplified (communities) graph
attr: name (g/c), name (v/c), weight (e/n)
Thus, we have ended up with a rather simple graph of 80 vertices and only 2391 edges,
which should not be hard to visualize. We invoke the tkplot() command (the R package
tcltk needs to be installed, but once installed igraph will load it automatically). The
results are shown in Figure 12.3 (for best results, maximize the screen and choose some
layout other than “Random” or “Circle” from the figure menu).
284
Text Mining and Visualization: Case Studies Using Open-Source Tools
> tkplot(grc)
Loading required package: tcltk
[1] 1
Although our communities graph is still very dense, we can visually distinguish the
central nodes from the peripheral ones, and the visualization is in line with what we might
expect: nodes representing highly used tools, environments, and tasks are positioned in
the central bulk, while nodes representing less widely used tools are pushed towards the
graph periphery. We notice that the visualization provided by the tkplot() function is
interactive, for example, one can move and highlight nodes and edges or change their display
properties etc.
We now turn to the second approach mentioned in the beginning of this Section, i.e., the
so-called egocentric visualizations focused on individual nodes.
Egocentric Visualizations
We can use egocentric visualizations in order to focus closer on our uncovered communities, keeping in mind the informal visualization rule of thumb mentioned before, i.e., that
we should try to keep the number of our visualized nodes under about 100. The following
code produces the subgraph of the “R” community (#18) discussed already:
> k.name <- "r" # the name of the node of which the community we want to
examine
> k.community <- imc$membership[which(V(gr)$name==k.name)]
> k.community.graph <- induced.subgraph(gr, (V(gr)$community==k.community))
> k.community.graph$name <- paste(k.name, "community graph")
> k.community.graph
IGRAPH UNW- 32 311 -- r community graph
+ attr: name (g/c), name (v/c), community (v/n), weight (e/n)
As expected, we end up with a 32-node graph, which should not be hard to visualize.
Although not clearly documented in the igraph package manual, it turns out that we can
indeed use an edge.width argument in the tkplot() function, in order to visualize the
graph edges proportionally to their weight. Experimenting with the corresponding coefficient
(invoking the edge weight as-is produces a graph completely “shadowed” by the edges), we
get:
> tkplot(k.community.graph, edge.width=0.05*E(k.community.graph)$weight)
with the results shown in Figure 12.4.
From Figure 12.4, we can immediately conclude that the vast majority of R-related
questions in our data have to do with plotting (and the ggplot2 package in particular),
as well as with the specificities of the data.frame structure. Also of notice is the rather
strong presence of the relatively new data.table package.
We can go a step further, and remove the r node itself from the plot; that way we expect
the edge width visualization to be less dominated by the presence of the “central” node (r
in our case), and possibly uncover finer details regarding the structure of the particular
community. To do that, we only need to modify the induced.subgraph() line in the
previous code as follows:
> k.community.graph <- induced.subgraph(gr,
+ (V(gr)$community==k.community & V(gr)$name != k.name))
> tkplot(k.community.graph, edge.width=0.05*E(k.community.graph)$weight)
Empirical Analysis of the Stack Overflow Tags Network
285
The results are shown in Figure 12.5. We can confirm, for example, that the presence
of latex in the “R” community is indeed not spurious, as the subject node is strongly
connected with both knitr and markdown, although our initial speculation for its relation
also to shiny turns out to be incorrect.
We can easily extend the above rationale in order to examine closer the relationship
between two or more communities, as long as we limit the investigation to relatively small
communities (remember our 100-node max rule of thumb for visualizations). Say we would
like to see how the “Big Data” and “Machine Learning” communities are connected, excluding these two terms themselves:
>
>
>
+
>
k.names <- c("bigdata", "machine-learning")
k.communities <- imc$membership[which(V(gr)$name %in% k.names)]
k.communities.graph <- induced.subgraph(gr,
(V(gr)$community %in% k.communities & !(V(gr)$name %in% k.names)))
tkplot(k.communities.graph, edge.width=0.05*E(k.communities.graph)$
weight)
The results in Figure 12.6 show that, apart from the mildly strong connections of
mahout with both cluster-analysis and k-means, the two communities are practically strangers.
12.8
Discussion
We have provided a brief demonstration of how graph and network approaches and tools
can be applied for the analysis of semi-structured text in the form of tags. In summary, the
steps we have undertaken in the present chapter are as follows:
• Starting from a raw XML file of about 26 GB in size, we parsed the data and built a
graph structure of about 2.7 million edges for the co-occurrences of all 36,942 tags.
