International Conference on Electrical, Electronics, and Optimization Techniques (ICEEOT) - 2016

A Brief Review on Sentiment Analysis

Neha Raghuvanshi Prof. J.M. Patil
PG Student, Dept of CSE, Assistant Professor, Dept of CSE,
SSGMCE, Shegaon, India SSGMCE, Shegaon, India
[email protected] [email protected]

Abstract— This paper presents a brief review on sentiment Twitter or article reviews on Digg. Another way of expressing
analysis. To mine the opinion on the web, it is essential to the popularity is to put a rating the popularity of the messages,
perform a well defined task, which helps us to retrieve the which can be related to the opinion expressed by the author.
information from the available data on the web. We have started Sentiment Analysis involves determining the
our discussion with the introduction on sentiment analysis, which evaluative nature of a piece of text. For example, a product
gives us a insight into sentiment analysis. The detail discussion on
review can express a positive, negative, or neutral sentiment
various methods proposed by different researchers is also
presented. Different types of sentiment analysis techniques give a (or polarity). Automatically identifying sentiment expressed in
research direction in different directions. Finally a method is text has a number of applications, including tracking sentiment
proposed based on the naïve bayes classifier. towards products, movies, politicians, etc., improving
customer relation models, detecting happiness and well-being,
and improving automatic dialogue systems. Over the past
Keywords—Sentiment analysis, Opinion based mining, Naïve decade, there has been a substantial growth in the use of micro
bayes classifier blogging services such as Twitter and access to mobile phones
world-wide. Thus, there is tremendous interest in sentiment
I. INTRODUCTION
analysis of short informal texts, such as tweets and SMS
Textual information on the internet is growing every day. To messages, across a variety of domains (e.g., commerce, health,
mine and search the textual information is becoming more military intelligence, and disaster management). Sentiment
difficult day by day. The text is prevalent data format on the analysis aims to uncover the attitude of the author on a
web, since it is easy to generate and publish. But extracting particular topic from the written text. Other terms used to
the information in the proper context from the vast ocean of denote this research area include “opinion mining” and
content is more difficult. The manual efforts are beyond the “subjectivity detection”. It uses natural language processing
human control as it takes more time. Therefore, the research and machine learning techniques to find statistical and/or
problem of automatic categorization and organizing data is linguistic patterns in the text that reveal attitudes. It has gained
more focused. Textual information is mainly divided into two popularity in recent years due to its immediate applicability in
main parts: facts and opinions. Here facts are only the business environment, such as summarizing feedback from the
subjective information, whereas the opinions represents the product reviews, discovering collaborative recommendations,
sentiments of their presenters. Initially, the research was in the or assisting in election campaigns.
direction of searching and categorizing the factual data. Now a Creating systems that can process subjective
days, web search engine retrieves the data from the web based information effectively requires overcoming a number of
on the keyword provided. For one keyword, it retrieves large novel challenges. To illustrate some of these challenges, let us
number of documents. For example, if we try to find the word consider the concrete example of what building an opinion- or
“nature” on the Google, google search engine searches review-search application could involve. As we have
different documents with mathing keywords. This time it will discussed, such an application would fill an important and
retrieve the documents with different meaning of keyword prevalent information need, whether one restricts attention to
provided to search. In recent years, we also contribute our blog search or considers the more general types of search that
opinion to large number of websites that expresses our views. have been described above. The development of a complete
Opinion can be expressed in different forms. One example review- or opinion-search application might involve attacking
may be web sites for reviewing products, such as Amazon, or each of the following problems.
movie review sites such as Rotten Tomatoes which enable (1) If the application is integrated into a general-purpose
rating of products, usually on some fixed scale as well as search engine, then one would need to determine whether the
leaving personal reviews. These reviews tend to be longer, user is in fact looking for subjective material. This may or
usually consisting of a few paragraphs of text. With respect to may not be a difficult problem in and of itself: perhaps queries
their length and comprehensiveness they tend to resemble blog of this type will tend to contain indicator terms like “review”,
messages. Other type of web sites contains prevalently short “reviews”, or “opinions”, or perhaps the application would
comments, like status messages on social networks like provide a “checkbox” to the user so that he or she could

