Online Paper Review Analysis - The Science and Information (SAI

Online Paper Review Analysis - The Science and Information (SAI
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
Online Paper Review Analysis
Doaa Mohey El-Din
Hoda M.O. Mokhtar
Osama Ismael
Information Systems Department
Faculty of Computers and
Information, CU
Cairo, Egypt
Information Systems Department
Faculty of Computers and
Information, CU
Cairo, Egypt
Information Systems Department
Faculty of Computers and
Information, CU
Cairo, Egypt
Abstract—Sentiment analysis or opinion mining is used to
automate the detection of subjective information such as
opinions, attitudes, emotions, and feelings.
Hundreds of
thousands care about scientific research and take a long time to
select suitable papers for their research. Online reviews on
papers are the essential source to help them. The reviews save
reading time and save papers cost. This paper proposes a new
technique to analyze online reviews. It is called sentiment
analysis of online papers (SAOOP). SAOOP is a new technique
used for enhancing bag-of-words model, improving the accuracy
and performance. SAOOP is useful in increasing the
understanding rate of review's sentences through higher
language coverage cases. SAOOP introduces solutions for some
sentiment analysis challenges and uses them to achieve higher
accuracy. This paper also presents a measure of topic domain
attributes, which provides a ranking of total judging on each text
review for assessing and comparing results across different
sentiment techniques for a given text review. Finally, showing
the efficiency of the proposed approach by comparing the
proposed technique with two sentiment analysis techniques. The
comparison terms are based on measuring accuracy,
performance and understanding rate of sentences.
Keywords—Sentiment analysis; Opinion Mining; Reviews; Text
analysis; Bag of words; sentiment analysis challenges
I.
INTRODUCTION
World Wide Web (www) has become the most popular
communication platforms to the public reviews, opinions,
comments and sentiments about products, places, scientific
books or papers and to daily text reviews. The number of
active user bases and the size of their reviews created daily on
online websites are massive. There are 2.4 billion active
online users, who write and read online and Internet usage
around the world [1]. Scientific research domain has a big
world in journals and conferences, there are more than 4000
rated conferences and 5000 ranked journals [2]. Each one of
them has thousand number of papers such as ACM, Springer
and Science direct. Notably, a large fragment of WWW
researchers makes their content public, allowing researchers,
societies, universities, corporations to use and analyze data.
According to a new survey conducted by Dimensional
Research, April 2013: 90% of customer’s decisions depends
on Online Reviews [3]. According to 2013 Study [4]: 79% of
customer’s confidence is based on online personal
recommendation reviews. As the result, a large number of
studies and research have monitored the trending new research
increasing year by year. In this work, trying to achieve trusted
scientific reviews evaluation to be useful for researchers and
facilitate the selection of the suitable papers.
Recently, several websites encourage researchers to
express and exchange their views, suggestions and opinions
related to scientific papers. Sentiment analysis [5] depends on
two issues sentiment polarity and sentiment score. Sentiment
polarity [6] is a binary value either positive or negative. On
the other hand, sentiment score which relies on one of three
models [7]. Those models are Bag-of-words model (BOW)
[8], part of speech (POS) [9], and semantic relationships [10].
BOW [11] is the most popular for researchers and based on
the representation of word but BOW neglects language
grammar. POS [12] which is grammatically tagging especially
verbs, adjectives and adverbs [13]. For example, (The research
is not good.) declaring in (The/DT research/NN is/VBZ
not/RB good/JJ. /.). In the example DT refers to "Determiner",
NN refers to "Noun", singular or mass, VBZ refers to "Verb",
RB refers to "Adverb", and JJ refers to "Adjective". But a
semantic relationship method is the most complex method,
which is based on the relationship between concepts or
meanings for example antonym, synonym, homonym etc.
There is a research gap the sentiment analysis accuracy
because of sentiment evaluation drawbacks and sentiment
analysis challenges [14].The evaluation sentiment drawbacks
that Reflected in language coverage. This paper focuses on
understanding text reviews and introduces solutions for some
sentiment challenges. The sentiment analysis challenges
summarized in ten challenges [15]. They are spam and fake
reviews Detection, Limitation of classification filtering,
Asymmetry in availability of opinion mining software
Incorporation of opinion with implicit and behavior data,
Incorporation of opinion with implicit and behavior data, and
Natural language processing overheads (ambiguity),
Generation of highly content lexicon database, handling of
bipolar sentiments, dealing with short Sentence like
abbreviations, Requirement of World Knowledge, Negation.
All challenges have a bad effect on the understanding of
reviews.
In this paper, the research aims to fill this research gap by
proposing the new technique for sentiment analysis of online
scientific papers reviews (SAOOP). The technique also
measures efficiency by making a comparison between
SAOOP, and other two sentiment analysis techniques [16].
Namely “Natural Language Toolkit-Text processing” (NLTK)
and “recursive deep models for semantic” (NLPS). The
results depend on comparing accuracy, performance and rate
of coverage of language through two datasets.
The rest of this paper is organized as follows: Section 2
represents related works. Section 3, the presentation of the
220 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
new technique “SAOOP”. In Section 4, outlines of the
Experiment as well as the sample used for comparison.
Section 5 highlights the comparison results. Finally, Section 6
concludes and proposes directions for future work.
II.
RELATED WORKS
The purpose of this paper is sentiment evaluation which
means to find the sentiment polarity (positive, negative, or
neutral) of a text reviews data and evaluate the sentiment score
of the text review. Generally a text review is divided into
single sentences (“sentence-based”) and words (“wordsbased”) or very short texts from a single source.
