Opinion Mining, Analysis and its Challenges

Opinion Mining, Analysis and its Challenges
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
Opinion Mining, Analysis and its Challenges
Nidhi R. Sharma , Prof. Vidya D. Chitre
Computer Department, University of Mumbai, Computer Department, University of Mumbai,
[email protected]
[email protected]
BVCOE, Navi Mumbai
Abstract--Any user , buyer or customer rely on the Web for their
opinions on various products and services which they have used,
it is very important to develop methods to automatically classify
and evaluate them. The task of classifying and analyzing such
collective data together is known as customer feedback or
review data, and is called as opinion mining.
Opinion Mining is a very challenging and promising discipline
which is defined as an intersection of information retrieval and
computational linguistic techniques to deal with the opinions
expressed in a document. The main aim at solving the problems
related to opinions about products, reviews ranking in movies,
Politian in newsgroup posts, review sites etc. In this paper we
are about to cover the source of data from where we take , its
classification, evaluation process and then grouping techniques,
tools used, and future challenges in opinion mining. Opinion
mining consists of various stages such as extraction of data from
various sources, text classification, grouping together and then
evaluating it to positive or negative or true or false value. On
the basis of our survey and analysis of the techniques, we
provide an overall picture of what is involved in developing a
software system for opinion mining.
related decisions. Multiple research fields like statistics,
machine learning, artificial intelligence and visualization are
involved to develop such automated systems [2-7].
A number of proficient ways are existing [4] to store the huge
volumes of data, computational techniques and models are
required to extract the hidden patterns and knowledge. These
techniques and tools are used to transform the data into
useful information, to make market analysis, fraud detection
and find the customer intentions etc. Such computational
tools and techniques are the subject of Knowledge Discovery
in Database and Data Mining [4-5].Text mining is an
interdisciplinary method used in different fields like machine
learning, information retrieval, statistics, computational
linguistic and data mining to form mining algorithms. Some
researchers defined text mining as tool to discover the new
knowledge from huge volume of natural language text using
computational algorithms. Web mining is a sub discipline of
text mining used to mine the semi structured web data in
form of web content mining, web usage mining and wed
structure mining.
Keywords-- Data mining, web mining, opinion mining,
sentiment classification, text classification, evaluations.
1. Introduction:
The World Wide Web is growing at an alarming rate not only
in size but also in the types of services and contents provided.
Each and every users are participating more actively and are
generating vast amount of new data. These new Web contents
include customer reviews and blogs that express opinions on
products and services – which are collectively referred to as
customer feedback data on the Web. As customer feedback
on the Web influences other customer’s decisions, these
feedbacks have become an important source of information
for businesses to take into account when developing
marketing and product development plans. This era is of
automated systems [1] and digital information every field of
life is evolving rapidly and generating data. As a result huge
amount of data produce in the field of science, engineering,
medical, marketing, finance etc [2]. Automated systems are
needed to automate analysis, summarization, and
classification of data. It also helps at enterprise level to take
Nidhi , Vidya
Fig. 1. Data Mining Hierarchical Model
This paper is organized as follows: section 2 covers opinion
mining. Section 3 is about the dataset source. Section 4 is
about the levels of sentiment classification. Section 5
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
describes text classification Section 6 is all about the
grouping feature Section 7 describes evaluation process and
Section 8 describes various recent tools used to do this.
2. Opinion Mining (O.M)
Opinion Mining is a promising discipline which is defined as
combination of information retrieval and computational
linguistic techniques deals with the opinions expressed in a
document. The field major goals is solving the problems
related to opinions about products, politics in newsgroup
posts, review sites, etc. There are different techniques for
summarizing customer reviews like Data Mining,
Information Retrieval, Text Classification and Text
Summarization [2], before World Wide Web users asked the
opinions of his family and friends to purchase the product . In
the very same way when any organizations need to take the
decision about their products they had to conduct various
surveys to the focused groups or they had to hire the external
consultants to do so [4]. Web 2.0 [7], ease the customers to
take decision to purchase the product by reviewing the posted
comments. Customers can post reviews on web communities,
discussion forums, twitters, blogs, product’s web site these
comments are called user generated contents. Web2.0 is
playing a vital role in data extracting source in opinion
mining. It facilitates users to know about the product from
other customer’s reviews who have already used it instead of
asking friends and families. Companies, instead of
conducting surveys and hiring the external consultants to
know about the clients opinions, extract opinionated text
from product web site [8]. An automated opinion
summarization model is needed to complete these tasks.