978-1-4673-9939-5/16/$31.00 ©2016 IEEE

2827

Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 19:58:24 UTC from IEEE Xplore. Restrictions apply.
 indicate directly that reviews are what is desired; but in features such as bag-of-words model, using large movie
general, query classification is a difficult problem — indeed, it reviews corpus, restricting to adjectives and adverbs, handling
was the subject of the 2005 KDD Cup challenge. negations, bounding word frequencies by a threshold, and
(2) Besides the still-open problem of determining which using WordNet synonyms knowledge. The performance is
documents are topically relevant to an opinion-oriented query, evaluated on accuracy of four machine learning methods -
an additional challenge we face in our new setting is Naive Bayes, Decision Trees, Maximum-Entropy, and K-
simultaneously or subsequently determining which documents Means clustering. V.K. Singh, R. Piryani, A. Uddin [2]
or portions of documents contain review-like or opinionated presents a new Feature-based Heuristic for Aspect-level
material. Sometimes this is relatively easy, as in texts fetched Sentiment Classification. This paper presents an experimental
from review aggregation sites in which review-oriented work on a new kind of domain specific feature-based heuristic
information is presented in relatively stereotyped format: for aspect-level sentiment analysis of movie reviews. Authors
examples include Epinions.com and Amazon.com. However, have devised an aspect oriented scheme that analysis the
blogs also notoriously contain quite a bit of subjective content textual reviews of a movie and assign it a sentiment label on
and thus are another obvious place to look (and are more each aspect. The scores on each aspect from multiple reviews
relevant than shopping sites for queries that concern politics, arethen aggregated and a net sentiment profile of the movie is
people, or other non-products), but the desired material within generated on all parameters. The SentiWordNet is used which
blogs can vary quite widely in content, style, presentation, and is based scheme with two different linguistic feature selections
even level of grammaticality. comprising of adjectives, adverbs and verbs and n-gram
(3) Once one has target documents in hand, one is still faced feature extraction. They have also used our SentiWordNet
with the problem of identifying the overall sentiment scheme to compute the document-level sentiment for each
expressed by these documents and/or the specific opinions movie reviewed and compared the results with results
regarding particular features or aspects of the items or topics obtained using Alchemy API. The sentiment profile of a
in question, as necessary. Again, while some sites make this movie is also compared with the document-level sentiment
kind of extraction easier—for instance, user reviews posted to result. The results obtained show that our scheme produces a
Yahoo! Movies must specify grades for pre-defined sets of more accurate and focused sentiment profile than the simple
characteristics of films — more free-form text can be much document-level sentiment analysis.
harder for computers to analyze, and indeed can pose rn Schuller, Yunqing Xia,
additional challenges; for example, if quotations are included Catherine Havas [3] presents New Avenues in Opinion