A. Sentiment Analysis: An Overview
The author in (Sentiment analysis of document based on
annotation) presented a tool which judges the quality of text
based on annotations on scientific papers [17]. The authors's
methodology declares in collective’s sentiment of annotations
in two approaches. This methodology counts all the annotation
produces the documents and calculates total sentiment scores.
The problem of this methodology appears in a relationship
between annotations that is complex. The technique needs to
have a big query knowledge base containing metadata. The
notion declares in that the values are not accurate enough such
as the value of “Good=0.875” has greater value than the value
of “Best=0.75” although the result is wrong in logical
meaning. Nevertheless, believing that collecting metadata and
evaluating them could be useful to achieve higher analysis
quality.
The researchers proposed a “Web Based Opinion Mining
system” for hotel reviews [18]. They introduced an evaluation
system for online user’s reviews and comments to support
quality controls into hotel management. The research is
capable of detecting and retrieving reviews on the web and
deals with German reviews. The multi-topic/multi-polarity is
the method of this research; the system would recognize the
neutral e.g., “don’t know” to “classify sentiment polarity that
as neutral” and the multi-topic cases identified in their corpus.
The major weakness illustrates in not handling some cases in
multi-topic segments. The authors [19] analyzed sentiments
reviews of mobile devices products. Their Machine learning
(ML) [20] system investigates the classification accuracy of
Naïve Bayes algorithm. In addition to Judge the product
quality and status in the market is advantageous. They use
three machine learning algorithms (Naïve Base Classifier, Knearest neighbor [21], and random forest [22] to calculate the
sentiments accuracy. The random forest improves the
performance of the classifier.
B. Sentiment Analysis Techniques
This section provides a brief description of the two
sentiment analysis techniques investigated in this paper.
These techniques are the most popular in the literature and
they cover diverse techniques such as the use of Natural
Language Processing (NLP) [23] in assigning polarity and
sentiment score.
1) Natural Language Toolkit: The authors aim at an
evaluation sentiment scores and polarity. They produce the
Natural Language Toolkit (NLTK) [12]. NLTK is a text
analysis technique that evaluates cognitive and constitutional
components of a given text reviews based on using lexicon
including words. They use hierarchal sentiment classification
level with two levels (Neutral, Positive, and Negative). The
drawback of this technique illustrated in low accuracy and
some logical errors. Because the technique needs to increase
handling of language coverage [24].
2) NLP Stanford sentiment (NLPS): The researchers
introduce recursive neural models have in common: word
vector representations and classification [25]. The authors
released a tool named “NLP Stanford” NLPS [26], which
develops an integration of learning techniques that produces
better results and higher accuracy training model empirically.
Their goal is based on Semantic word spaces have been very
beneficial but NLPS cannot express the meaning of longer
phrases in a primary way. So they improve this technique by
detection the sentiment requires wider supervised training and
evaluation resources.
III.
SENTIMENT ANALYSIS OF ONLINE PAPERS (SAOOP)
In this section, Sentiment analysis of online papers
“SAOOP” will be presented. SAOOP is used in opinion
mining [27] and based on a new English lexical dictionary
[28]. This lexical dictionary groups adjectives, nouns, verbs,
adverbs, adjectives, prefixes, suffixes and other grammatical
classes into synonym. The proposed technique is an
enhancement on Bag-Of-Words (BOW) model [29] in
sentiment analysis to achieve high accuracy, which depends
on word weight replacing term frequency of each word. The
proposed technique solves two important Bag-of-words
weaknesses.
The standard bag of words is not automatic in
classification and creating polarity lexicons because BOW
model needs to create manual lists of 'positive' and 'negative'
words [30]. That means the review judgment is based on the
probability of positive or negative words. The second is low
accuracy because the standard BOW model neglects text
grammatically. Sentiment classification levels will be divided
into five classes (very positive, positive, negative, very
negative and Neutral).
The proposed technique makes the sentiment classification
levels are more detailed and easy by word percentage of each
class. The goal of SAOOP is for inferring the polarity of
common meaning and polarity concepts from natural language
text at a word level, rather than at the syntactic level. SAOOP
also classifies reviews into some categorizations based on
papers parameters. In addition, the estimation rank of each
paper based on evaluation some parameters.
221 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
A. SAOOP Overview
Fig. 1. SAOOP Overview
“Fig.1” shows that SAOOP model consists of two
components sentiment score and system score. SAOOP can
evaluate any paper based on the components. Sentiment score
depends on total reviews evaluation score. And system score
which depends on the sum of total scores of three parameters
of paper (place of publication), citation number of paper and
paper publishing date. SAOOP technique helps researchers to
select the suitable paper with the total paper score.
detection for each one and each sentence and calculate the
total score of sentiment review score. In classification phase,
that’s splitting into two parts, first the reviews classification
into five sentiment classification levels (very positive,
positive, negative, very negative and objective (neutral), also
having degrees of each sentiment level with scale from [-1, 0,
1]. There is also another classification which declares each
review categorization based on five meaning classes (topic,
date, author, citation, and place of publication). The benefit
from the extracted data to memorize them and make
relationships between evaluated papers and reviews and
categorize reviews logically based on topic domain
parameters. Output is the sentiment evaluation score of all
reviews with all papers with caring of number of reviews
parameter, and evaluation of scientific paper parameters score
which is based on metadata of each paper (place of
publication, publishing date, and number of citation). So the
consequent result is ranking to each paper with the total score
of sentiment and system scores with graphical reports of
results.
B. SAOOP Methodology
SAOOP can assign polarity based on this approach,
considering the words weight replacing term frequency, by
assuming each word has two values and polarity with this
assumption equation,
( ) ∑( ( )
( ))
.
