International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.

2, April 2015

A SURVEY OF SENTIMENT CLASSIFICATION

TECHNIQUES USED FOR INDIAN REGIONAL
LANGUAGES
Pooja Pandey and Sharvari Govilkar
Department of Computer Engineering, University of Mumbai, PIIT, New Panvel, India

ABSTRACT
Sentiment Analysis is a natural language processing task that extracts sentiment from various text forms
and classifies them according to positive, negative or neutral polarity. It analyzes emotions, feelings, and
the attitude of a speaker or a writer towards a context. This paper gives comparative study of various
sentiment classification techniques and also discusses in detail two main categories of sentiment
classification techniques these are machine based and lexicon based. The paper also presents challenges
associated with sentiment analysis along with lexical resources available.

KEYWORDS
NLP, sentiment, sentiment analysis, classification techniques, challenges, lexical resources, features,
machine learning, lexicon based.

1.INTRODUCTION
Sentiment Analysis (SA) is a natural language processing task that deals with finding orientation
of opinion in a piece of text with respect to a topic [1]. It deals with analyzing emotions, feelings,
and the attitude of a speaker or a writer from a given piece of text. Sentiment Analysis involves
capturing of users behaviour, likes and dislikes of an individual from the text. The main goal
behind sentiment analysis is to identify sentiment associated with the text by extracting
sentimental context from the text.
The purpose of sentiment analysis is to determine the attitude or inclination of a communicator
through the contextual polarity of their speaking or writing. Their attitude may be reflected in
their own judgment, emotional state of the subject, or the state of any emotional communication
they are using to affect a reader or listener. It is trying to determine a persons state of mind on
the subject they are communicating about. This information can be mined from various data
sources like: texts, tweets, blogs, social media, news articles, product comments.
There are different classification levels in SA: document-level, sentence-level and aspect-level.
Document-level SA aims to classify an opinion of the whole document as expressing a positive or
negative sentiment. Sentence-level SA aims to classify sentiment expressed in each sentence
which involves identifying whether sentence is subjective or objective. Aspect-level SA aims to
classify the sentiment with respect to the specific aspects of entities which is done by identifying
the entities and their aspects for instance researchers need a tool to generate summaries for
deciding whether to read the entire document or not and for summarizing information searched by
user on internet. News groups can use multi document summarization to cluster the information
from different media and summa.
DOI:10.5121/ijcsa.2015.5202

13

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

The paper presents a detail survey of various sentiment classification techniques. Related work
done and past literature is discussed in section 2.Baseline algorithm is defined in section 3 along
with challenges associated in performing sentiment analysis. Two main categories of sentiment
classification techniques which are machine based SA and lexicon based SA are discussed in
detail in section 4 along with the comparison of each method cited in Indian regional languages.
Finally, section 5 concludes the paper.