• We made an informed reduction of our data, in order to end up with a graph that
can be feasibly analyzed using only commodity hardware; in doing so, we retained
90% of our 21 million tag occurrences while ending up with a much more manageable
reduced graph of 3711 nodes and about 1 million edges.
• Despite the fact that the relevant metrics (diameter = 3, average path length =
1.85) suggested a very dense and tightly packed structure for our reduced graph, we
successfully employed the state of the art Infomap algorithm for community detection,
in order to uncover 80 distinct communities that, upon close inspection, look indeed
meaningful and coherent.
• Utilizing the detected communities in what stands effectively as a dimensionality
reduction operation on our reduced graph, we managed to end up with a coarsened
graph of only 80 nodes and about 2,500 edges, which could subsequently be visualized
and used for further exploratory data analysis.
• Finally, we exploited further the uncovered community structure of our graph, in order
to demonstrate how further micro-analysis in the neighborhood of selected nodes is
possible and potentially insightful.
There are certainly different choices than can be made in several key points of our analysis;
286
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 12.3: Visualization of our communities graph.
Empirical Analysis of the Stack Overflow Tags Network
FIGURE 12.4: Visualization of the “R” community.
287
288
Text Mining and Visualization: Case Studies Using Open-Source Tools
FIGURE 12.5: The “R” community, with the r node itself removed.
Empirical Analysis of the Stack Overflow Tags Network
289
FIGURE 12.6: The “Big Data” and “Machine Learning” communities (excluding these
terms themselves).
290
Text Mining and Visualization: Case Studies Using Open-Source Tools
for example, we could have pruned our initial graph not only according to the frequency of
the corresponding tags, but also according to the edge weights (which are also based on a
frequency measure). That way, we would have ended with a more clustered and less densely
connected graph, with a higher enough diameter and average path length. We chose not to
do so mainly in order to fully demonstrate the capabilities of the Infomap algorithm even
in such a dense graph, since the community detection is definitely easier for a graph that
possesses already some clustering on its own. With the step-by-step instructions and code
provided, the reader can easily explore this direction by herself.
We hope that we have presented a convincing case regarding the possible merits of the
graph representation, whenever appropriate, and of the handy tools provided by the R
language and the igraph library. Given the increasing presence of such representations in
modern data analysis, we certainly consider that a basic familiarity with the relevant tools
and concepts is a must-have skill for the practitioner.
12.9
Appendix: Data Acquisition & Parsing
Here we describe how to obtain the original data and how to parse them, in order to end
up with the two R workspace files, tag freq.RData and adj matrix.RData, which are
necessary for the analysis described in the chapter’s main body.
Data Acquisition
The whole Stack Exchange dump is available from the Internet Archive6 in 7z format.
The files necessary for our analysis are:
• stackoverflow.com-Posts.7z (about 5 GB, which contains the 26 GB uncompressed file Posts.xml for Stack Overflow)
• stackoverflow.com-Tags.7z (contains the Tags.xml file, with the tags and
their frequencies)
• readme.txt (with various information useful for parsing the data)
Parsing the Tag List & Frequencies
Nowadays, R offers extensive functionality for working with XML files [22], including
the XML package. After a quick inspection of the XML structure in Tags.xml, it is almost
trivial to parse the file and get the individual tag names with their corresponding frequencies
(notice that the Tags.xml file needs to be in the current R working directory, and the XML
package needs to be installed):
>
>
>
>
>
library(XML)
f <- ’Tags.xml’
doc <- xmlParse(f)
tag <- xpathSApply(doc, ’//row’, xpathSApply, ’@TagName’)
freq <- as.integer(xpathSApply(doc, ’//row’, xpathSApply, ’@Count’))
Subsequently, we just combine the variables tag and freq in a data frame, we sort it
in decreasing frequency order, and we save the resulting data frame in an R workspace file:
6 https://archive.org/details/stackexchange.
Empirical Analysis of the Stack Overflow Tags Network
291
> tag_freq <- data.frame(tag, freq,
+ stringsAsFactors=FALSE, row.names = NULL)
> # sort in descending frequency:
> tag_freq <- tag_freq[order(tag_freq$freq, decreasing=TRUE),]
> # save in R workspace file:
> save(tag_freq, file=’tag_freq.RData’)
File tag freq.RData should now be in R’s current working directory.
Parsing the Tag Co-occurences
In order to parse the tag co-occurrences present in the file Posts.xml, we need to
download and install BaseX7 , a free open-source XML database. Installation is straightforward, and so is the creation of a new database from Posts.xml (although it can take a
while); the resulting database is about 19 GB in size.