(1)
V (w) is value of word, W (p) refers to positive value and
W (n) refers to negative word, the selection between positive
or negative polarity Influenced by the meaning of words and
each other polarity. But the sentence contains negative that
differs in the word value. If the word is positive, convert to
negative polarity and the negative score will be as in the
equation,
( )
( )
.
(2)
And if the word is negative, the score will be calculated by
V (w) = ( )
. The selection of 0.2 because this disison
is suitable for the five sentiment class’s levels [18]. The
proposed technique also creates papers ranks with calculating
sentiment and measuring domain parameters. By assuming,
Fig. 2. SAOOP Technique overview
“Fig.2” declares SAOOP Technique overview. The input
is scientific paper website link. In data extraction [31] level
two parts: first, using Easy web extract tool which is web
scraping tool to extract data of paper from scientific papers
website online. Part two is data reformat from Excel sheet
which is one output of EasWebExtract tool [32] suitable with
SAOOP database format. In text analysis level, SAOOP
applies some functions of text analysis on reviews of each
paper. In the first, applying the splitting sentences function,
tokenizing words function, and checking of stop list and
removing them [33].
In review understanding (NLP) level, the proposed
technique understands the sentences meaning and check words
in vocabulary lexicon with similarity and differences
algorithms. In estimation phase, showing the evaluation
sentiment score for each word into text review and the polarity
( ) ∑ ( ( )
( )).
(3)
In the equation, P (TS) refers to a total score of each paper,
T (SA): is a total score of sentiment score of all reviews on
each paper with caring of number of positive reviews. In the
next equation,
(
( ))
) ∑
.
(4)
The calculation of the total score of all reviews depends on
the score of each review. There is a difficult problem between
large number of reviews and evaluating sentiment polarity of
each one, this problem is improper the most review number
having assessment higher score. For example, one paper
publishing in 2013 that’s mean from 2 years and this paper has
twenty reviews, not equal evaluation one paper publishing in
2005, that’s mean from 10 years and the second paper has
twenty reviews. The first one is the top rated because the
evaluation number of reviews in short time. In other example,
one paper publishing from 2 years and having twenty negative
(
222 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
reviews, not equal evaluation other one publishing from 10
years and having positive twenty reviews. The second one has
maximum rated because the evaluation numbers of positive
reviews is larger than the one, although the second is the
oldest. As mentioned before double trouble with reviews
number and the relationship between date and other relation
between sentiment polarity of reviews and number of reviews.
That interprets difficulty of evaluation domain parameters.
The proposed technique faces these challenges and
evaluates the percentage of positive reviews over total scores.
But still there is a problem in relationship between date and
number of reviews, for example: one paper publishing from 2
years which has twenty positive reviews, not equal evaluation
other one publishing from 10 years which has positive twenty
reviews. Actually that is not equal their selves because the
recent has bigger reviews number. So SAOOP presents a
solution for date relation with reviews number, according with
two parameters number of positive reviews and the recent
paper. T (SS): is a total score of system score parameters that
are evaluated logically of paper parameters according to this
equation,
(
(
)
( )
( )
) ∑(
).
(5)
V (SS) expresses the value of systems score. S (PP) means
the score of publication place, S(C) refers to the score of paper
citation number, and S (D) means the score of paper publish
date. Assuming λ is a constant equal 2, dividing into λ and 2λ
to determine the priority of evaluation of the parameter. The
evaluation topic parameters process does not ease because of
depending on the logical meaning of each one. So the research
focuses on scientific papers domain to put the foundation of
evaluation parameters to achieve the fact value of each paper
to support researcher with sentiment analysis by ranking
papers based on total score of them. There is inverse
relationship between publishing date and number of citation of
the paper, which declare in this equation,
( )
.
(6)
The result is not true the highest citation number having
the highest evaluation score of it. For example, one paper
publishing in 2013 that’s mean from 2 years which has ten
citations, not equal evaluation one paper publishing in 2005,
that’s mean from 10 years which has ten citations. The first
one is the highest score because number of citations in shortly
is high, this first paper will be predicated if the paper has the
same time 10 years, it mostly has 50 reviews not 10 reviews
such as the second paper. In other words, the first paper has 5
papers into each year but the second has 1 into each year. To
evaluate score of publishing place conference which depends
on ACM conferences tiers with a sample into computer
science conferences, such as “VLDB: Very Large Data Bases
is in the top tier: tier 1”, “ER: Intl Conf. on Conceptual
Modeling (Conf. on the entity Relationship Approach)” is in
next tier which is in lower tier: Tier 2, and “IDEAS: Intl
Database Engineering and Application Symposium” is in a
lower tier: Tier 3” [34].
C. SAOOP & Sentiment Challenges
SAOOP enables to make solutions to most significance
sentiment analysis challenges [35]. The proposed technique
can produce some solutions for main challenges to reach to
higher accuracy. The discussion of the solutions in the
following:
1) Topic domain independence
Domain-dependent [36] is a difficult challenge to
recognize topic nature. There are some words have many
meanings and different sentiment values relevant to the topic.
There is also a problem shows in extracting keyword or
features and how to evaluate words based on each topic. One
feature set may give very good performance in one domain, at
the same time it performs very poor in some other domain.
The produced solution suitable with a small scale by applying
the proposed technique on one topic domain and examine
domain parameters evaluation by categorization reviews
because they also give different meaning with the same word.
This research presents a technique to recognize topic nature
automatically. The proposed technique is based on extracting
keywords and relevant features of each topic. In addition, to
produce a solution for some words have many meanings and
different sentiment values relevant to the topic. The proposed
technique is based on Classification review of each domain
features and keywords.