2.LITERATURE SURVEY
In this section we cite the relevant past literature of research work done in the field of sentiment
analysis for Indian languages.
Amitava Das and Bandopadhya developed SentiWordNet for Bengali language, which is an
automatically constructed lexical resource in which WordNet synset are assigned positive and a
negative score. SentiWordNet and Subjectivity Word List are used to generate merged sentiment
lexicon in which duplicate words are removed. Bengali SentiWordNet is created by applying
word level lexical-transfer, using an English-Bengali dictionary, on content available in
SentiWordNet. [4]
Authors have [1] developed a modified approach to identify the sentiments associated with Hindi
content by handling negation and discourse relation. They updated the existing Hindi
SentiWordNet (HSWN) by when specific sentiment words where not found in existing HSWN by
extracting same meaning word from English SentiWordNet. Through handling of negation and
discourse associated with text, their proposed algorithm achieved approximately 80% accuracy on
classification of Hindi reviews.
Amandeep Kaur and Vishal Gupta Proposed Algorithm for Sentiment Analysis for Punjabi Text.
Their used the Hindi WordNet to develop the Subjective Lexicon for the Punjabi language.[8]
They are using three popular methods used for the generation of subjective lexicon-Use of BiLingual Dictionary, Machine Translation, Use of Word net .Then they devise an Algorithm
Combining the unigram method and simple scoring method which provides better efficiency. The
overall efficiency of the proposed algorithm is 54.2%.
Aditya Joshi and Pushpak Bhattacharyya [2] proposed a fallback strategy for finding sentiment
associated with Hindi language. Machine Translation, In-language and Resource Based SA where
the three approach proposed for SA in Hindi. Through WordNet linking, words present in English
SentiWordNet were replaced by similar Hindi words to construct Hindi SentiWordNet
(HSWN).To determine the polarity of the opinion associated in text SVM classifier was used to
perform In-language SA. In a machine translation based method Google translator was used to
translate Hindi corpus into English and resulted corpus was used as input to the classifier to
determine polarity. In resource based SA approach the synset corresponds to the English
SentiWordNet is used in the corresponding synset in Hindi to build the SentiWordNet (H-SWN)
for Hindi. 78.14 was the best accuracy achieved through in-language sentiment analysis for Hindi
documents.
Kishorjit and Sivaji Bandyopadhyay proposed verb based Manipuri sentiment analysis. They are
using the conditional random field (CRF) approach. It is an unsupervised approach where the
system learns by giving some training and can be used for testing other texts. Then they processed
text for part of speech tagging using CRF. Here With the help of POS tagger the verbs of each
sentence are identified and the modified lexicon of verbs is used to notify the polarity of the
sentiment in the sentence, because the sentiment of the sentence is highly dependent on the verbs.
Their proposed algorithm achieved approximately 75% accuracy on sentiment analysis of
Manipuri. [9]
14

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

Authors have used self learning neural network which takes linguistic and part of speech emotive
features as input, for detection of sentiment associated in Tamil content. The primary inputs to the
neural network are ; first the outputs of domain classifier which retrieves noun and verbs and term
domain frequencies, second the two schmaltzy analyzers which are Negation Scorer and Flow
scorer which assigns sore to each document based on the pleasantness of the words in it ,and and
lastly noun, verb and urichol the three taggers . Unsupervised learning was selected and Hebbian
learning was incorporated since emotion recognition has to deal with emotional intelligence. Two
dimensional animated face generator is used to show resultant emotion which is identified by
assigning weights for features based on their affective influence [7].
Das and Bandopadhya [5] used different strategies to predict the sentiment of a word in the given
text. In one of the strategies they annotated the words with their associated polarity manually. In
another strategy, to determine the polarity of the text, Bi-Lingual dictionary for English and
Indian Languages was used. In next strategy synonym and antonym relations of words in
WordNet are used, to determine the polarity. Final strategy used learning from pre-annotated
corpora to find the polarity of the text.
Authors have used lexicon based approach for extracting sentiment from Urdu text. Sentimentannotated lexicon based approach for analyzing sentiment based on SentiUnits. SentiUnits are the
expressions made of one or more words, which carry the sentiment information of the whole
sentence. Shallow parsing is used for identification and extraction of SentiUnits from the given
text. Two types of SentiUnits were used i.e. Single adjective phrase and multiple adjective
phrases. Adjectives, modifiers and orientation are the three attributes associated with SentiUnits.
Process of sentiment analysis is composed of three phases; pre-processing where normalization
and segmentation of text was performed then next phase involves use of shallow parsing for
extracting SentiUnits and finally extracted SentiUnits are compared with lexicon and their
polarities are calculated for classification as positive or negative and overall polarity is obtained
by combining polaritys.[6]
Authors have built a robust sense based classifier [3] which is a supervised document level
sentiment classifier on basis of semantic space based on WordNet senses. Through the use of
similarity metrics unknown synset in the training set was replace by similar synset in test set.
Words in the corpus were annotated with their senses using combinations of manual sense
annotation and automatic iterative WSD. In synset replacement algorithm a synset encountered in
a test document is not found in the training corpus, it is replaced by one of the synsets present in
the training corpus. The substitute synset is determined on the basis of its similarity with the
synset in the test document. The synset that is replaced is referred to as an unseen synset as it is
not known to the trained model.
Authors have proposes Cross-Lingual Sentiment Analysis(CLSA) [10] using WordNet senses as
features for supervised sentiment classification where machine translation is not available for
translation between specific languages .Concept of linked WordNet is used which bridge the gap
between those two languages. WordNet of Hindi and Marathi was developed using an expansion
approach having same synset identifier The words in the corresponding synsets represent
translations of each other in specific contexts. Words of the training as well as the test corpus
were mapped to their WordNet synset identifiers. A classification model is learnt on the training
corpus and tested on the test corpus leads to a new corpora which is represented in the common
feature space i.e. sense space. Accuracy of 72% and 84% for Hindi and Marathi was obtained for
sentiment classification of these languages.

