IEEE SPONSORED 2ND INTERNATIONAL CONFERENCE ON ELECTRONICS AND COMMUNICATION

SYSTEMS(ICECS ‘2015)

Extracting Aspects and Mining Opinions in Product

Reviews using Supervised Learning Algorithm
A.Jeyapriya (P.G Scholar) C.S.Kanimozhi Selvi (Associate Professor)
Department of Computer science and Engineering Department of Computer science and Engineering
Kongu Engineering College Kongu Engineering College
Erode, Tamilnadu, India. Erode, Tamilnadu, India.
dauntlessjeya@gmail.com kanimozhi@kongu.ac.in

Abstract— Social media is emerging rapidly on the internet.

This media knowledge helps people, company and organizations A. Opinion Mining
to analyze information for important decision making. Opinion Opinion mining is extracting people’s opinion from the
mining is also called as sentiment analysis which involves in web. It analyzes people’s opinions, appraisals, attitudes, and
building a system to gather and examine opinions about the emotions toward organizations, entities, persons, issues,
product made in reviews or tweets, comments, blog posts on the actions, topics, and their attributes in Liu [1]. Opinion is
web. Sentiment is classified automatically for important quintuple(e j , a jk , so ijkl , h i , t l ) where e j is target entity , a jk is
applications such as opinion mining and summarization. To make aspect of entity , h i is opinion holder , t l is the time when the
valuable decisions in marketing analysis where implement opinion is expressed , so ijkl is sentiment orientation of opinion
sentiment classification efficiently. Reviews contain sentiment holder h i on feature a jk of entity e j at time t l . Users express
which is expressed in a different way in different domains and it
their opinions about products or services they consume in blog
is costly to annotate data for each new domain. The analysis of
posts, shopping sites, or review sites. It is useful for both the
online customer reviews in which firms cannot discover what
exactly people liked and did not like in document-level and
consumers as well as for the producers to know what general
sentence-level opinion mining. So, now opinion mining ongoing public think about a particular product or service. Sentiment
research is in phrase-level opinion mining. It performs finer- analysis and opinion mining aim to automatically extract
grained analysis and directly looks at the opinion in online opinions expressed in the user-generated content. There are
reviews. The proposed system is based on phrase-level to examine many social media sites reporting user opinions of products in
customer reviews. Phrase-level opinion mining is also well-known many different formats. Monitoring these opinions related to a
as aspect based opinion mining. It is used to extract most particular company or product on social media sites is a new
important aspects of an item and to predict the orientation of challenging one. Opinion mining tools allow businesses to
each aspect from the item reviews. The projected system understand new product opinion, product sentiments, brand
implements aspect extraction using frequent itemset mining in view and reputation management. These tools help users to
customer product reviews and mining opinions whether it is perceive product opinions or sentiments on a global scale.
positive or negative opinion. It identifies sentiment orientation of Supervised learning algorithms that require labeled data have
each aspect by supervised learning algorithms in customer been successfully used to build sentiment classifiers for a given
reviews. domain. However, sentiment is expressed differently in
different domains, and it is expensive to interpret data for each
Keywords— aspect based opinion mining, frequent itemset novel domain.
mining, sentiment orientation .

I. INTRODUCTION B. Levels of opinion Mining

Data mining research has successfully shaped numerous Opinion mining is a method of tracking feel of the civic
methods, tools, and algorithms for handling huge volume of about a particular item, company, events and issues. This
data to solve real world problems. The key objectives of the organization analyzes which part has opinion expressing, who
data mining process are to effectively handle large scale data, wrote the opinion and what is being commented online
mine actionable rules, patterns and gain insightful knowledge. reviews. There are three general categorizations for opinion
The explosion of social media has created extraordinary mining tasks: document-level, sentence-level, and phrase-level
opportunities for citizens to publicly voice their opinions. in Liu [1]. Document-level tasks are mainly formulated as
Because societal media is widely used for diverse purposes, classification problems where the input document should be
huge content of user created data exist and can be made an classified into a few predefined categories. In subjectivity
accessible for data mining. Recent researches in data mining classification, a document is classified as subjective or
focus on opining mining. objective. In sentiment classification, a subjective document is
classified as positive, negative, or neutral. Opinion helpfulness
prediction classifies an opinion as being helpful or not helpful.
Finally, opinion spam detection classifies opinions as spam and