Opinion Mining or Sentiment Analysis is the area to extract
the opinionated text datasets and summarize in
understandable form for end user [8]. Opinion mining is used
to extract the positive, negative or neutral opinion summary
from unstructured data. It involves subjectivity in text and
computational management of opinion. It is the subdiscipline of web content mining, which involves Natural
Language Processing and opinion extraction task to find out
the polarity of any product consumers feedback [4]. Figure 2
describes the object model of Opinion Mining.
Fig. 2. Opinion Mining Model
In the above diagram there are five components i.e a
customer giving the broader reviews from various sources ,
it is the sentiment, views or judgment about any object based
on knowledge or experience , then is the Opinion Holder
which is the person, organization that expresses its views or
sentiments about any object and the Object which is an
entity (person, topic, product or organization) about which
the opinion expressed.
3. Data Source
People and companies across disciplines exploit the rich and
unique source of data for varied purposes. The major decisive
factor for the improvement of the quality services rendered
and enrichment of deliverables are the user consumers
opinions. Review sites, blogs and micro blogs provide a good
understanding of the reception level of products and services.
3.1 Review Sites
Opinions are the major and actual data or more precise a
decision for any user in making a purchase. The user
generated reviews for products and services are mainly
available on internet. The sentiment classification uses
reviewer’s data are gathered and composed from the websites
likewww.gsmarena.com (mobile reviews),www.amazon.com
(product reviews), www.CNETdownload.com (product
reviews), which hosts millions of product reviews by
consumers. [9]
3.2 Blogs
The name associated to universe of all the blog sites is called
blogosphere. People write about the topics they want to share
with others on a blog. Blogging is a happening thing because
Nidhi , Vidya
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
of its ease and simplicity of creating blog posts, its free form
and unedited nature. We find a large number of posts on
virtually every topic of interest on blogosphere. Sources of
opinion in many of the studies related to sentiment analysis,
blogs are used.[10]
3.3. Micro-blogging
A very accepted communication tool among Internet users is
micro-blogging. We use this as one of the data source formed
as a dataset of collected messages from Twitter. Twitter
contains a very large number of very short messages created
by the users, consumers of this micro blogging platform.
Millions of messages appear daily in well-liked web-sites for
micro-blogging such as Twitter, Tumblr, Facebook. Twitter
messages sometimes express opinions which are used as data
source for classifying sentiment. [11]
4. Sentiment Classification
4.1 Document level:
Document level sentiment classification is based on the
sentiments executed on the overall sentiments expressed by
authors. Documents classified according to the sentiments
instead of topic. It is very useful in summarizing the whole
document as positive or negative polarity about any object
(camera, fridge, mobile, car, movie, and politician). In [12]
authors proposed a new approach “classification of opinion
documents by a vote system” based on combining text
representations using key-words related to bigrams.
Sentiment Classification Using Phrase Patterns in used
Special tags opinion words. System constructed some phrase
patterns and compute sentiment orientation using
unsupervised learning algorithm. Proposed system achieved
86% accuracy. Investigated perspective from which a
document was written. They build Naïve Bayes based model
and test on Israeli-Palestinian conflict. Their corpus consists
of articles published on the bitter lemons website. They used
NB-B (full Bayesian inference) and NB-M (Maximum a
4.2 Sentence level:
Sentence level sentiment classification models is used for the
extraction of the sentences contained in the opinionated
terms, opinion holder and opinionated object. It is one level
deep to document level and just concerns to the opinionated
Nidhi , Vidya
words but not the features. Total number of positive and
negative words are counted from the extracted and classified
sentences and if positive words are maximum then opinion
about object is positive and if the negative words are more
than opinion object is negative otherwise the opinion object
will be neutral. To mine the customer reviews on a product
proposed unsupervised algorithm is used and in this the
algorithm find frequent features using Apriori algorithm.