15

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

3.SENTIMENT ANALYSIS
Sentiment analysis is computational study of emotions, opinions and mainly the sentiment
expressed in the text by user. Sentiment analysis is a challenging task due to many challenges
which are associated while processing natural language. Any sentiment analysis system needs
first to extract feature i.e. sentimental words or phrases from the given text and then using
suitable text classifier overall sentiment associated with the text is extracted.

3.1.Challenges for sentiment analysis

3.1.1.Contextual Information
Identifying the context of the text becomes an important challenge to address in SA.
Behaviour/use of the word changes in a great aspect based on the context.
Ex-1 The journey was long.
Ex-2 Seminar was long.
Ex-3 Battery life of Nexus 5 is long.
In all the above 3 examples, meaning of long is same it indicates the duration or passage of time.
In Ex-1 and Ex-2 long indicates bored hence a Negative expression whereas in Ex 3 long
indicates efficiency hence a Positive expression. In Ex 3 long indicates efficiency hence a
Positive expression.
3.1.2.Sarcasm Detection
Sarcasm involves statement and a remark which is usually indirect taunt towards any object or an
appraisal in a negative way. Detecting sarcasm is a tough task for humans and equally harder for
machine. Some examples of sarcasm: Ex- Amazing presentation by Mr. X, I wont ever attend
such presentation again.
3.1.3.Word Sense Disambiguation
Word sense disambiguation (WSD) is the problem of determining in which sense a word having a
number of distinct senses is used in a given sentence. The same word can have multiple
meanings, and based on the sense of its usage the polarity of the word also changes. [19] For
example, the word "cold" has several senses and may refer to a disease, a temperature sensation,
or an environmental condition. The specific sense intended is determined by the textual context in
which an instance of the ambiguous word appears. In "I am taking aspirin for my cold" the
disease sense is intended, in "Let's go inside, I'm cold" the temperature sensation sense is meant,
while "It's cold today, only 2 degrees", implies the environmental condition sense.
3.1.4.Word Order:
Word order plays a vital role in deciding the polarity of a text, in the text same set of words with
slight variations and changes in the word order affect the polarity aspect. For example X is
efficient than Y conveys the exact opposite sentiment from Y is efficient than X.
3.1.5.Identify subjective portions of text:
The same word can be treated as subjective in one context and objective in some other. This
makes it difficult to identify the subjective (sentiment-bearing) portions of text. Consider
following examples:
Ex-1 the language of the author was very crude.
Ex-2 Crude oil is extracted from the sea beds.
16

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

3.1.6.Indirect negation of sentiment

Negation of sentiment [17] is assigned to words like no, not, never, etc. But there are certain
words tend to reverse the sentiment polarity implicitly. For example the sentence, This movie
avoids all predictable and boring drama found in most of the bollywood movies. The negative
sentiment associated with words predictable and boring is reversed by associating word avoid
with those words.
3.1.7.Entity Recognition
Same entity is not seen for all the texts in a document. When multiple entities are being
mentioned about in a single document, the overall document polarity does not make much sense.
We need to separate out the text about a particular entity and then analyze its sentiment. Ex- I
hate Heat, but I like ICE.