in a newspaper article, care must be taken to attribute the Mining and Sentiment Analysis. The findings in this paper are
views expressed in each quotation to the correct entity. as follows. Gradually, sentiment analysis research is
(4) Finally, the system needs to present the sentiment distinguishing itself as a separate field, falling between NLP
information it has garnered in some reasonable summary and natural language understanding. Context-/intent-level
fashion. This can involve some or all of the following actions: analysis ensures the relevance of the opinions gathered. Social
(a) aggregation of “votes” that may be registered on different context will continue to gain importance, and an intelligent
scales (e.g., one reviewer uses a star system, but another uses system will have access to the comprehensive personal
letter grades) (b) selective highlighting of some opinions (c) information of vast numbers of people. Opinion mining will
representation of points of disagreement and points of be specific to each user’s or group of users’ preferences and
consensus (d) identification of communities of opinion holders needs. Opinions won’t be generic, but will reflect their source
(e) accounting for different levels of authority among opinion (for example, a relevant circle of friends or users with similar
holders. Note that it might be more appropriate to produce a interests, or the selection of a camera for trekking rather than
visualization of sentiment data rather than a textual summary for night shooting). Gilad Katz, Nir Ofek, Bracha Shapira [4]
of it, whereas textual summaries are what is usually created in presents Context-based sentiment analysis in 2015. A ConSent
standard topic based multi-document summarization. model, a novel context-based approach for the task of
sentiment analysis is presented. Proposed approach builds on
II. LITERATURE REVIEW techniques from the field of information retrieval to identify
key terms indicative of the existence of sentiment.. Whilst
The research community has studied almost all main aspects most researchers focus on assigning sentiments to documents,
of the problem. The most well studied sub problem is opinion others focus onmore specific tasks: finding the sentiments of
orientation classification (i.e., at the document level, sentence words (Hatzivassiloglou & McKeown 1997), subjective
level and feature level). In this section, we introduce different expressions (Wilson et al. 2005; Kim&Hovy 2004), subjective
methods of sentiment analysis. Kuat Yessenov [1] presents sentences (Pang&Lee 2004) and topics (Yi et al. 2003;
sentiment analysis of movie review comments. They have Nasukawa & Yi 2003; Hiroshi et al. 2004). These tasks
presented an empirical study of efficacy of machine learning analyze sentiment at a fine-grained level and can be used to
techniques in classifying text messages by semantic meaning. improve the effectiveness of a sentiment classification, as
Authors have use movie review comments from popular social shown in Pang & Lee (2004).
network Digg as the data set and classify text by Chen et al. [9] presented a visual analysis system
subjectivity/objectivity and negative/positive attitude. using multiple coordinated views, such as decision trees and
Different approaches have been proposed to extract text

2828

Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 19:58:24 UTC from IEEE Xplore. Restrictions apply.
 terminology variation, to help users to understand the weight and added up) and using this sum as exponent.
dynamics of conflicting opinions. Wanner et al. [10] described Decision tree is a tree in which internal nodes are represented
a concise visual encoding scheme to represent attributes, such by features, edges represent tests to be done at feature weights
as the emotional trend of each RSS news item. Both works for and leaf nodes represent categories which results from above
analyzing text contents are efficient by using word matching tests. It categorizes a document by starting at the tree root and
methods. However, they lack semantic analysis. Draper et al. moving successfully downward via the branches (whose
[11] developed an interactive visualization system to allow conditions are satisfied by the document) until a leaf node is
users to visually construct queries and view the results in real reached. The document is then classified in the category that
time. For sentiment mining and analysis, Gregory et al. [12] labels the leaf node. Decision Trees have been used in many
proposed a user-directed sentiment analysis method to applications in speech and language processing. Support
visualize affective document contents. Although they analyze Vector Machines has outperformed other classi- fiers such as
and visualize emotion, they only use statistical methods. To Nave Bayes. While SVM has become a dominant technique
demonstrate and predict the trend for an event, we suggest that for text classification, other algorithms such as Winnow and
rules about the evolution of public sentiments related to the AdaBoost have also been used in previous sentiment
participants about hot topic types should be modeled and classification studies. SVM gives highest accuracy results in
discovered. Collective behavior has many characteristics, such text classification problems.
as being spontaneous, zealous, unconventional, and transient.
Sentimental contagion and imitation are the main B. Unsupervised Learning Algorithms
psychological mechanisms of the collective behaviors. Hoyst These are also known as lexicon based techniques. This
et al. [13] and Sznajd-Weron et al. [14] proposed two different involve learning patterns in the input when no specific output
opinion dynamics models using the aforementioned theory. values are supplied, this means that the learner only receives
For example, when discussing a debatable topic on forums, an unlabelled set of examples. K-Means tries to find the
some participants’ sentiments can easily be affected by others, natural clusters in the data, by calculating the distance from
which might result in booing or other extreme actions. Several the centers of the clusters. The position of centers is iteratively
approaches focus on the visual exploration of blogs, forum changed until the distances between all the points are minimal.
posts, and Web logs. Adnan et al. [15] used frequent closed The centers are initially randomly assigned.
patterns to model and analyze data, and create a social
network. They also analyzed Web logs by integrating data C. Hybrid Techniques
mining and social network techniques [16]. Indratmo et al. In Hybrid Techniques both combination of machine learning
[17] visualized Web tags and comments arranged along a time and lexicon base approaches ate used. Researchers have
axis. Dork et al. [18] provided faceted visualization widgets proved that this combination gives improved performance of
for visual query formulation according to time, place, and classification. Mudinas et al. proposed a concept-level
tags. Ong et al. [19] proposed an interactive Web-based tree sentiment analysis system, called pSenti, which is developed
map, News map, to represent the relative number of articles by combining lexicon based and learning-based approaches.
per news item. Fisher et al. [20] found the evolution of topical The main advantage of their hybrid approach using a lexi-
trends in social media by using line graphs indicating term con/learning symbiosis is to obtain the best of both worlds-
trends. The aforementioned works focus on social networks, stability as well as readability from a carefully designed
text analysis and knowledge representation of social networks lexicon, and the high accuracy from a powerful supervised
to analyze microblog and forum content without sentiment learning algorithm.
analysis. Once the initial pre-processing is done, there are
number of approaches to perform the sentiment analysis. The
III. DIFFERENT METHODS IN SENTIMENT most common categories of review analysis are as follows:
ANALYSIS
A) Word Level Sentiment Analysis
Sentiment analysis mainly classified in three techniques
mainly supervised, unsupervised and hybrid. It is mostly used and effective sentiment analysis technique.
A. Supervised Learning Algorithms An effective encoding is done between the sentiment words
Some of the most predominant Supervised Learning and the class. Such as (Brillient, Awesom, Very Good) =>
techniques in Sentiment Analysis have been SVM, Nave Positive Sentiment
Bayesian Classifiers and other Decision Trees. A Nave Bayes There are number of databases that represents the
classifier is a simple probabilistic model based on the Bayes adjective and the respective class. This adjective extraction
rule along with a strong independence assumption. The Nave comes under the lexical analysis of the review. The class
Bayes model involves a simplifying conditional independence generation is a kind of clustering. In the general form, two
assumption. Maximum Entropy is conditional exponential classes are formed to identify the positive and the negative
classifier. It maps each pair of feature set and its label to a reviews. The reliability of this approach depends on the
vector. It is also called as loglinear classifier because they adjective or the sentiment word set. The word set must include
work by extracting some set of features from the input, all the synonyms and antonyms relative to the word. Here the
combining them linearly (each feature is multiplied by its

2829

Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 19:58:24 UTC from IEEE Xplore. Restrictions apply.
 synonym represents the positive sentiments and the antonym http://digg.com/movies/Star_Trek_The_best_prequel_ever : a
represents the negative sentiments. review of “Star Trek” movie (669 diggs.) etc. Based on the
B) Sentence Level Sentiment Analysis techniques, the proposed model is as follows.
In this sentiment analysis approach, different levels of 1. Number of documents are considered
granularity are analyzed over the review. A rule based analysis 2. The keyword extraction from the reviews
is required to perform the sentence based sentiment ¾ Elimination of the stop words eg. am,is are etc
identification. These rules include the negation rules ¾ Identification of the keywords
extraction approach. It means the sentence or the review 3. From these keywords, the adjectives and the other opinion
having the negative words such as no, not and never are used oriented keywords are identified
to represent the negative perspective of the sentiment. Some of 4. Opinion class is identified for each keyword
the verbs that shows the negative sense also represents the ¾ Number of opinion keywords in each class is
negative reviews such as "stop", "problem" etc. These verbs identified and based on which the weightage is
are also analyzed in different verb forms and the combination. assigned to each belonging opinion class
C) Feature Level Sentiment Analysis ¾ Opinion class that will have more number of
It is one of the most intelligent analysis over the movie keywords will be assigned by higher weightage
review. This analysis process defines the feature identification ¾ The aggregative weightage is applied on each opinion
from the review. This feature is compared from the review set class and the collective decision is taken.
and based on which the orientation score is identified. Each Naive Bayes is a simple but effective classification algorithm.
positive feature is assigned with positive weightage and the The Naive Bayes algorithm is widely used algorithm for
negative feature is assigned by the negative weight age. Once document classification
all the features are collected, the aggregation on the weightage
is performed to identify the overall featured score. If the score
is positive, the review is considered positive otherwise it is IV. CONCLUSION
considered negative. The feature analysis approach is based on
the statistical or the mathematical formula based on which This paper gives a brief insight into different methods
overall prediction of the sentiment feature will be done. available to mine the data on the web. This data is sometimes
D) Document-level sentiment analysis used as opinions, placed for different things. As the data on
Given a set of documents D, a sentiment analysis algorithm the web is increasing rapidly, it is better to process effectively.
classifies each document d from D into one of the two classes, We have proposed sentiment analysis model based on naïve
positive and negative. Positive label denotes that the document bayes model. This technique is proposed based on the study of
d expresses a positive opinion and negative label means that d different methods proposed by different researchers. Different
expresses a negative opinion of the user. The document-level types of sentiment analysis model based on word, sentence,
sentiment classification attempts to classify the entire feature and document is also discussed and found that
document (such as one review) into ‘positive’ or ‘negative’ document level model is more promising for better results.
class. The approaches based on SentiWordNet targets the term
profile of the review document and extract terms having
desired POS label (such as adjectives, adverbs or verbs). This References
clearly shows that before applying the SentiWordNet based
formulation; the review text should be applied to a POS tagger
[1] [1] Kuat Yessenov, Sasa Misailovic,“Sentiment Analysis of Movie
which tags each term occurring in the review text. Then some Review Comments”, 6.863 Spring 2009 final project, pp.1-17.
selected terms (with desired POS tag) are extracted and the [2] [2] Vivek Kumar Singh, Rajesh Piryani, A. Uddin, “Sentiment analysis
sentiment score of each extracted term is obtained from the of movie reviews: A new feature-based heuristic for aspect-level
SentiWordNet library. The scores for all extracted terms in a sentiment classification”, DOI: 10.1109/iMac4s.2013.6526500,
JANUARY 2013.
review are then aggregated using some weightage and
[3] [3] Erik Cambria, Bjorn Schuller, Yunqing Xia, Catherine Havasi, “New
aggregation scheme. Thus two key issues are to decide (a) Avenues in Opinion Mining and Sentiment Analysis”, IEEE Computer
which POS tags should be extracted, and (b) how to decide the Society, March/April 2013, pp. 15-21.
weightage of scores of different POS tags extracted while [4] [4] Gilad Katz, Nir Ofek, Bracha Shapira, “ConSent: Context-based
computing the aggregate score. sentiment analysis”, Knowledge-Based Systems 84 (2015) 162–178.
Different datasets which can be used to test the [5] [5] Hatzivassiloglou, V. & McKeown, K. R. (1997), “Predicting the
performance are 100 Hindi movies from the popular movie semantic orientation of adjectives”, In Proceedings of the 8th conference
on european chapter of the association for computational linguistics (pp.
review database website www.imdb.com, NLTK 0.9.9 174– 181). Madrid, Spain.
including its movie review corpus, machine learning facilities, [6] [6] Hiroshi, K., Tetsuya, N., & Hideo, W. (2004), “Deeper sentiment
and WordNet bindings, movie reviews of recent blockbuster analysis using machine translation technology”, In Proceedings of the
movies: 20th international conference on computational linguistics (COLING
2004), August 23 – 27, 2004 (pp. 494–500). Geneva, Switzerland.
http://digg.com/movies/Quantum_of_Solace_disappoints : a
[7] [7] Nasukawa, T. & Yi, J. (2003), “Sentiment analysis: capturing
review of new James Bond movie “Quantum of Solace” (684 favorability using natural language processing”, In Proceedings of the
diggs)and

2830

Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 19:58:24 UTC from IEEE Xplore. Restrictions apply.
 2nd international conference on Knowledge capture, October 23–25, [14] K. Sznajd-Weron, “Sznajd model and its applications,” Acta Phys.
2003. (pp. 70–77). Florida, USA. Polonica B, vol. 36, p. 2537, 2005.
[8] Turney, P. D. (2002), “Thumbs up or thumbs down? semantic [15] M. Adnan, R. Alhajj, and J. Rokne, “Identifying social communities by
orientation applied to unsupervised classification of reviews”, In frequent pattern mining,” in Proc. 13th Int. Conf. Inf. Vis., 2009, pp.
Proceedings of the 40th annual meeting of the Association for 413–418.
Computational Linguistics (ACL), July 6–12 , 2002 (pp. 417–424). [16] M. Adnan, M. Nagi, K. Kianmehr, M. Ridley, R. Alhajj, and J. Rokne,
Philadelphia, PA, USA. “Promotingwhere,when and what? An analysis ofweb logs by
[9] C. Chen, F. Ibekwe-SanJuan, and E. SanJuan, “Visual analysis of integrating data mining and social network techniques to guide
conflicting opinions,” in Proc. IEEE Comput. Soc. Symp. Visual eCommerce business promotions,” J. Social Netw. Anal. Mining, vol. 1,
Analytics, Chicago, IL, USA, 2006, pp. 59–66. no. 3, pp. 173–185, 2010.
[10] F. Wanner, C. Rohrdantz, F. Mansmann, D. Oelke, and D. Keim, [17] J. Vassileva and C. Gutwin, “Exploring blog archives with interactive
“Visual sentiment analysis of RSS news feeds featuring the US visualization,” in Proc. Conf. Adv. Vis. Interfaces, 2008, pp. 39–46.
presidential election in 2008,” presented at theWorkshop on Visual [18] M. Dork, S. Carpendale, C. Collins, and C. Williamson, “VisGets:
Interfaces to the Social and the Semantic Web, Sanibel Island, FL, USA, Coordinated visualizations for Web-based information exploration and
2008. discovery,” IEEE Trans. Vis. Comput. Graph., vol. 14, no. 6, pp. 1205–
[11] G. Draper and R. Riesenfeld, “Who votes for what? A visual query 1212, Nov./Dec. 2008.
language for opinion data,” IEEE Transa. Vis. Comput. Graphics, vol. [19] T. Ong, H. Chen,W. Sung, and B. Zhu, “Newsmap: A knowledge map
14, no. 6, pp. 1197–1204, Nov./Dec. 2008. for online news,” Decision Support Syst., vol. 39, no. 4, pp. 583–597,
[12] M. Gregory, N. Chinchor, P.Whitney, R. Carter, E. Hetzler, and A. 2005.
Turner, “User-directed sentiment analysis: Visualizing the affective [20] D. Fisher, A. Hoff, G. Robertson, and M. Hurst, “Narratives: A
content of documents,” in Proc.Workshop Sentiment Subjectivity Text, visualization to track narrative events as they develop,” in Proc. IEEE
2006, pp. 23– 30. Symp. Vis. Anal. Sci. Technol., 2008, pp. 115–122.
[13] J. Hoyst, K. Kaceperski, and F. Schweitzer, Annual Reviews of
Computational Physics IX. Singapore:World Scientific, 2001.

2831

Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 19:58:24 UTC from IEEE Xplore. Restrictions apply.