Chinese WordNet set classify opinion words in clauses (pos,
neg or neutral) to summarize the comments. Sentence level
opinion mining uses subjective and polarity (orientation) to
find strength of opinions at the clause level. [13], all these are
a notable work in this regard. To find the strength of opinions
a new idea of syntactic clues is used. They use a wide range
of features to find the strength of opinions. The system is
about to provide tools and support for information analysts in
government, commercial, and political domains, who want to
be able to automatically track attitudes and feelings in the
news and on-line forums. Opinion Analysis based on Lexical
Clues and their Expansion to improve combination of rulebased algorithms and machine learning techniques. [14]
proposed semi-supervised learning method based on highly
precise seed rules. Subjectivity discovered at the sentence
level. Polarity of the sentence defined as Positive, Negative
or Neutral as well as opinion holders identified. The
experimental results demonstrate that system achieved 45%
Accuracy to extract opinionated sentences and 35% Accuracy
to identify opinion holders.
4.3 Feature based level:
In customer reviews document, reviewer express positive,
negative or both sentiments about the object and attributes.
Document level and sentence level classification does not tell
the likes and dislikes of consumer about particular attributes
of object [16]. When consumer comment on object (product,
person, and topic, organization) he comment on the features
of object[15].For example, if users commented on a Mobile
Phone they basically comment on Camera result, LCD size,
speaker, weight etc. On camera output 125 comments
express the positive opinions and 25 comments may be
negative. If a new customer is interested in camera quality of
mobile he can take decision easily to purchase the product or
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
not. To explore the detailed opinion on product or any topic,
a detailed opinion mining study is required that is called
feature based opinion mining [17]. Statistical Opinion
Analyzer (SOA) extract the polarity of online customer
reviews using Bayesian probability and frequency
distribution. The proposed system helps the new customer to
purchase the product and manufacturer to enhance the
product’s functionality. Reviews crawled, preprocess, tagged
(GO tagger) and insert in SOA to find the positive and
negative opinion probability and frequency distribution as
well. The proposed system originated the very promising
results. In a web based system SUMView crawled reviews
from Amazone.com, decompose into sentences and tagged to
find the nouns and noun phrases. Product features extracted
using Hu and Liu (2004) method and top five extracted
features were suggested to the users on the basis of
Here we describes the shortcomings of sentiment
classification at documents, sentence and feature level.
Document level classification
It does not give details of what people likes or
dislikes because writer comments only on the
specific aspects of product.
It is also not applicable on forums and blogs as they
contains only few opinionated sentences on features
of object.
It defines the polarity of document , but a positive
phrase does not indicates that the user likes
everything and similarly a negative phrase does not
indicate that the opinion holder dislikes everything.
Sentence level classification
 It is a fine-grained level of classification in which
polarity of the sentence can be given by three
categories as positive, negative and neutral.
 In this the identification features indicate whether
sentences are on-topic which is kind of co-reference
Feature level classification
 It is defined as product attributes or components.
Nidhi , Vidya
In this approach positive or negative opinion is
identified from the already extracted features.
It is a fine grained analysis model among all other
It is having a drawback that it could really cut very
badly if there used any grammatically incorrect
Text Classification
Now a day’s huge and vast structured and unstructured
volume of online text is available through different websites,
internet news feed, emails, cooperate databases and digital
library. The main problem is to classify text documents from
such massive databases. Using set of training labeled
examples statistical text learning algorithms can be trained to
approximately classify documents. The news articles and
web pages were automatically catalogued by these text
classification algorithms.
Naïve Bayes Classifier is a well known probabilistic
classifier which describes its application to text. In order to
incorporate unlabelled data, the foundation Naïve Bayes was
build. The task of learning of a generative model is to
estimate the parameters using labeled training data only. The
estimated parameters are used by the algorithm to classify
new documents by calculating which class the generated the
given document belongs to. Biological data is modeled using
probabilistic models, such as HMM (Hidden Markov Model)
or Bayesian networks which are efficient and robust
procedures for learning parameters from observations. There
are various sources for missing values such as in medical
diagnosis, missing data for certain tests or gene expression
clustering due to intentional omission of gene-to-cluster
assignments in the probabilistic model. Such error does not
occur in EM algorithm.[18,19]
Grouping Features
Regular expressions: Regular expressions are highly
specialized programming language, in which the rules are
specified for the set of possible strings that can be matched
and the set might contain English sentences, or e-mail
addresses etc. String processing tasks which are performed
using regular expressions becomes very complicated because
regular expressions language is relatively small and
restricted. This mechanism was motivated to avoid silly
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
errors by automated systems, particularly machine learning
models. For example, opening salutations, such as “Dear
Prabhat”, were falsely assigned with instructions by a
machine learning model in some runs of cross-validation
tests, possibly due to the frequent occurrences of person
names in instructions. With this mechanism, “dear
__NAME__”, a normalized form of “Dear Prabhat”, was
compared against all such normalized instances in the
training data, and false assignment of instructions could be
avoided after reviewing emotions assigned to the found
training instances, specifically, by confirming more than twothirds of the found training instances were not assigned with
instructions. [20]
Clustering: Clustering is the natural technique used to
discover hundreds of feature expressions from text for an
opinion mining application. Similarity measures used for
clustering are usually based on some form of distributional
similarity. There are two main kinds of similarity measures
those relying on pre-existing knowledge resources (e.g.,
thesaurus, and semantic networks) and those relying on
distributional properties of words in corpora.
First, a pre-processing pass could build a list of words and
phrases that appear frequently in the review of a particular
restaurant but are uncommon in the wider corpus. This
should find phrases like the name of a dish that many people
are talking about. Second, given the narrow domain of the
problem, it should also be possible to hand-build a list of
common ideas a reader might want to know about, like
service, food, and price. Extracting these combined, specific
features should lead to purpose-built vectors that form
clusters around relevant concepts. [21]
Purity: Purity measures the extent that a cluster
contains only data from one gold-partition. The
cluster purity is computed with Equation (c). The
total purity of the whole clustering (all clusters) is
computed with Equation (d) [22]
Similarly for the evaluation of sentiment classification using
regular expression is usually measured by precision and
recall. Precision is the fraction of relevant retrieved instances,
while recall is the fraction of retrieved relevant instances.
Therefore precision and recall are based on an understanding
and measure of relevance.
Where, TP - number of true positives
TN-number of true negatives
FP - number of false positives
FN - number of false negatives. [20]
Evaluation Measures
All since the problem of grouping feature expressions is a
clustering task, two common measures for evaluating
clustering are used the study, Entropy and Purity. Below, we
briefly describe entropy and purity. Given a data set DS, its
gold partition is G = {g1....,…,gj....,…gk}, where k is the
given number of clusters. The groups partition DS into k
disjoint subsets, DS1,…, DSi, …, DSk.
Entropy: For each resulting cluster, we can measure
its entropy using Equation (a), where P i(gi) is the
proportion of gi data points in DSi. The total entropy
of the whole clustering (which considers all clusters)
is calculated by Equation (b)
Nidhi , Vidya
A Red Opal is a tool that enables users to find products based
on features. The features from customer reviews are used for
scoring each product. Opinions on web are analysed and
compared using Opinion observer. The product opinions are
displayed feature by feature in graph format.
Automation of aggregation sites is done by Review Seer tool.
The extracted features are assigned score by Naïve Bayes
classifier as positive and negative review. The crawled pages
are not classified properly by this tool. Result is displayed in
the form of attribute and its score.
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
Product features are extracted in Web Fountain using
beginning definite Base Noun Phrase (bBNP) heuristic. The
sentiment lexicon and sentiment pattern database are used to
assign sentiments to feature. Sentiment extraction patterns
are defined in sentiment pattern database and polarity of
terms is defined in sentiment lexicon.