Posts.xml contains all the information about posts, including the answers. From
the provided readme file, we can see that PostTypeID=1 refers to questions, while
PostTypeID=2 refers to answers. Since tags are included only in the questions, we are
not interested in the latter. After some experimenting in order to arrive to the desired output format (one line of tags per post, with proper separation), we end up with the following
simple query (file get tags.xqy):
for $x in //row[@PostTypeId="1"]
return ($x/@Tags/string(),’&#xa;’)
Pasting the above code to the BaseX query window (double-clicking the provided script
file also opens the query to a new instance of the BaseX GUI) and running it, we get a sample
of the desired output displayed in the relevant window of the BaseX GUI. Although the
report window indicates that only 250,000 hits are returned, saving the output results indeed
to a text file with the tag co-occurrences of all 7,214,697 posts included in Posts.xml.
Building the Adjacency Matrix
Once we have the tag co-occurrences in a text file, we can proceed to build the graph
adjacency matrix. It is not exactly a straightforward job, and we will not describe it here in
detail - interested readers can always consult the provided script parse posts.R. Here we
highlight some challenges that needed to be addressed, using them also as a demonstration
of good programming approaches in R.
In theory, with enough memory available (at least 8 GB), it is possible to load the
whole post tags.txt file produced in the previous step, and parse it to directly build the
(sparse) graph adjacency matrix G; unfortunately, it turns out that in practice one encounters the following problem: the sparse matrix G, as it gets populated, it gets increasingly
“heavier” for incremental updates, leading to unacceptably high execution times, even in
machines with powerful processors and lots of memory.
The solution we applied to overcome this issue was twofold: first, we read the
post tags.txt file via a file connection, taking in only a bunch of rows at a time in
an external while loop; second, in each loop iteration, we initialized an all-zeros temporary co-location matrix G temp, which was the one being incrementally updated, and was
added to the adjacency matrix G only at the end of each loop iteration. This approach
speeded up things significantly, since G temp never grew up to sizes that slowed down its
7 http://basex.org/.
292
Text Mining and Visualization: Case Studies Using Open-Source Tools
incremental updates, while the matrix addition of G and G temp proved to be of practically
no computational cost.
In our setting, it turned out that a value of 1000 rows per iteration was optimal, leading
to 7215 while loop iterations; users with slower hard discs may wish to experiment with
these settings, since there is a disk read operation in each loop iteration. The resulting
solution also has the merit that it is usable to machines with low memory, since the number
of rows read at each loop iteration can be as low as desired if the value of 1000 we used
turns out to be prohibitive with the memory available.
Been forced to use an external while loop for the reasons just mentioned, we aimed to
completely avoid any for loops in the parsing procedure, since they are notoriously slow in
R (see “The Dreaded for Loop” in [23]). We thus made extensive use of the lapply family
of R functions. Novice R users usually find lapply and its relatives rather hard to master
initially, but once comprehended, they provide an excellent set of tools for writing efficient
R code.
Utilization of the lapply functions is also important for one more reason, leading
naturally the discussion towards the third ingredient we used for speeding up our code:
parallelization, which is rather trivial nowadays in R, even in Windows systems (notice
that the parallel package is now included by default in any R distribution). Indeed,
incorporating parallelization in a multicore machine is almost as easy as the following code
snippet indicates:
library(parallel)
load("tag_freq.RData")
# initialize parallelization:
num_cores <- 4
cl <- makeCluster(num_cores)
clusterExport(cl, varlist="tag_freq")
Now, if we have already put the effort to use lapply functions in our code instead of
for loops, converting them in order to effectively use the existing parallelizatiuon backend
is trivial: just compare the single-threaded version shown below, where we build a temporary
variable by applying our function getTags to the raw file input doc
temp <- apply(as.matrix(doc), 1, getTags)
to its parallel version
temp <- parApply(cl, as.matrix(doc), 1, getTags)
On the contrary, and despite the functionality available in the foreach package, parallelizing for loops is by no means that easy and straightforward.
Bibliography
[1] A.-L. Barabási. Network Science.
networksciencebook/.
2014.
http://barabasilab.neu.edu/
[2] D.J. Watts and S.H. Strogatz. Collective dynamics of ”small-world” networks. Nature,
393(6684):440–442, 1998.
Empirical Analysis of the Stack Overflow Tags Network
293
[3] A.-L. Barabási and Réka A. Emergence of scaling in random networks. Science,
286(5439):509–512, 1999.