3.2.Features for sentiment analysis

Converting a piece of text to a feature vector is the basic step in any data driven approach to SA.
Some commonly used features used in Sentiment Analysis are:
3.2.1.Term Presence vs. Term Frequency:
Term frequency is mostly used in different Information Retrieval (IR) and Text Classification
tasks. But Pang-Lee [11] found that term presence is more important to Sentiment analysis than
term frequency as for term presence binary-valued feature vectors are used, and those vectors the
entries merely indicate whether a term occurs indicated as value 1 or do not occur indicated by
value 0. Polarity based classification is very useful for SA ,as overall sentiment may not usually
be highlighted through repeated use of the same term an also occurrences of rare word can have
more sentimental value compared to frequently repeated word.
3.2.2.Term Position:
More sentimental value is given to certain words based on their position in a sentence or a
document. Generally words appearing in the 1st few sentences and last few sentences in a text
are given more weightage than those appearing elsewhere in the document.
3.2.3.Parts of speech (POS):
Part-of-speech (POS) information is commonly exploited in sentiment analysis which deals with
finding adjectives, adverbs in the text as they are important indicators of sentiment in a given text.
Amongst all parts of speech like nouns, verbs, adjectives, etc most important POS is considered
as adjective to represent sentimental feature or simply subjective text. Adverbs when occurs along
adjectives improves the probability of finding exact sentiment of a given text.
3.2.4.Topic-Oriented Features:
Interactions between topic and sentiment play an important role in extracting sentiment from the
content. For example, in a hypothetical article on Rebook, the sentences Reebok reports that
profits rose and Target reports that profits rose could indicate completely different types of
news regarding the subject of the document. Topic information can also be incorporated into
features set as they also help in specifying sentiment.
17

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

3.2.5.Opinion words and phrases:

There are words like; good or bad, like or hate, happy or sad which are mostly used to express
opinions towards an object. Also sometimes there are phrases which expresses opinions without
using an of the opinion words in the phrase. For example: It cost me my full month salary.

3.3.Lexicon resources for sentiment lexicon

Lexicon resources provide [17] a polarity value of associated with words. Lexical resources can
be developed either manually by humans or automatically by training machines. Manually built
lexical resources tend to be more accurate but time consuming task, automatically built resources
can attain much higher coverage in less time. WordNet is a lexical hierarchical database with
nodes represented by word meaning instead of word itself, and relationships between different
synonym synsets representing the edges between the nodes. Multiple words can be assigned to
single node and a same word might be present in more than one node at a time.
3.3.1.Manually built Lexical Resources
General Inquirer consists of 11000+ words compiled from two different sources with roughly
2000 positive and 2000 negative words. Each word is tagged either as positive, negative or none.
The precision with which Inquirer tags a word is very high; because of its low coverage the recall
rate is very low. Bing- Liu Lexicon consists of around 2000 positive and 4000 negative words
which are frequently seen in social media. Bing-Liu lexicon has advantage over General Inquirer
in terms of coverage as it contains more sentiment-bearing words but has a low precision as less
overall words. Subjectivity Lexicon is developed from multiple sources and sentiment which are
manually tagged and contains 8000+ sentiment words, each tagged as either positive or negative.
Problem with manually-tagged sentiment lexicons is that there is no concept of like mildly
positive or strongly negative to classify in deep.
3.3.2.SentiWordNet
SentiWordNet is a lexical resource [17] which is an extension of WordNet. It appends sentiment
information to every synset present in the WordNet. Positive, Negative and Objective score are
tagged to each synset where there combined score sum up to 1.0 overall for a synset. Here
synsets sentiment are not fixed to a single category i.e. the same synset has non-zero score for
both the classes as it is positive in particular context and negative in another context.
3.3.3. SentiWS
SentimentWortschatz or SentiWS [15] is a German-language resource for sentiment analysis. It
contains list of positive and negative sentimental words which are assigned weighted between
[1; 1] and also part of speech tag and their inflections are associated. Currently SentiWS
contains 1,650 negative and 1,818 positive words in the list. The list contains adjectives and
adverbs which explicitly expresses a sentiment with addition of nouns and verbs.
3.3.4.WordNet Affect
WordNet Affect is a lexical resource [12] that represents the affective content of synsets by
dividing them into affective categories. Thus, it gives more affective information as compared to
SentiWordNet and is used when analysis to be done is with respect to emotions like sad, anger,
happy, joy and others.
18

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

3.4. Baseline Algorithm

Baseline algorithm for sentiment analysis consists of tokenization, feature extraction and
sentiment classification [18] using classifier such as SVM, Naive Base, etc.