Challenges in Opinion Mining
1. Product reviews, comments and feedback could be in
different languages (English, Urdu, Arabic, french
etc), therefore to tackle each language according to
its orientation is a challenging task.
2. As noun words are considered as feature words but
Verbs and adjectives can also be used as feature
words which are difficult to identify.
3. If a customer-One comments on mobile phone, “the
voice quality is excellent” and customer-Two
comments, “sound quality of phone is very good”.
Both are talking about same feature but with
different wording. To group the synonym words is
also a challenging task.
6. Orientation of opinion words could be different
according to situation. For example “Camera size of
mobile phone is small”. Here adjective small used in
positive sense but if customer parallel said that “the
battery time is also small”. Here small represent
negative orientation to battery of phone. To identify
the polarity of same adjective words in different
situation is also a challenging task.
7. As the customer comment in free format, she can
use abbreviation, short words, and roman language
in reviews. For example u for you, cam for camera,
pic for picture, f9 for fine,b4, before, gud for good
etc. To deal with such type of language need a lot of
work to mine opinion.
8. Different people have different writing styles, same
sentence may contain positive as well as negative
opinion, so it is difficult to parse sentence as
positive or negative in case of sentence level opinion
mining .
9. In Bing Liu approach opinion always classified only
in two categories positive and negative but Neutral
opinion also expressed sometimes. Liu considers
Nidhi , Vidya
only adjective as opinion words but opinion can also
expressed as adverb, adjectives and verb. For
example “like” is a verb but also an opinion word.
His approach finds the implicit features because it
extracts the sentences contain at least one feature
word. So the features commented by customer
indirectly are ignored [17].
Lexicon based methods use for opinion mining has
not an effective method to deal with context
dependent words. For example the word “small” can
express the either positive or negative opinion on the
product features. For a mobile phone if customer
comments that “size of mobile phone is small” this
sentence does not show either size is positively
opinioned or negatively.
To finding of spam and fake reviews, mainly
through the identification of duplicates.
The comparison of qualitative with summary
reviews and the detection of outliers, and the
reputation of the reviewer.
The combination of opinion with behavior to
validate data and provide further analysis into the
data ahead of opinion expressed.
The continuous need for better usability and userfriendliness of the mining systems.
10. Conclusions
Opinion miming is an emerging field of data mining used to
extract the pearl knowledge from huge volume of customer
comments, feedback and reviews on any product or topic etc.
A lot of work has been conducted to mine opinions in form
of document, sentence and feature level sentiment analysis It
is examined that now opinion mining trend is moving to the
sentimental reviews of twitter data, comments used in
Facebook on pictures, videos or Facebook status. In future,
Opinion Mining can be carried out on a set of reviews and set
of discovered feature expressions extracted from reviews.
The state-of-art for current methods, useful for producing
better summary based on feature based opinions as positive,
negative or neutral is the Expectation Maximization
algorithm based on Naïve Bayesian is the most efficient
method. The efficiency of EM algorithm can be increased by
augmenting it, to reassign classes of the labelled set.
International Journal of Innovations & Advancement in Computer Science
ISSN 2347 – 8616
Volume 3, Issue 1
April 2014
The natural language text can be processed based on machine
learning toolkit called as OpenNLP library. The NLP tasks,
such as tokenization, part-of-speech (POS) tagging, named
entity extraction, parsing, chunking, sentence segmentation,
and co reference resolution are provided by Open NLP
library. The advanced text processing services are built using
these tasks. OpenNLP also includes perceptron and
maximum entropy based machine learning.After POS
tagging, opinion retrieval can be performed by extracting
product candidate feature, related opinion and producing
opinion feature pairs. The keywords extracted from Opinion
Retrieval Module can be used to perform similarity check
with the database dictionary. The similarity check can use
semi supervised learning.
T. Khushboo “Mining of Sentence Level Opinion Using Supervised
Term Weighted Approach of Naïve Bayesian Algorithm” ,Int. Journal.
Computer Technology & Applications, Vol 3 IJCTA | MAY-JUNE
N, Anwer and A, Rashid “Feature Based Opinion Mining of Online
Free Format Customer Reviews Using Frequency Distribution and
Bayesian Statistics” Networked Computing and Advanced Information
Management (NCM), 2010 Sixth International Conference on 16-18
Aug. 2010.
B. Seerat and F.Azam “Opinion Mining: Issues and Challenges (A
survey)” International Journal of Computer Applications (0975
– 8887) Volume 49– No.9, July 2012.
N. M. Shelke, S. Deshpande and V. Thakre “Survey of Techniques for
Opinion Mining” International Journal of Computer Applications (0975
– 8887) Volume 57– No.13, November 2012.
D. S. Deshpande “ A Survey on Web Data Mining Applications”
Emerging Trends in Computer Science and Information Technology 2012(ETCSIT2012) Proceedings published in International Journal of
Computer Applications® (IJCA).
R. Kohavi and F.Provost “A research on web datamining and its
application in electronic commerce
‘ computational Intelligence and software Engineering 2009. CiSE
2009. International Conference on Date of conference 11-13 Dec 2009.
Singh and Vivek Kumar, A clustering and opinion mining approach to
Nidhi , Vidya
socio-political analysis of the blogosphere, Computational Intelligence
and Computing Research (ICCIC), 2010 IEEE International
Alexander Pak and Patrick Paroubek, Twitter as a Corpus for
Sentiment Analysis and Opinion Mining,
Zhai Z, Liu B, Xu H, and Jia P, Grouping Product Features Using
Semi-supervised Learning with Soft-Constraints, in Proceedings of
COLING. 2010
Khairullah Khan and Baharum B. Baharudin, Identifying Product
Features from Customer Reviews using Lexical Concordance, Research
Journal of Applied Sciences Engineering and Technology, 2012
Boris Kraychev and Ivan Koychev, Computationally Effective
Algorithm for Information Extraction and Online Review Mining, 2010
ZhongchaoFei, Jian Liu, and Gengfeng Wu: “Sentiment Classification
Using Phrase Patterns”, Proceedings of the Fourth International
Conference on Computer and Information Technology in 2004.
Wilson, T., Wiebe, J. and Hwa, R, “Just how mad are you? Finding
strong and weak opinion clauses”, Proceeding of National Conference
on Artificial Intelligence in 2004.
Hiroshi, K., Tetsuya, N., and Hideo, W. “Deeper sentiment analysis
using machine translation technology”. In Proceedings of the 20th
international Conference on Computational Linguistics (Geneva,
Switzerland) in 2004.
M Fan, G WU “Opinion Summarization of Customer comments”
International conference on Applied Physics and Industrial Engineering
in 2012.
B. Liu “Sentiment Analysis and Opinion Mining”, April 22, 2012.
Kim, Y. and Myaeng, S., “Opinion Analysis based on Lexical Clues
and their Expansion”, Proceedings of NII Test Collection for
Information Retrieval in 2007.
Chuong B Do & Serafim Batzoglou, What is the expectation
maximization algorithm?, 2008 Nature Publishing Group
Kamal Nigam, Andrew Kachites Mccallum, Sebastian Thrun and Tom
Mitchell, Text Classification from Labeled and Unlabeled Documents
using EM, Machine Learning, 39, 103–134, 2000.2000 Kluwer
Academic Publishers. Printed in The Netherlands.
Sunghwan Sohn,1,* Manabu Torii,2,* Dingcheng Li,1 Kavishwar
Wagholikar,1 Stephen Wu,1 and Hongfang Liu1, A Hybrid Approach
to Sentiment Sentence Classification in Suicide Notes, Biomed Inform
Insights. 2012; 5(Suppl. 1): 43–50. Published online 2012 January 30.
doi: 10.4137/BII.S8961
YuanbinWu, Qi Zhang, Xuanjing Huang, LideWu, Phrase Dependency
Parsing for Opinion Mining, Proceedings of the 2009 Conference on
Empirical Methods in Natural Language Processing, pages 1533–1541,
Singapore, 6-7 August 2009. c 2009 ACL and AFNLP
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

Related manuals

Download PDF