[4] Easley D. and J. Kleinberg. Networks, Crowds, and Markets: Reasoning About a Highly
Connected World. Cambridge University Press, New York, NY, 2010.
[5] M.O. Jackson. Social and Economic Networks. Princeton University Press, Princeton,
NJ, 2008.
[6] M. Newman. Networks: An Introduction. Oxford University Press, Inc., New York,
NY, 2010.
[7] E.D. Kolaczyk. Statistical Analysis of Network Data: Methods and Models. Springer,
2009.
[8] D.J. Watts. Six Degrees: The Science of a Connected Age. W.W. Norton & Company,
2004.
[9] R Core Team. R: A Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna, Austria, 2015.
[10] G. Csárdi and T. Nepusz. The igraph software package for complex network research.
InterJournal, Complex Systems:1695, 2006.
[11] E.D. Kolaczyk and G. Csárdi. Statistical Analysis of Network Data with R. Springer,
2014.
[12] D. Bates and M. Maechler. Matrix: Sparse and Dense Matrix Classes and Methods,
2014. R package version 1.1-4.
[13] A. Barrat, M. Barthélemy, R. Pastor-Satorras, and A. Vespignani. The architecture of
complex weighted networks. Proceedings of the National Academy of Sciences of the
USA, 101(11):3747–3752, 2004.
[14] L.C. Freeman. A set of measures of centrality based on betweenness. Sociometry,
40(1):3–415, 1977.
[15] S. Fortunato. Community detection in graphs. Physics Reports, 486(3-5):75–174, 2010.
[16] M. Girvan and M.E.J. Newman. Community structure in social and biological networks.
Proceedings of the National Academy of Sciences of the USA, 99(12):7821–7826, 2002.
[17] A. Clauset, M.E.J. Newman, and C. Moore. Finding community structure in very large
networks. Physical Review E, 70:066111, 2004.
[18] M.E.J. Newman and M. Girvan. Finding and evaluating community structure in networks. Physical Review E, 69(2):026113, 2004.
[19] M. Rosvall and C.T. Bergstrom. Maps of random walks on complex networks reveal
community structure. Proceedings of the National Academy of Sciences of the USA,
105(4):1118–1123, 2008.
[20] M. Rosvall, D. Axelsson, and C.T. Bergstrom. The map equation. The European
Physical Journal Special Topics, 178(1):13–23, 2009.
[21] I. Robinson, J. Webber, and E. Eifrem. Graph Databases. O’Reilly Media, Inc., 2013.
294
Text Mining and Visualization: Case Studies Using Open-Source Tools
[22] D. Nolan and D.T. Lang. XML and Web Technologies for Data Sciences with R.
Springer, 2014.
[23] Norman Matloff. The Art of R Programming. No Starch Press, 2011.
This page intentionally left blank
Computer Science & Engineering
The contributors—all highly experienced with text mining and open-source soft ware—explain how text data are gathered and processed from a wide variety of
sources, including books, server access logs, websites, social media sites, and
message boards. Each chapter presents a case study that you can follow as part
of a step-by-step, reproducible example. You can also easily apply and extend
the techniques to other problems. All the examples are available on a supplemen tary website.
The book shows you how to exploit your text data, offering successful applica tion examples and blueprints for you to tackle your text mining tasks and benefit
from open and freely available tools. It gets you up to date on the latest and most
powerful tools, the data mining process, and specific text mining activities.
T ex M ining and
V isualiza ion
t
Text Mining and Visualization: Case Studies Using Open-Source Tools pro vides an introduction to text mining using some of the most popular and powerful
open-source tools: KNIME, RapidMiner, Weka, R, and Python.
Chapman & Hall/CRC
Data Mining and Knowledge Discovery Series
t
“The timing of this book could not be better. It focuses on text mining, text be ing one of the data sources still to be truly harvested, and on open-source tools
for the analysis and visualization of textual data. … Markus and Andrew have
done an outstanding job bringing together this volume of both introductory and
advanced material about text mining using modern open-source technology in a
highly accessible way.”
—From the Foreword by Professor Dr. Michael Berthold, University of Konstanz,
Germany
T ex t M ining and V isualizat ion
Chapman & Hall/CRC
Data Mining and Knowledge Discovery Series
Case Studies Using
Open-Source Tools
Hofmann
Chisholm
Edited by
Markus Hofmann
Andrew Chisholm
K23176
w w w. c rc p r e s s . c o m
K23176_cover.indd 1
11/3/15 9:18 AM
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertisement