Sentiment Analysis Based Approaches For Understanding User Context in Web Content
Sentiment Analysis Based Approaches For Understanding User Context in Web Content
Abstract-- In our day to day lives, we highly value the understanding of human languages for machines’ use.
opinions of friends in making decisions about issues like Sentiment analysis extracts opinions, sentiments, and
which brand to buy or which movie to watch. With the emotions from text and analyses them.
increasing popularity of blogs, online reviews and social Sentiment classification can be done at three levels, at
networking sites, the current trend is to look up reviews,
the document level, at sentence level and at feature level.
expert opinions and discussions on the Web, so that one can
make an informed decision. Sentiment analysis, also known In document level classification, a document can be
as opinion mining is the computational study of opinions, classified into two classes, positive and negative based on
sentiments and emotions expressed in natural language for overall sentiment expressed by its writer [1].
the purpose of decision making. Sentiment analysis applies Classification can be done based on four pairs of human
natural language processing techniques and computational emotions, namely, “Joy ֞ Sadness”, “Acceptance ֞
linguistics to extract information about sentiments expressed Disgust”, “Anticipation ֞ Surprise” and “Fear ֞ Anger”
by authors and readers about a particular subject, thus as suggested in the Emotion Model proposed by Plutchik.
helping users in making sense of huge volume of
[2]
unstructured Web data. Applications like review
classification, product review mining and trend prediction
Sentence level sentiment analysis has two tasks-
benefit from sentiment analysis based techniques. This paper subjectivity classification and sentiment classification.
presents a study of different approaches in this field, the Information in a sentence can be of two types, objective
state of the art techniques and current research in Sentiment information and subjective information. Subjectivity
Analysis based approaches for understanding user’s context. classification involves identifying whether the sentence is
subjective or objective. Sentiment classification is further
Keywords—Sentiment Analysis, Computational Linguistics, classifying the subjective information as positive or
Natural Language Processing, Web Content Analysis negative. For example consider the following snippet of
text - “I bought an iPhone a few days ago. It is a great
phone.” The information in the first sentence is neutral,
I. INTRODUCTION
and hence it is objective whereas the second sentence
There are two types of textual information on the Web - speaks positively about the product, therefore it is
facts and opinions. Currently available search engines subjective.
search for facts, using machine readable information such It has been found that document level and sentence
as metadata and content within the page’s HTML tags like level classification are not enough to identify each and
title and headings. The same cannot be applied to every one detail about sentiments expressed in a
opinions. In today’s Web, a lot of opinionated text is document as sentiments may be expressed with respect to
available in various forms, for example, as reviews, blogs, different features. For example, a phone may have a rating
news articles, discussion groups and social networking of 4 out of 5 for speed, 2 out of 5 ease of use, 3 out of 5
sites. for battery, etc. Hence, it is necessary to extract certain
Sentiment analysis also known as opinion mining plays features from the input text.
a crucial role in determining the sentiments involved in Feature level classification comprises of three main
various Web content. Analysing opinions is very tasks. First step is to identify and extract the features. The
important for making decisions. For example, if one wants next step is to determine whether the opinions on the
to buy a new cell phone, a Web savvy buyer will almost features are positive, negative or neutral. Final task is to
always first peruse reviews about it in order to make an group the feature synonyms. A supervised learning
informed buying decision based on others experiences. algorithm can be used to build a graph of adjectives linked
Sentiment analysis is currently a very significant trend in by the same or different semantic orientation [3]. A
the area of natural language processing. Natural language conventional clustering algorithm can be used to partition
processing involves giving artificial intelligence to the adjectives into two subsets, one set containing positive
computers and is concerned with promoting an adjectives and the other containing the negative ones.
Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 18:46:48 UTC from IEEE Xplore. Restrictions apply.
This paper is organized as follows. In section II, we assigns numerical valence values and output sense value
introduce some existing approaches to sentiment analysis. for each sentence. The input paragraph is divided into a
In section III, we present a discussion on approaches set of sentences and each sentence is further divided into
available to perform sentiment classification. In section triplets. Valence values are assigned to the words in the
IV, we present certain applications that can employ triplet. These triplets are then processed to calculate the
sentiment analysis based approached. Section V discusses sentence level sentiment valence.
current trends and techniques in performing sentiment
analysis on Web data, followed by conclusion and III. CLASSIFICATION BASED APPROACH
references. Due to extensive popularity of the Web in the business
world, sentiment analysis on product reviews, articles
II. APPROACHES TO SENTIMENT ANALYSIS speaking about current market trends will be beneficial in
Sentiment analysis classifies the opinions into positive making business decisions. However, an overall view
and negative categories. Keke et al [4] focus on the about a document does not reveal the sentiments about all
technique to detect the topic related to the positive and aspects of a topic. For example, a person might be happy
negative opinions. Knowing the reasons behind with the camera, music, games in his cell phone but its
classifying the sentiment provides better perception. battery life may be a problem. Many sentiment analysis
These reasons are called as sentiment topics associated approaches have been proposed for analysing reviews,
with the sentiment. The proposed method collects web blogs, discussion forums, etc.
content and extracts snippets from them. Snippets are A document may contain information about several
keywords like brand names. Then a sentiment score is topics and sentiments about them. Mapping the sentiment
calculated for each snippet based on which they are to the correct topic is quite a challenge. The Sentiment
classified into different categories to create a sentiment Analyzer algorithm presented by Nasukawa et.al [8]
taxonomy. Topics related to each category are identified. addresses the above issues. It extracts the features related
Point wise mutual information and mutual support are to a topic, and then extracts sentiments of each sentiment
used to find words for a particular topic, to evaluate the bearing phase. It associates this topic, feature and
importance of a word in a category. Then, the word with sentiment to the document.
highest frequent value and highest point wise mutual An approach to classify news video stories and rank
information value is chosen as the topic. them has been presented by Chunxi et.al [9]. In their
Hasan et.al [5] propose an approach that is called approach, the stories were divided into two classes –
proximity based sentiment analysis. The proposed positive class and negative class. Fig. 1 shows an example
algorithm performs sentiment analysis of movie reviews to illustrate the working of their clustering algorithm. The
based on three proximity based features - proximity algorithm forms two clusters - one containing positive
distribution, shared information between proximity types, adjectives and other containing negative ones. A graph
and proximity patterns. Sentiment classification is based semi-supervised learning approach has been used
performed using an unsupervised approach, a mean & for this purpose. Similarity between words is calculated to
median approach (supervised approach) and a machine find the sentiment words. The selected sentiment words
learning approach. are used as features for classification. For the visual part,
Hogenboom et al. [6] proposed a method which an Affinity Propagation clustering approach is used to
considers the negation scope and strength of a word while determine the ranking of the videos. A linking matrix is
classifying whether a word has positive or negative effect used to check similarity between videos. Both text and
on the sentence. For example, let us consider two visual information are combined to rank the video.
sentences “I am happy with your performance” and “I am
not that happy with your performance”. The first sentence
expresses a positive emotion. If we just consider the
negative keyword “not” then the second sentence would
be equivalent to “I am not happy with your performance”
which is not correct. If scope and strength of the negative
keywords are considered while deciding its effect then it
would give better results. The proposed approach uses two
algorithms; the first one is used to calculate sentence score Fig. 1 A sample showing clusters formed using clustering algorithm
for each word. In the second algorithm, the sentence score
is calculated using the word sense and word score with Lin et.al [10] considered a reader’s view to form
respect to each negative keyword. If the calculated emotion categories as compared to conventionally
sentence score is less than zero, then it is assigned to a considering writer’s view for analysis. These categories
negative class. can be integrated with a web search engine to provide
Methods to analyse sentiment include machine results according to category needed by the user. In the
learning, statistical methods, building a knowledge base proposed method, they used Yahoo’s eight emotion model
and identifying keywords. To recognize effective - happy, sad, angry, surprising, boring, heart warming,
information from text, sentence level analysis is required. awesome, and useful. Features are analysed with respect
Shaikh et al. [7] developed a tool called SenseNet, that to these categories. A Support Vector Machine (SVM)
608
Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 18:46:48 UTC from IEEE Xplore. Restrictions apply.
was used as the classifier algorithm. The other models tries finds the hyper plane such that each training point is
used for comparison are Naïve Bayes classifier (NB), correctly classified and the hyper plane is as far as
passive-aggressive classifier (PA), bigram (BI), word possible from the points closest to it. It has a drawback of
(WD), metadata (MT), affix similarity (AS), word dimensionality.
emotion (WE) and Cui’s combined word n-grams (CN). 2) Naïve Bayes Method: Naïve Bayes is suitable
The highest accuracy was achieved when the models when the size of training data is less. It is a probabilistic
SVM, BI, WD, MT, AS and WE were used together. classifier based on applying Bayes theorem with strong
The accuracy achieved by using symbolic techniques independence assumptions. Its running time is linear in
used for classification can be improved by using the size of input. Maximum entropy classifiers do not
clustering based approach as suggested Li and Liu [11]. assume statistical independence of the independent
Symbolic techniques basically calculate sentiment score variables (as in Bayes classifier) that serve as predictors.
for a document by finding the average or sum which is not Learning in this model is slower than for a Naïve Bayes
that accurate. The suggested approach even reduces classifier, and thus may not be suitable if the number of
human effort unlike supervised learning approach where classes to learn is very large.
training the data is required. It uses the k-means clustering 3) Maximum Entropy Classifiers: Learning in a
algorithm to cluster the documents into positive and Naive Bayes classifier involves counting the number of
negative clusters. Raw data is given as input to TF-IDF co-occurrences of features and classes. In a maximum
(term frequency – inverse document frequency) [12] entropy classifier, the weights, which are typically
weighting method. This method calculates the importance maximized using Maximum-a-Posteriori (MAP)
of each word in the document. It considers two factors - estimation, must be learned. Unsupervised methods make
the number of times the word appears in the document use of PMI (Point wise mutual information) for co-
(term frequency) and the general importance of the word occurrence of a word with positive or negative word.
(inverse document frequency). Clustering is performed 20 Dictionary based methods can be used for expanding the
times and a voting mechanism is used to further improve set of positive and negative sentiment words and
the efficiency. If a document gets more than 10 positive classifying the words. But the drawback here is that a
votes then it is regarded as positive and vice-versa. Term word may be used positively in some domain and
scores from symbolic techniques are used to boost the negatively in some other domain.
performance. Further, in case of sentence level classification,
Feature level classification analyzes with respect to knowing the main aspect in the sentence is important
various features. It involves three steps - first the topic otherwise sentence level classification becomes futile.
related features are extracted, then sentiment bearing Another issue is that, it is not always a case that a
phrases are extracted. The main challenge here lies in the sentence will talk only about one topic or have only one
third step which involves associating a particular feature opinion. In such cases, feature level classification is better
with its sentiment. For the third step, Nasukawa et al.[8] than sentence level classification.
showed that analysis of grammatical sentence structures
and phrases give better results than statistical approaches. IV. SENTIMENT ANALYSIS APPLICATIONS
It is more complicated than document level classification. News articles written by different websites about a
In document level classification the document is analyzed topic may express different views. For certain topics,
with respect to a certain target. All the positive and some websites may hold a biased view due to which they
negative opinions about that particular topic are analyzed. may select and edit information according to their own
Zhang et al. [13] used a method where, based on keyword standpoints before delivering articles. Lacking conscious
entered by users, a sentiment graph of sentiment vectors awareness of websites’ sentiment bias may result in blind
of articles that keyword is plotted. The sentiment graph obedience to the reported information. With the evolution
gives an idea about inclination of articles towards various of internet as a huge platform as social media, the scope
sentiments. of biasness to manipulate the thinking of readers online
A. Machine Learning in Document Level Classification politically, economically, etc. has reached great levels.
To deal with certain controversial topics, or may be to
In document level classification, machine learning
divert the way the common man thinks about a certain
techniques can be employed to carry out sentiment
topic, the electronic media plays a influential role in
analysis. As presented in [17] supervised methods like
forming public opinion, which makes the analysis of the
support vector machine (SVM), Naïve Bayes, Maximum
document from a reader’s point of view very crucial.
entropy can be used.
Given a topic, Zhang et al [13] proposed a system that
1) Support Vector Machines: SVM is a
extracts relevant subtopics and presents sentiment
discriminative classifier and is better suited for classifying
difference between different subtopics. The system
reviews. To improve its quality large amount of training
analyses a given sentiment in four dimensions, which is
data is required. SVM is a non probabilistic binary
more similar to human emotion than conventional
classifier. It operates by separating points in a d-
positive-negative sentiment and detects sentiment bias. In
dimensional space using a (d-1)-dimensional hyper plane.
the system, articles are crawled and the part of speech
Given a set of training data, the SVM classifier finds a
tagging is done on them. Weight for each extracted word
hyper plane with the largest possible margin; that is, it
from article is calculated using (1).
609
Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 18:46:48 UTC from IEEE Xplore. Restrictions apply.
ሺ୵ǡ୧ሻ inbound links, etc. Some newspaper sites may provide
ሺǡ ሻ ൌ Ǥ (1)
ሺ୧ሻ ሺ୵ሻ good information about sports, some other sites about
politics, and so on. A site may be given more preference
where N(w, Pi) is the number of times that word w due to its age and genuineness of information that it
appears in article Pi, N(Pi) is the number of words provides.
extracted from Pi, N is the number of all collected news
articles, and N(w) is the number of articles in which word TABLE 1: ORIGINAL SENTIMENT WORDS FOR THE FOUR DIMENSIONS
w appears.
Dimensions Original sentiment words(e1 e2)
Then, a sentiment dictionary is constructed which (e)
contains a word and its sentiment value. Sentiment value Joy pleasure, be pleased, glad, happy, enjoy,
consists of scale value and weight value for four Sadness blessing, bless sad, feel sorry,
dimensions. Sentiment value is calculated using sadness, sorrow
probability functions for each article. For a particular year
Acceptance agreement, agree, acknowledge, consent,
(Y) edition for a particular newspaper, the number of Disgust acknowledgment, acceptance, accept
articles which include any word in the set e of original disgust, dislike, hate, be unpleasant,
sentiment words in Table 1 be df(Y, e), and the number of antipathy, evasion, evade
articles which include both target word w and any word in Anticipation expectation, expect, anticipation,
e be df (Y, e&w). Surprise anticipate, forecast surprise, be
surprised, astonishment, astonish,
ௗሺǡƬ௪ሻ admiration, admire
ܲሺܻǡ ݁Ƭݓሻ ൌ (2)
ௗሺǡሻ Fear fear, be scary, misgivings, have
Anger misgivings, be frightened anger, get
Next the interior division ratio is calculated using (3). angry, resentment, resent, rage, enrage
Positive sentiment words correspond to dimension e1 and
negative ones to dimension e2. It uses the Table 1 for V. CURRENT DEVELOPMENT IN SENTIMENT ANALYSIS
calculating joint probability function P(Y, e &w) in (2).
Here R e1֞ e2 (Y, w) = 0 if the denominator is 0. E-learning is related to distance learning, and is a
ሺǡଵƬ௪ሻ method of learning which can be through e-mail,
ܴ݁ͳ ֞ ݁ʹሺܻǡ ݓሻ ൌ (3)
ሺǡଵƬ௪ሻାሺǡଶƬ௪ሻ electronic forum, videoconferencing, chat rooms along
with other computer-based communication means. The
σ ୖୣଵ֞ୣଶሺଢ଼ǡ୵ሻ teachers and students do not need to be physically present
ሺሻ ൌ σೞ (4)
ೞ ୣଵ֞ୣଶሺଢ଼ǡ୵ሻ at the same place. There is no space and location
constraint, for example, in video conferencing people can
Scale value is calculated using (4). T e1֞ e2 (Y,w) is 0 if sit at home and attend the conference. For the system to
both df(Y, e1&w) and df(Y, e2&w) are 0, and T e1֞ e2 be successful many factors like ease of access to system,
(Y,w) is 1 otherwise. A word may appear in a number of interactivity, proper infrastructure play an important role.
editions and a number of times in various editions. To Analyzing e-learning blogs and reviews can help in
consider this, weight factor is calculated using (5). providing better services to the users and improve the
teaching -learning process.
ሺሻ ൌ ͳʹǤ ȭୣୢ୧୲୧୭୬ୱ ሺͳ ֜ ݈ʹሺǡ ሻ ൈ Kechaou et al. [14] proposed an approach to evaluate a
ͳͶͶǤ ȭୣୢ୧୲୧୭୬ୱ ሺǡ ͳƬݓሻ user’s opinion on e-learning systems. Three feature
݂݀ሺܻǡ ݁ʹƬݓሻ (5) selection methods MI (Mutual Information), IG
(Information Gain), and CHI statistics (CHI) have been
The sentiment vector O(P) of an article P has the form examined and advanced along with their proper HMM
(Oa(P), Ob(P), Oc(P), Od(P)). Considering P as a set of and SVM-based hybrid learning method. Their results
words extracted from it by the morphological analysis. A showed that IG (Information Gain) performed the best.
sentiment value Oe(P) of article P on dimension e is Applying data mining techniques on e-learning reviews
calculated as follows: and studying e-learning blogs are some of the challenges
faced in improving the accuracy of the proposed system
σೢאು ௌሺ௪ሻൈȁଶௌሺ௪ሻିଵȁൈெሺ௪ሻ further.
ܱሺሻ ൌ σೢאು ȁଶௌሺ௪ሻିଵȁൈெሺ௪ሻ
(6)
Sentiment analysis of twitter messages can help in
making financial, marketing, political decisions. People
The scale value Se(w) and weight Me(w) can be looked use tweets (short messages) to express their opinion about
up in the sentiment dictionary. Then based on search, a something. Jensen et al. [15] proposed a technique by
sentiment graph is generated which corresponds to which about 150,000 twitter messages were analyzed. The
sentiment vectors of articles containing topic and sub results obtained conveyed that 19% mentioned a brand
topic. When a user enters any keyword based on weight name, and 20% expressed sentiments about brands,
assigned to various words relevant sub keywords are among which about 50% spoke positively and 33% spoke
extracted. Weight for a word is calculated using (1). negatively. Celikyilmaz et al. [16] considered that twitter
Credibility information can be added to detect the messages are of two types - polar and non polar (neutral).
correctness of information like site age, number of They present a probabilistic model based sentiment
610
Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 18:46:48 UTC from IEEE Xplore. Restrictions apply.
analysis approach for twitter messages. Their technique International Conference on , vol.1, no., pp.265-271, 9-12 Dec.
2008
analyzes sentiments of polar text. As the twitter messages
[5] Hasan, S.M.S.; Adjeroh, D.A.; , "Proximity-based sentiment
are human generated, it is very difficult to interpret its analysis," Applications of Digital Information and Web
meaning correctly sometimes even by humans and there Technologies (ICADIWT), 2011 Fourth International Conference
may be a lot of noise in it, in the form of slang, shorthand on the , vol., no., pp.106-111, 4-6 Aug. 2011
etc. The method proposed first does text normalization [6] Hogenboom, A.; van Iterson, P.; Heerschop, B.; Frasincar, F.;
followed by pronunciation based clustering. For example, Kaymak, U.; , "Determining negation scope and strength in
sentiment analysis," Systems, Man, and Cybernetics (SMC), 2011
4get is same as forget. Then, polarity lexicon extraction is IEEE International Conference on , vol., no., pp.2589-2594, 9-12
done using a mixture model. The authors state that this Oct. 2011
analysis can be further improved by interpreting the [7] Mostafa Shaikh; Prendinger, H.; Ishizuka, M.; , "An analytical
similarity distance between words; for example, love, approach to assess sentiment of text," Computer and information
lovwww, loveee and luv as one entity ’love’. technology, 2007. iccit 2007. 10th international conference on ,
vol., no., pp.1-6, 27-29 Dec. 2007
VI. CONCLUSION AND FUTURE WORK [8] Yi, J.; Nasukawa, T.; Bunescu, R.; Niblack, W.; , "Sentiment
analyzer: extracting sentiments about a given topic using natural
Extensive research has been carried out in the field of language processing techniques," Data Mining, 2003. ICDM 2003.
sentiment analysis - text sentiment classifiers, effect Third IEEE International Conference on , vol., no., pp. 427- 434,
19-22 Nov. 2003
analysis, automatic survey analysis, opinion extraction, or
[9] Chunxi Liu; Li Su; Qingming Huang; Shuqiang Jiang; , "News
recommender systems. In this paper, we have presented video story sentiment classification and ranking," Multimedia and
different approaches available to analyze sentiment at Expo (ICME), 2011 IEEE International Conference on , vol., no.,
different levels. Based on the needs of the data to be pp.1-6, 11-15 July 2011
analyzed, a particular approach can be chosen. For [10] Lin, K.H.-Y.; Changhua Yang; Hsin-Hsi Chen; , "Emotion
example, to analyze reviews about a mobile, feature-level Classification of Online News Articles from the Reader's
Perspective," Web Intelligence and Intelligent Agent Technology,
sentiment analysis can be carried out. This will help in 2008. WI-IAT '08. IEEE/WIC/ACM International Conference on ,
knowing user’s opinion with respect to various features. If vol.1, no., pp.220-226, 9-12 Dec. 2008
the inclination of document is to be analyzed as a whole, [11] Gang Li; Fei Liu; , "A clustering-based approach on sentiment
to determine its polarity, then document level analysis," Intelligent Systems and Knowledge Engineering (ISKE),
2010 International Conference on , vol., no., pp.331-337, 15-16
classification is more suited. Nov. 2010
Currently, we are working towards applying Sentiment [12] G. Salton and C. Buckley, Term-weighting approaches in
Analysis based approaches for analyzing bias in online automatic text retrieval, Information Processing & Management,
content. The analysis of sentiments expressed by vol. 24, issue.5: 513–523, 1988.
information publishers is very useful for assisting users in [13] Zhang, J.; Kawai, Y.; Nakajima, S.; Matsumoto, Y.; Tanaka, K.; ,
judging the credibility of information. We plan to design "Sentiment Bias Detection in Support of News Credibility
Judgment," System Sciences (HICSS), 2011 44th Hawaii
and develop a system for detecting and visualizing International Conference on , vol., no., pp.1-10, 4-7 Jan. 2011
sentiment bias in online articles. The proposed system will [14] Kechaou, Z.; Ben Ammar, M.; Alimi, A.M.; , "Improving e-
be able to dynamically summarize the sentiment for learning with sentiment analysis of users' opinions," Global
different subtopics and for different websites. Any Engineering Education Conference (EDUCON), 2011 IEEE , vol.,
no., pp.1032-1038, 4-6 April 2011
background knowledge on a website can be used to assist
[15] B. J. Jensen, M. Zhang, K. Sobel, and A. Chowdury, “Twitter
users in determining news credibility. We plan to power: Tweets as electronic word of mouth,” Journal of the
construct a model which can automatically calculate American Society for Information Science and Technology, vol.
credibility scores for articles based on sentiment 60, no. 11, pp. 2169–2188, 2009
difference between subtopics and between websites. [16] Celikyilmaz, A.; Hakkani-Tu r, D.; Junlan Feng; , "Probabilistic
Since, sentiment with respect to a topic may also vary model-based sentiment analysis of twitter messages," Spoken
Language Technology Workshop (SLT), 2010 IEEE , vol., no.,
with time, the aspect of time will also be considered as a pp.79-84, 12-15 Dec. 2010
criterion for detecting bias in our proposed system. Our [17] Hajmohammadi, M., Ibrahim, R., Ali Othman, Z.. Opinion Mining
effort is towards detecting and visualizing any sentiment and Sentiment Analysis: A Survey. International Journal of
differences between different subtopics and between Computers & Technology, North America, 2, jun. 2012
different websites, and to make use of these to evaluate
published online information’s credibility.
REFERENCES
611
Authorized licensed use limited to: Trial User - Warsaw University (Uniwersytet Warszawski). Downloaded on March 17,2023 at 18:46:48 UTC from IEEE Xplore. Restrictions apply.