For example, “IEEE is [great +] publication for your
paper”, SAOOP can put IEEE is in a place of publication
classification (based on feature name of publication) and the
polarity is positive. “The publishing conference is [great+]”,
this review refers to the place of publication classification
(based on keywords) and the polarity refers to positive. In
other example, “The paper publishing date is [old-]”, this
reviews refers date classification (based on date attribute) and
“Old” having the negative score. “The author is [old -] in this
field”, but SAOOP can categorize the last review in author
classification that is meaning the author is expert in this field
so “Old” will be had positive score.
SAOOP improves the sentiment score to be more accurate
and fair. By assuming some words have 0 value because of
depending on classifications of each sentence of each review,
there are some groups of words having a polarity and score to
relate with the detected classification.
2) Negation
Negation is the biggest challenge in sentiment analysis
[37]. The new technique produces a solution to improve
evaluation negative with the enhanced bag of words
technique. This research handles the two techniques: explicitly
and implicitly negative [38]. First: explicitly is deliberately
formed and are easy to self-report and by keywords. Second
implicitly [5] is the unconscious level, are involuntarily
formed and are typically unknown to us without any keywords
of negative. In addition, the handling the negative meaning of
some conjunctions such as “not only”, and “But”. The dual
negative is the most important case which cares to achieve the
total sentiment polarity. Reverses polarity of mid-level terms:
great V.S not great.
A method often followed in handling negation explicitly in
sentences like:
“I do [not like+] − the paper”, is to detect the negative
polarity because the word (not) and convert the sentence
223 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
operator to negative. But this does not work for “I do [not
like+] − this research but I [like+] the field”. But still there can
be problems.
1.
2.
3.
4.
5.
6.
Other example, “I find the functionality of this new
methodology [less -] practical”, this review refers to explicit
comparative negative. “This algorithm is [not great +]-]”, the
proposed technique handles in this review the positive and
negative evaluation which declares in [not great! = bad] but
[not great = good]. Implicitly negative such as “This research
is [very [complex -] -]” this example does not have any
negative keyword, but the meaning has negative and the
polarity will be negative of this sentence.
7.
8.
9.
There are sentences having keywords of negation and they
don’t have the negative polarity such as “[Not only+] I [like+]
this algorithm, but also [easy+] to understand and apply.” the
polarity is not reversed after “not” due to the presence of
“only”. So this type of combinations of “not” with other words
like “only” has to be kept in mind while designing the
algorithm.
There is a difference between “not only” and not because
not only strengths the meaning (more positive or less
negative) based on the polarity of the sentence. In this
example other case of implicit negative, I [wish -] to work
[harder -]”. In the last review, the new technique presents
future words e.g. wish refers to the negative polarity but first
must check the polarity of the next sentence polarity because
maybe changed the polarity depends on meaning.
3) Creation lexicon
The proposed SAOOP yields an improvement over prior
published bag of words built lexicons. This technique also
provides an improvement in calculation technique used in
reviews sentiment analysis. SAOOP technique presents a
solution to take care of grammar (which is one of limitation of
Bag-Of-Words) and to save time took is N-gram algorithm to
create subsequences of terms. There are two phases that will
be produced:
 Phase 1. Data Preparation Phase
Less number of words in vocabulary lexicon to fast search
based on similarity and differences algorithms. SAOOP
neglects verbs tenses or word formula (singular or plural),
that’s meaning neglecting English grammar and syntax
because of the comparison and differentiation with the
infinitive verbs, and singular words with most letters
similarity.
 Phase 2. Lexicon Development Phase
Evaluation words /terms: is based on enhanced bag of
words: the proposed technique doesnot depend on term
frequency. This phase is based on assuming each word has
two values and the total of them equal 1. Each term has 2
polarities (+/-).
10.
11.
12.
13.
Negative Algorithm
For each review R in paper P sets
For each sentence Sent in Review R
Apply Pre-Processing: Remove the stop words
Convert all words to Upper case
Check on expressions have “No or Not” e.g Not Only
Check on first Negation keywords list, which effects on
the polarity of the words e.g “Not” I don’t like this
paper.
assuming the negative value for positive word
( )
nd assume the positive value for the negative word
Check on the next word after negative e.g “like” has
positive value and polarity but here it will take negative
polarity and value.
Detect the polarity and value.
Check on the second list of negation keywords, which
effects on the polarity of sentence e.g. never, yet, neither.
Convert polarity of the sentence by multiplication with(1) ,
(
) Check on future words e.g “wish/hope”.
14. Check on the next sentence polarity.
15. End for
16. Detect sentence value and polarity.
17. End For
18. calculate review value and polarity
(Note: knowing our attention of review classification.)
4) World knowledge requirement
SAOOP technique produces a solution for Knowledge
about worlds’ facts, events, people are often required to
correctly classify the text. Trying to achieve higher accuracy
and get the evaluation for some neutral reviews. The World
knowledge challenge solution is based on the hierarchical
database of nouns. Semantic (hierarchal) relationships
between nouns to achieve the polarity, score and meaning.
Also to differ between them and keywords or features.
Consider the following example, “the author is a [lion-] in this
field”, the previous review present negative polarity because
lion is a name of animal but in real evaluation in the review
refers to a positive polarity. In the next review, “Bing is really
[Einstein?]” evaluation sentiment analysis without world
knowledge classifies above sentence as neutral, but this
review is an objective sentence because Einstein is the name
of the famous scientist, so it refers a positive polarity also.
This review is very hard for software to understand that
automatically. SAOOP creates a huge lexicon database to
contain the world knowledge especially related to researchers
and the most common in the reviews. The solution of world
knowledge also assumes values of the words based on the
most common meaning. The evaluation of these world
knowledge depends on keywords and classification of
reviews.
5) Spam and Fake Reviews:
The WWW contains both realistic and spam contents. For
effective Sentiment classification, this spam content should be
eliminated before processing.
224 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
SAOOP can be done by empty or identifying duplicates,
by detecting outliers and by considering the reviewer
reputation. The proposed Technique enhances reviews spam
and fake. SAOOP technique can avoid and cure the most of
them by:
 Remove empty reviews: To calculate the real number of
reviews.
 Delete duplicate reviews by considering the same
reviewer: To Calculate the real sentiment score of the
paper.
For example, one paper has 10 reviews, 3 of them for the
same text review and with the same user, and 2 is empty
reviews , in most sentiment application, if having 10 reviews
number and the same repeated reviews will calculate together,
the sentiment score is not real because having fake reviews
and the results became fake also. And also there are some
reviews are general are not related to the paper actually.
SAOOP can produce solution for the case study on
citeulike.com website, through making quaternary relationship
between a set of paper parameters “paper name”, “author
names”, “review” and “Username” (who is a review writer)
with taking into consideration review written time, if the
review is repeated by the same review writer with ensuring if
the review is fake by all parameters and time, the proposed
technique will delete the spam review before calculating the
sentiment analysis. SAOOP can also deal with fake reviews if
it empty and deleted.
In this paper, showing the implementation of SAOOP
technique using C# programming language working on
Microsoft visual studio 2010 platform. The newly created
lexicon is based on SQL Server Management Studio 2008.
IV.
EXPERIMENT
The second part is a test data: which is a set of data to evaluate
sentiment with hide class label around 5000 text reviews.
Citeulike receives in excess of 200,000 distinct visits (defined
by Google Analytics as a group of page views by a unique
user with timeout after 30 minutes inactive) monthly, with
each visit originating an average of 2.77 page views [41]. Of
that 200,000 around distinct users who have previously visited
the site on multiple occasions.
There are currently 505,402 items posted in the database
(counting n people post the same article); 1,676,130 tags
(counting n if there are 'n' tags applied to an article); and
130,548 distinct words used these numbers are growing
exponentially. This sample set allows us to study the
responses to noticeable past texts. In addition, to evaluate the
improved levels of techniques, methodologies in sentiment
analysis. SAOOP can handle ten cases to ease to understand
text review accurately by CiteULike users they illustrated in
table 1. SAOOP can care and evaluate of some English
grammar to improve BOW model.
2) Verified dataset
The second dataset which is called verified data set is a
real data set but they can’t be known the evaluation before.
The dataset has around 10.000 text reviews in this sample.
This data set is splitting into two parts of verified data reviews
as positive and negative. These datasets include a wide range
of online papers texts reviews: general reviews. In Table 2,
the sample reviews of online scientific papers. SAOOP
technique can evaluate sentiment score with the relationship of
reviews categorization. With applying on this human-verified
sample set [29], by fitting to quantify the range with different
sentiment analysis techniques can accurately evaluate polarity
of text reviews.
TABLE I.
In this section, the discussion of the comparison between
the proposed technique and two sentiment analysis techniques.
This comparison shows the accuracy and performance results
based on two datasets. This comparison also compares with
the effects and solutions of sentiment analysis challenges.
A. Datasets
The comparison uses two different datasets: 1) real data
set: which splits into two data sets with training set (1000 text
reviews) and test set (5000 text reviews), 2) verified data set:
which is a real set with unknown evaluation around10.000 text
reviews (including more than 5.000 positive words, 5.000
negative words).
1) Real dataset
The first sample set is a sample of WWW.citeulike.com
papers reviews and Metadata posted by computer science
papers branch [39, 34]. The comparison in real data set in
computer science scope including two parts: training data and
test data [40]. Training data is a set of data to evaluate
sentiment around 1000 reviews, knowing the values before.
TABLE ENGLISH LANAUGE COVERAGE HANDLING BY SAOOP
Cases
1. Expressions
2.Topic objects[3]
3.Negation
4.Suffixes &prefixes
5.Verbs
6.Adjective&
Adverb
7.Nouns
8.Comparative [4]
comparisons
9. Phrases
10. Some special
(Need, Wish)
Definition & Examples
Based on syntax (e.g., Not Only, No one.)
Features (e.g., lot of contributions)
Implicit (e.g., Independent)) and explicit (e.g. not
bad, does not very good)meaning
The beginning or end letters of word to have
different meaning (e.g., dislike, opposed to, useless)
Converting verbs tenses into infinitive
e.g., Well, improved, highly
e.g., algorithms, improving, enhancing
e.g., easier convert to easy. (“More”; “higher”)
and (“most”, “highest”).
e.g., very good, the professional work
(e.g., hope, wish): in the most times, they have
negative polarity.
225 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
Fig. 3. Coverage Rate of ten cases understanding cases with the three techniques
TABLE II.
which positive reviews are predicted to be positive (R),
whereas the true negative rate is the rate at which negative
reviews are predicted to be negative.
SAMPLE OF REVIEWS
Sample of reviews
This paper is very well.
It’s not great
The best web point
I am interesting in this field
Extremely good
It’s not only hot research area but also having new scientific contributions.
This point research is more affected in web mining than using in neural
network.
high accuracy
It’s not have good value enough
V.
Citation is valuable
B. Comparsion Measures
In order to define the evaluation of accuracy and
performance of the three techniques, which will consider in
the following table.3:
TABLE III.
MEASUREMENT TABLE
Actual observation
Positive Negative
Predicted
expectation
Positive
Negative
The accuracy represents the rate at which the method
predicts results correctly (A). The precision also called the
positive predictive rate, calculates how close the measured
values are to each other (P). The comparison also provides the
F-measure results, since it is a standard way of summarizing
precision and recall (F). Ideally, a polarity identification
method reaches the maximum value of the F- measure, which
is 1, meaning that its polarity classification is perfect. The yaxis is a percentage of the understanding sentence rate.
x
z
y
w
Let present True positive (x) was defined when a text was
correctly classified as positive, False Positive (y) is a
negative text which was classified as positive, False Negative
(z ) is a positive text but was classified as negative, and the
last one True Negative (w) is a correctly classified as negative
[42]. In order to compare and evaluate the techniques, by
considering the following metrics, commonly used in
information retrieval: true positive rate or recall: R = x/(x +
z), false positive rate or precision: P = x/(x + y), accuracy:
A = (x + w)/(x + y + z + w), and F-measure (performance): F
= 2 • (P • R)/(P + R). In many cases simply use the Fmeasure, as it is a measure of a test’s accuracy and relies on
both the precision and recall [10]. By reporting, all the
measurement mentioned above by practical interpretation. The
true positive rate or recall can be understood as the rate at
COMPARISON RESULTS
In order to facilitate understanding the advantages,
disadvantages, and limitations of the various sentiment
analysis techniques [43]. This section also presents the
comparison results among them.
Understanding of word coverage: in the beginning, the
comparison of the coverage of English grammar cases across
the representative scientific reviews from CiteULike website.
Then examination the intersection of the covered reviews
cases across the techniques were in table 1. “Fig. 3 (a)” shows
the result for the proposed technique SAOOP, which explain
in section 4. “Fig. 3 (b)” declares the NLTK technique. NLTK
which is a teaching tool works in, computational linguistics
using Python [44]. And “Fig.3 (c)” shows NLP technique.
NLPS technique which is predicting the sentiment of reviews
based on a recursive model.
As shown in the figure, SAOOP has the highest
understanding sentence coverage with 82.5 % with two data
sets with three data sets samples, respectively, followed by
NLPS which can’t evaluate the total sentiment score but with
detecting word by word polarity its percentage is 72%.
NLTK can interpret less than 10% of all relevant reviews.
In addition, we compare with the percentage of handling
sentiment analysis challenges to high accuracy and
performance of sentiment analysis of the three techniques of
the text reviews depicted in “Fig. 3”.
226 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
Fig. 4. Perecentage of handling sentiment analysis with the three techniques
According to “Fig.4 (d)” in fact, SAOOP had a new
solution for some sentiment challenges but NLPS and NLTK,
they and can’t produce methodology to solve them expect
some cases in negation but they have many logical errors, that
shown by “Fig.4 (E) and (F)”. The analysis results in table 3,
refers to the: Percentage of accuracies between techniques
based on different data set size. Also we examine the average
result analysis of the two big data set that spirited into three
data sets, that illustrate the highest average results with
sentiment score of the proposed SAOOP technique then NLPS
and the lowest one is a NLTK Technique. Finally, the
summarization the results with the average of the three data
sets (real and verified sets), we find the average of sentiment
score of the proposed technique improve the results. Because
of working binary analysis solutions of some important
challenges and evaluate some technical cases in the text which
have a problem in evaluation to be more accurate. In next
section, we discuss the accuracy results of the comparison.
a) Accuracy: With the examination of the percentage
degree of different techniques accuracy on text reviews
content. In order to compute the accuracy of each technique,
by calculating the intersections of the positive or negative
proportion given by each technique. Table.4 presents the
percentage of accuracy for the three compared techniques. For
each technique in the first column, showing the estimation
from the two data sets of reviews. Finding that some
techniques have a high coefficient as in the case of SAOOP
(82.5%), while others have least overlap such as NLTK (62%)
and NLPS (70.2%).
The last “column” of the table shows on average to what
extent each technique agrees with the other two samples. The
last “row” quantifies how other methods agree with a certain
technique, on average.
With the results of table 4, they illustrate differences
between accuracy and performance of the three techniques.
Table 4 shows techniques recall, precision, accuracy and
performance.
“Fig.5” is shown the accuracy results of them. In a
summary, the result indicates that existing tools vary widely in
terms of accuracy about sentiment score, with scores ranging
from 60% to 80%.
TABLE IV.
AVERAGE RESULTS FOR ALL DATASETS
Metric
SAOOP
NLTK
NLPS
Recall
Precision
Accuracy
F-measure
0.856
0.867
0.817
0.846
0.571
0.845
0.629
0.665
0.253
0.846
0.715
0.729
Fig. 5. Differences between Accuracies of three techniques
TABLE V.
PRECENTAGE OF ACCURACY BETWEEN TECHNIQUES
Dataset
Training set 1.000
Test set 5.000
Teal set 10.000
SAOOP
83.5%
81.99%
82.5%
NLTK
62%
61%
60%
NLPS
72%
70%
71.56%
Average
72%
70.99%
71.186%
Average
82.5%
61.514%
71.604%
-
b) Perfromance: In this section, showing an
evaluation of the performance of the three compared
techniques. For comparing the performance results,
Table.5 which gives the average of the results obtained for
all datasets. For the F-measure, a score of 1 is ideal and 0
is the worst possible.The technique with the highest Fmeasure was faced sentiment analysis challenges and
cover ten cases of each text review (0.846), which had the
highest sentiment accurate and understanding text
coverage. The second rated technique in the understanding
of F-measure is NLPS, which obtained a much higher
227 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
coverage than understanding and challenges. It is
important to note this problem that it can’t be interpreted
into of total score of the text review. For observation better
performance on data sets that contain more expressed
sentiment, such as text reviews (e.g., papers online) and
the lowest performance is NLTK technique.
VI.
[2]
[3]
[4]
[5]
[6]
[7]
[9]
[10]
CONCLUSION
Sentiment analysis is the most important source in decision
making. Almost people becomes depends on it to achieve the
efficient product. Thousands of researchers rapidly year by
year that focuses on scientific online reviews for papers to
help them. So the researchers introduce a new sentiment
technique. In this paper, the researchers create a new
technique is called sentiment analysis of online papers
“SAOOP”. The proposed technique will be a suitable and
efficient solution to analyze online reviews. The target of
technique to improve accuracy and achieve to accurate review
meaning. The proposed SAOOP approach is based on two
methods: evaluation and analysis reviews (sentiment analysis)
and solve some sentiment analysis challenges. In order to
serve researchers in selecting efficient papers. In addition, it
evaluates topic domain parameters of scientific papers (place
of publication, publishing date, and a number of citation
paper) to evaluate the total score of papers. To evaluate
SAOOP efficiency, making a comparison between it and two
famous techniques. The results have a comparison between
the accuracy and performance between the three techniques
when the researchers apply the techniques on three data sets
(training, test and verified). The comparison results illustrate
how proposed technique can increase accuracy and
performance with facing many language coverage cases and
solving some sentiment analysis challenges. The accuracy
results show in NLTK (62%) and NLPS (70%) to 82%
(SAOOP) with the proposed technique.
[1]
[8]
REFERENCES
Suman, D.R, Wenjun, Z., Social Multimedia Signals: A Signal
Processing Approach to Social Network Phenonmena, ISBN-13: 9783319091167, Springer International Publishing Switzerland, 2015.
Larsen, Peder Olesen, and Markus von Ins. “The Rate of Growth in
Scientific Publication and the Decline in Coverage Provided by Science
Citation Index.”Scientometrics 84.3 (2010): 575–603. PMC. Web. 25
Sept. 2015.
Peng, L., Cui, G., Zhuang, M., & Li, C. (2014). What do seller
manipulations of online product reviews mean to consumers? (HKIBS
Working Paper Series 070-1314). Hong Kong: Hong Kong Institute of
Business Studies, Lingnan University.
Thomas, B., Keep Social Honest , “What Consumers Think about brands
on social media, and what bunesses need to do about it” Report, 2013.
Namita, M., Basant, A., Garvit, C., Nitin, B., & Prateek, P., “Sentiment
analysis of Hindi Review based on Negation and Discourse Relation”,
International Joint Conference on Natural Language Processing,
Nagoya, Japan, 2013.
Theresa,W. ,Janyce, W. ,& Paul, H., “Recognizing Contextual Polarity
in Phrase-Level Sentiment Analysis”, proceeding HLT’05 proceedings
of the conference of Human Language Technology and Empirical
Methods in Natural Language Processing, 2006.
Ian, H.W., Eibe, F., & Mark A. H., Data Mining: Practical machine
learning tools and techniques. The Morgan Kaufmann series in data
management systems, 3rd Edition, Morgan Kaufmann Publishers is an
imprint of Elsevier, 2011.
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
Yin, Z., Rong, J., & Zhi-Hua, Z., “Understanding Bag-of-Words Model:
A Statistical Framework”, International Journal of Machine Learning
and Cybernetics, 2010.
Diana, M., & John.C.,“Disambiguating Nouns, verbs, and adjectives
using automatically acquired Selectional preferences”, Association for
Computational Linguistics , Vol. 29, 2003.
Presutti, V., Stankovic, M., Cambria, E., Cantador, I., Di Iorio, A., Di
Noia, T., Lange, C., Reforgiato Recupero, D., Tordai, A., “Semantic
web evaluation challenge. 1st Edition, Series title: Communications in
Computer and Information Science, Springer International Publishing,
2014.
Bing, L., “Sentiment Analysis and Subjectivity”. In Nitin, I. & Fred, J.
(eds). Handbook of Natural Language Processing. 2nd Ed, Machine
Learning & pattern recognition series, Chapman& Hall/CRC, 2010.
Samih,Y., Erdogan,Y.,& Halife, K., “Tagging Accuracy Analysis on
Part-of-Speech Taggers”, Journal of Computer and Communications,
2014.
Mitchell, P.M., Mary A.M., and Beatrice, S., “Building a Large
Annotated Corpus of English: The Penn Treebank”, Computational
Linguistics Journal , Voul. 19, Number 21993.
Saifee, V., & Jay, T., “Applications and Challenges for Sentiment
Analysis: A Survey”, International Journal of Engineering Research &
Technology (IJERT), Vol. 2 Issue2, 2013.
Hassena, R.P., “Opinion Mining and Sentiment Analysis Challenges and
Applications”, International Journal of Application or Innovation in
Engineering & Management (IJAIEM), Volume 3, Issue 5, 2014.
Bing, L., Sentiment Analysis and Opinion Mining. Morgan & Claypool
Publishers, 2012.
Archana, S., “Sentiment analysis of document based on annotation”,
CORR Journal, Vol. abs/1111.1648, 2011.
Walter, K., & Mihaela, V., “Sentiment analysis for hotel reviews”,
proceedings of the computational linguistics-applications, Jacharanka
Conference, 2011.