978-1- 4788-7225 -8/15/$31.00©2015 IEEE 548

not spam. Sentence-level opinion mining is performed at the for sentiment classification. In topic-based classification, all
sentence level. In opinion search & retrieval and in opinion three classifiers have been reported to achieve accuracies of
question answering, sentences are usually retrieved and ranked 90% and above for particular categories.
based on some criteria. Opinion summarization aims to select a
set of sentences which summarizes the opinion more Turney [5] measured the co-occurrences between a word
accurately. Finally, opinion mining in comparative sentences and a set of manually selected positive words (e.g., good, nice,
includes identifying comparative sentences and extracting excellent and so on) and negative words ( e.g., bad, nasty, poor
information from them. Phrase-level opinion mining is also and so on ) using pointwise mutual information to compute
known as aspect based opinion mining. It performs finer-
the sentiment of a word.
grained analysis and directly looks at the opinion. The goal of
this level of analysis is to discover sentiments on aspects of
items. Aspects that are explicitly mentioned as nouns or noun In Kanimozhi Selvi et al [6] proposed an approach to
phrases in a sentence are called as explicit aspects. e.g., obtain the frequent itemsets involving rare items by setting the
‘resolution’ aspect in the review sentence “The resolution of support thresholds automatically.
this camera is nice”. Implicit Aspects are not explicitly
mentioned in a sentence but are implied, e.g., ‘price’ in the Kanayama et al [7] proposed an approach to build a
sentence “This camera is so expensive.” domain-oriented sentiment lexicon to identify the words that
express a particular sentiment in a given domain. By
Mining opinions at the document-level or sentence-level is construction, a domain specific lexicon considers sentiment
useful in many cases. However, these levels of information are orientation of words in a particular domain. Therefore, this
not sufficient for the process of valuable decision-making (e.g.
method cannot be readily applied to classify sentiment in a
whether to buy the product). For example, a positive review on
a particular item does not mean that the reviewer likes every different domain.
aspect of the item. Likewise, a negative review does not mean
that the reviewer dislikes everything. In a typical review, the Ding et al [8] focused on customer reviews of products. In
reviewer usually writes both positive and negative aspects of particular, the author reviewed the problem of determining the
the reviewed item, although his general opinion on the item semantic orientations (positive, negative or neutral) of
may be positive or negative. In fact, document-level and opinions expressed on product features in reviews. So, the
sentence-level opinions cannot provide detailed information for author proposed holistic approach that can accurately infer the
decision making. To obtain such information, a finer level of semantic orientation of an opinion word based on the review
granularity is needed. Hence, the proposed method focused on context. It provided a new function which is used to combine
aspect based opining mining in which concentrates on explicit multiple opinion words in the same sentence.
aspects. Section IV contains the proposed idea and techniques
used. Section V shows the experimental results. In Pang et al [9] focused on the methods that seek to
II.RELATED WORKS address the new challenges raised by sentiment aware
applications, as compared to those that are already present in
The automatic analysis of user generated contents such as more traditional fact based analysis. This paper includes a
reviews, online news, blogs and tweets can be extremely material on summarization of evaluative text and on broader
valuable for tasks such as mass opinion estimation, corporate issues regarding privacy, manipulation, and economic impact
reputation measurement, political orientation categorization, that the development of opinion oriented information access
stock market prediction, customer preference and public services gives rise to. To facilitate future work, a discussion of
opinion study. Liu et al [2] proposed a method to summarize benchmark datasets is also provided.
all the customer reviews of a product. It focused on mining
product features on reviews by user commented content. The Ramage et al [10] introduced Labeled LDA, a topic model
drawback is that there is no group features according to the that constraints Latent Dirichlet Allocation by defining a one–
strength of the opinions. to–one correspondence between LDA’s latent topics and user
tags. This allows Labeled LDA to directly learn word tag
The projected system focused an approach called Dynamic correspondences. Labeled LDA outperforms SVMs by more
Adaptive Support Apriori in Kanimozhi Selvi et al [3] to than 3 to 1 when extracting tag specific document snippets.
calculate the minimum support for mining class association
rules and to build a simple and accurate classifier. Zhang et al [11] focused on mining features. Double
propagation works well for medium-size corpora. However,
In sentiment classification, a classifier is trained using for large and small corpora, it can result in low precision and
labeled data, annotated from the domain in which it is applied. low recall. To deal with these two problems, two
Pang et al [4] examined whether it is sufficient to treat improvements based on part-whole and “no” patterns are
sentiment classification simply as a special case of topic-based introduced to increase the recall. It can rank feature candidates
categorization or whether special sentiment-categorization by feature importance which is determined by two factors:
methods need to be developed. This approach used three feature relevance and feature frequency.
standard algorithms: Naive Bayes classification, maximum
entropy classification, and support vector machines (SVMs)

549
 Daume et al [12] proposed a semi-supervised (labeled data Input: Online reviews
in source, and both labeled and unlabeled data in target) Output: aspects and sentiment orientation
extension to a well-known supervised domain adaptation Main procedure ()
approach. This semi-supervised approach to domain Data preprocessing ()
adaptation is extremely simple to implement, and can be Aspect Extraction ()
applied as a pre-processing step to any supervised learner. Sentence and Aspect Orientation ()
End
In Edison et al [13] focused on aspect based opinion Function Data preprocessing ()
mining in the proposed system. Tourism product reviews are Stop words removal
used as dataset in the system. Hotel and Restaurants corpus is Stemming
taken as dataset to mine reviews in aspect level. The task of Pos tagging
mining opinions and summarization is performed to provide End
customers a decomposed view of rated aspects. Function Aspect Extraction (pos tagged input
reviews)
III.PROBLEM DEFINITION if word is in noun then
The people cannot analyze exact information in the extract (word)
document and sentence level opinion mining on customer endif
reviews. Aspect level opinion mining is one of the solutions to count numbers of each word
problem. This gives fine detail information in aspect level. set a minimum support count
The goal of the task is to extract aspects on customer reviews. if aspect count < minimum support
Mining opinions on online customer reviews whether it is count
positive or negative opinion. The projected system identifies display (word)
the number of positive and negative opinions of each aspect in else
online reviews. remove (word)
endif
IV.PROPOSED SYSTEM End
The architectural overview for our working model of the Function Sentence and Aspect Orientation ()
proposed system is shown in figure 4.1. Identify opinions using Naive Bayesian
algorithm
End
Figure 2. Proposed Algorithm

C. Stop Word Removal

Most frequently used words in English are not useful in
text mining. Such words are called stop words. Stop words are
language specific functional words which carry no
information. It may be of types such as pronouns, prepositions,
conjunctions. Stop word removal is used to remove unwanted
words in each review sentence. Words like is, are, was etc.
Reviews are stored in text file which is given as input to stop
word removal. Stop words are collected and stored in a text
file. Stop word is removed by checking against stop words list.
D. Stemming
Figure 1. Working of Proposed System Architecture Stemming is used to form root word of a word. A
stemming algorithm reduces the words "longing", "longed",
The following section gives a detailed view of the and "longer" to the root word, "long". It consist many
proposed work. The proposed system uses customer reviews algorithms like n-gram analysis, Affix stemmers and
to extract aspect and mine whether given is positive or Lemmatization algorithms. Porter stemmer algorithm is used
negative opinion. Each review is split into individual to form root word for given input reviews and store it in text
sentences. A review sentence is given as input to data file.
preprocessing. Next, it extracts aspect in each review sentence.
E. POS Tagging
Stop word removal, stemming and pos tagging are data
preprocessing. Sentiment orientation is used to identify The Part-Of-Speech of a word is a linguistic category that is
whether it is positive or negative opinion sentence. Then it defined by its syntactic or morphological behavior. Common
POS categories in English grammar are: noun, verb, adjective,
identifies the number of positive and negative opinions of each
adverb, pronoun, preposition, conjunction, and interjection.
aspect.
POS tagging is the task of labeling (or tagging) each word in a

550
sentence with its appropriate part of speech. POS tagging is an Opinion word rule in figure 3 gives that, if word is
important phase of opinion mining, it is essential to determine matched with positive opinion words then positive count get
the features and opinion words from the reviews. POS tagging increment, or it is negative opinion word then negative count
can be done either manually or with the help of POS tagger get increment.
tool. POS tagging of the reviews by human is time consuming. In figure 3 Negation rules have a negation word or phrase
POS tagger is used to tag all the words of reviews. Stanford which usually reverses the opinion expressed in a sentence.
tagger is used to tag each word in an online review sentences. Two rules must be applied:
Every one sentence in customer reviews are tagged and stored 1. Negation Negative->Positive. This will increment
in text file.
positive count.
F. Aspect Extraction 2. Negation Positive ->Negative. This will increment
Frequent itemset mining is used to find all frequent item negative count.
sets using minimum support count. Here, every sentence is After comparing all the words of the sentence, the found
assigned as single transaction. Noun Words in each sentence is probabilities of the positive and negative counts are compared
assigned as item sets for single transactions. Aspect extraction in the following manners.
is implemented using figure 2. This algorithm first extracts a) If the probability of positive count is greater than the
noun and noun phrases in each review sentence and store it in negative count, then the sentence or opinion is positive.
a text file. Minimum support threshold is used to find all b) If the probability of negative count is greater than the
frequent aspects for a given review sentences. Aspects like positive count, then the sentence or opinion is negative.
pictures, battery, resolution, memory, lens etc. Then, the c) If the probability of positive count minus probability
frequent aspects are extracted and stored in text file. of negative count is zero, then it is neutral.
Finally system identifies the number of positive and
G. Sentence and Aspect Orientation negative opinion of each extracted aspect in customer reviews.
The proposed system first determines the number of
positive and negative opinion sentence in reviews using V.EXPERIMENTAL SETUP
opinion words. The positive and negative labels are collected The following section describes the dataset used in our
labels in opinion words. Examples of positive opinion words experiments and the results obtained.
are long, excellent and good and the negative opinion words
are like poor, bad etc. And the next step is to identify the H. Dataset Descriptions
number of positive and negative opinions of each extracted The proposed system uses customer review dataset about a
aspect. Both sentence and aspect orientations are implemented product effectively. A review is a subjective text containing a
using Naïve Bayesian algorithm using supervised term sequence of words describing opinions of reviewer regarding a
counting based approach. The probabilities of the positive and specific item. Review text may contain complete sentences,
negative count are found according to the words using Naïve short comments, or both. Product reviews are collected from
Bayesian classifier. websites like www.amazon.com, www.epinions.com and
Naïve Bayesian algorithm www.cnet.com. Each review in websites is assigned with a
Steps are as follows: different rating like 0-5 stars, a review label and date, a
1. The positive labels, negative labels and review reviewer name and location, a manufactured goods name, and
sentences are stored in separate text file. the review content. Canon camera product reviews are used in
2. Split the sentence into the combination of words. It the system. This dataset consists of product name and review
means first combination of two words and then single words. text. Reviews are split into individual sentences. The details of
3. First compare the combination of two words, if it the dataset used in the proposed system are shown in table 1 as
matched then delete that combination from the opinion. Again follows,
start comparing of single word. Table 1. Corpus Details
4. Initially, the probabilities of positive and negative Corpus Canon
count to zero [positive=0, negative=0]. The sentiment Camera
orientation algorithm is as follows: Reviews 100
Total Sentences 400
if word is in opinion_words then Positive Sentence 231
mark(word)
orientation ĸ$SSO\2SLQLRQ:RUG5XOH Negative Sentence 108
end if Total Opinion Sentences 339
if word is near a negation word then
Opinion sentences(Percentage) 84.75%
orientation ĸ$SSO\1HJDWLRQ5XOHV
end if I. Parameter For Evaluation
return orientation
The performance of the system is evaluated. Precision,
Figure 3. Sentiment Orientation Algorithm recall and F-measure are the parameters used in the system for
evaluation. Precision is the measure of retrieved instances that

551
are relevant. Recall is the fraction of relevant instances that are negative opinion of each extracted aspect. The number of
retrieved. F-measure is a measure of test’s accuracy. Precision, positive and negative opinions in review sentences is estimated.
recall and F-measure are defined as follows, Sentiment orientation gives a good accuracy. In future, it is
proposed to summarize the aspects based on the relative
(1) importance of the extracted aspect. By using this, it is possible
to analyze the customers interesting aspects on products.
(2)
ACKNOWLEDGMENT
(3) Our sincere thanks to the experts who supported and guided
us with their valuable domain knowledge.
To calculate these measures, true values in reviews are REFERENCES
identified manually. The proposed system mines the aspects
and opinion (extracted values). Using this, precision, recall [1] Bing Liu (2012), ‘Sentiment Analysis and Opinion Mining’, Synthesis
and F-measure are calculated for product customer reviews. Lectures on Human Language Technologies, Morgan & Claypool
Publishers.
J. Results [2] Hu, Minqing and Bing Liu (2004), ‘Mining opinion features in customer
reviews’, In Proceedings of the national conference on artificial
Aspect extraction gives accuracy of 80.36% using frequent intelligence, Vol.4, No.4, pp.755-760..
itemset mining. Sentiment orientation provides 92.37% of [3] Selvi, Kanimozhi, and A. Tamilarasi (2007), ‘Association rule mining
accuracy for given dataset. Precision, Recall and F-measure with dynamic adaptive support thresholds for associative classification’,
for aspect extraction and sentiment orientation are shown in In Conference on Computational Intelligence and Multimedia
figures 4&5 as follows, Applications, International Conference, vol. 2, pp. 76-80.
[4] Pang, Bo, Lillian Lee and Shivakumar Vaithyanathan (2002), ‘Thumbs
up?: sentiment classification using machine learning techniques’, In
Proceedings of the ACL-02 conference on Empirical methods in natural
language processing, Vol.10, pp. 79-86.
[5] Turney and Peter D (2002), ‘Thumbs up or thumbs down?: semantic
orientation applied to unsupervised classification of reviews’, In
Proceedings of the 40th annual meeting on association for computational
linguistics, pp. 417-424.
[6] Sadhasivam, Kanimozhi SC, and Tamilarasi Angamuthu (2011), Mining
Rare Itemset with Automated Support Thresholds’,Journal of Computer
Science 7, vol 3, pp. 394-399.
[7] Kanayama, Hiroshi and Tetsuya Nasukawa (2006), ‘Fully automatic
lexicon expansion for domain-oriented sentiment analysis’, In
Proceedings of the 2006 Conference on Empirical Methods in Natural
Figure 4. Parameters for Evaluation of Aspect Extraction Language Processing, Association for Computational Linguistics, pp.1-
9.
[8] Ding, Xiaowen, Bing Liu and Philip S. Yu (2008), ‘A holistic lexicon-
based approach to opinion mining’, In Proceedings of the 2008
International Conference on Web Search and Data Mining, Association
for Computing Machinery, pp. 231-240.
[9] Pang, Bo and Lillian Lee (2008), ‘Opinion Mining and Sentiment
Analysis’, Foundations and Trends in Information Retrieval, Vol. 2, No.
1/2,pp.1-135.
[10] Ramage, Daniel, David Hall, Ramesh Nallapati and Christopher D.
Manning (2009), ‘Labeled LDA: A supervised topic model for credit
attribution in multi-labeled corpora’, In Proceedings of the 2009
Conference on Empirical Methods in Natural Language Processing,
Association for Computational Linguistics,Vol.1, pp.248-256.
[11] Zhang, Lei, Bing Liu, Suk Hwan Lim and Eamonn O'Brien-Strain
(2010), ‘Extracting and ranking product features in opinion documents’,
In Proceedings of the 23rd international conference on computational
Figure 5. Parameters for Evaluation of Sentiment Orientation linguistics: Posters, Association for Computational Linguistics, pp.
1462-1470.
VI.CONCLUSION AND FUTURE WORK
[12] Daumé III, Hal, Abhishek Kumar and Avishek Saha (2010),
The proposed system extracts aspects in product customer ‘Frustratingly easy semi-supervised domain adaptation’, In Proceedings
reviews. The nouns and noun phrases are extracted from each of the 2010 Workshop on Domain Adaptation for Natural Language
Processing, Association for Computational Linguistics, pp.53-59.
review sentence. Minimum support threshold is used to find all
frequent aspects for the given review sentences. Naïve [13] Marrese-Taylor, Edison, Juan D. Velásquez and Felipe Bravo-Marquez
(2014), ‘A novel deterministic approach for aspect-based opinion mining
Bayesian algorithm using supervised term counting based in tourism products reviews’, Expert Systems with Applications, Vol.41,
approach is used to identify whether sentence is positive or No.17, pp.7764-7775.
negative opinion and also identifies the number of positive and [14] http://www.cs.uic.edu/~liub/

552