A.Tamilsevi, & M.Parveen, T., “Sentiment analysis of micro blogs using
opinion mining classification algorithm”, International Journal of
Science and Research (IJSR), Vol. 2 Issue 10, 2013.
Nisha.J, & Dr.E. Kirubakaran, “M-Learning sentiment analysis with
Data Mining Techniques”, International Journal of Computer Science
and Telecommunications, Volume 3, Issue 8, 2012.
G¨unes, E., Ahamed, H., Qian, D., & Dragomir, R., “Improved Nearest
Neighbor Methods for Text”, Technical Report CSE-TR-576-11,
University of Michigan, Department of Electrical Engineering and
Computer Science, 2011.
Frederick, L., “Implementation of Breiman’s Random Forest Machine
Learning Algorithm”, ECE591Q Machine Learning Journal Paper, 2005.
Christof, M., “Machine learning for query formulation in question and
answering”, Natural Language Engineering, Cambridge University,
2011.
Steven, B., Ewan, K., and Edward, L., Natural Language Processing
with Python, 1st Edition, O’Reilly Media publisher, 2009.
Richard, S., Alex, P., Jean, Y.W., Jason, C., Christopher, D.M., Andrew,
Y.N. & Christopher, P., “Recursive Deep Models for Semantic
Compositionality Over a Sentiment Treebank”, Conference on Empirical
Methods in Natural Language Processing, 2013.
I. Hemalatha , Dr. G. P Saradhi V., & Dr. A. Govardhan, “Preprocessing
the Informal Text for efficient Sentiment Analysis” , International
Journal of Emerging Trends & Technology in Computer Science
(IJETTCS),Volume 1, Issue 2, 2012.
Dushyant, B.R., & Samrat, K., “A Review on Emerging Trends of Web
Mining and IT’s Application”, International Journal of Engineering
Development and Research | IJEDR, 2013.
Bo, P., & Lillian, L., “Opinion mining and sentiment analysis”, Journal
Foundations and Trends in Information Retrieval, Vol. 2, 2008.
Bing, L., “Sentiment Analysis: A Multi-Faceted Problem”, IEEE
Intelligent Systems, 2010.
228 | P a g e
www.ijacsa.thesai.org
(IJACSA) International Journal of Advanced Computer Science and Applications,
Vol. 6, No. 9, 2015
[30] Hang, C., Vibhu, M., & Mayur, D., “Comparative experiments on
sentiment classification for online product reviews”, American
Association for Artificial Intelligence (AAAI Conference), 2006.
[31] Mr. Akshay, A. Adsod, & Prof. Nitin R.C., “A review on web mining”,
International Journal of Engineering Trends and Technology (IJETT)”,
Volume 10 Number 3, 2014.
[32] Alberto, H.F.L., Berthier, A.R-N, Altigran, S. d. S., and Juliana, S. T.,
“A brief survey of web data extraction tools”, ACM SIGMOD, Vol.31,
Issue 2,, New York , USA, June 2002
[33] Farah, B., Carmine, C., & Diego.R., “Sentiment analysis: Adjectives and
Adverbs are better than Adjectives Alone”, International Conference on
Weblogs and Social Media -ICWSM , Boulder, CO USA, 2007.
[34] Utkarsh, G., Clemente, I., and Mike P. W.,Understanding Factors
Inuencing the Citation Count of Networking Conference Papers, in proc.
CHI 09 proceedings of the SIGCHI Conference on human factors in
computing systems, 2009.
[35] Sujata, R. & Patreek, K., “Challenges of Sentiment Analysis and
Existing State of Art”, International Journal of Innovation and Research
in Computer Science (IJIRCS), 2014.
[36] Chenghua, L., and Yulan, H., “Joint Sentiment/Topic Model for
Sentiment Analysis”, CIKM '09 Proceedings of the 18th ACM
conference on Information and knowledge management, Hong Kong,
China, ACM 978-1-60558-512-3/09/11, 2009.
[37] Andreas, H., Gerhard, P., and Andreas, N., “A Brief Survey of Text
Mining”, LDV Forum - GLDV Journal for Computational Linguistics
and Language Technology, 2005.
[38] Bas, H., Paul, v.I., Alexander, H., Flavius, F., & Uzay, K., “Accounting
for Negation in sentiment analysis”, DIR, Amsterdam, the Netherlands,
2011.
[39] John, M. B., “Counting publications - Journals vs. Conferences in
Computer Science”, in NIK-2013 conference, 2013.
[40] Quoc, L., & Tomas, M., “Distributed Representations of Sentences and
Documents”, Proceedings of the 31 st International Conference on
Machine Learning, Beijing, China. JMLR: W&CP volume 32, 2014.
[41] Umer, F.,Yang, S., John, M. C., and Lee, G.,“Social Bookmarking for
Scholarly Digital Libraries”, Published by the IEEE Computer Society,
Pennsylvania State University, December 2007.
[42] Pollyanna, G., Matheus, A., Fabrico, B., & Meeyoung, C.,“Comparing
and Combining Sentiment Analysis Methods”, ACM Association for
Computing Machinery, 2013.
[43] Mark, C., Oliver, D., & Fatih, U., “Potential and Limitations of
Commercial Sentiment Detection Tools”, Emotion and Sentiment in
Social and Expressive Media-ESSEM, 2013.
[44] Steven, B., Ewan, K., Edward, L., & Jason, B., “Multidisciplinary
Instruction with the Natural Language Toolkit”, Proceedings of the
Third Workshop on Issues in Teaching Computational Linguistics
(TeachCL-08),
Columbus,
Ohio,
USA,
2008.
229 | P a g e
www.ijacsa.thesai.org
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