Figure 1. Baseline Algorithm

We first have to tokenize the sentence which involves breaking a stream of text up into words,
phrases, symbols, or other meaningful elements called tokens. It is done by segmenting text by
splitting it by spaces and punctuation marks, and forms a bag of words. Care should be taken so
that short forms such as dont, Ill, shed will remain as one word.
Then important features are identified. Features such as Terms presence and frequency, Parts of
speech (POS), Opinion words and phrases, Negations are identified. We need to take care of
negations, since they will reverse polarities and decide whether we want to use only adjective,
adjectives plus adverbs or simply all the words as features. Lexicon-based or statistical feature
selection methods can be used to select features from documents which treat document as Bag of
Words (BOW) or string. Stemming and removal of stop-words are mostly common feature
selection step.
After we've tokenized and decided which features to use we need to classify the sentiment. Is it
good or bad? Classification can be done with different algorithms. For example: Nave Bayes,
Support Vector Machines, or Max Entropy. Lexical resources like dictionary, WordNet,
SentiWordNet are uses by these classifier algorithms. This technique attempts to determine
whether a text is objective or subjective and whether a subjective text contains positive or
negative sentiments. The system automatically collects, cluster, categorizes, and summarizes
news from several sites on the web on a daily basis. A summarization machine can be viewed as a
system which accepts either a single document or multiple documents or a query as an input and
produces an abstract or extract summary.

4.SENTIMENT CLASSIFICATION TECHNIQUES

Sentiment classification is a task under Sentiment Analysis (SA) that tags text as positive,
negative or neutral automatically. Thus, a sentiment classifier tags the sentence the movie is
entertaining and totally worth your money! in a movie review as positive with respect to the
movie. On the other hand, a sentence The movie is so boring that I was dozing away through the
second half. is labelled as negative. Finally, The movie is directed by Nolan is labelled as
neutral [2]. There are two main techniques for sentiment classification: machine learning based
and lexicon based. Better performance can be obtained by combining these two methods.
19

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

4.1.Machine Learning Approach

The machine learning method uses several learning algorithms to determine the sentiment by
training on a known dataset. Training and a test set are the two document sets which are mostly
needed for machine learning based techniques. Training set is used by classifier to understand the
different characteristics associated with documents, and to check the overall performance of the
classifier test set are used.
4.1.1.Supervised learning

The supervised learning methods depend on the existence of labelled training documents.
Supervised learning process: two Steps; Learning (training): Learn a model using the
training data testing: Test the model using unseen test data to assess the model accuracy.
There are different types of supervised classifiers like: Rule-based Classifiers, Decision
Tree Classifiers, Linear Classifiers and Probabilistic Classifiers.
a. Probabilistic Classifier
Mixture models for classification are used by Probabilistic classifiers where it assumes that each
class of the content is a component of the particular mixture. Each mixture component can be
referred as generative model which provides the probability of sampling a particular term for that
component.
1.Naive Bayes Classifier (NB)
Naive Bayes classification model computes the posterior probability of a class is computed in
Naive Bayes Classifier which is based on the way words are distributed in the particular
document. The positions of the word in the document are not considered for classification in this
model as it uses BOWs feature extraction technique. Bayes Theorem is used to predict the
probability where given feature set belongs to a particular label of the content.
P(labelfeatures)=(P(label)*P(features|label))/(P(features))
P (label) signifies the prior probability of a label. P (features | label) signifies the prior probability
that a particular feature set is being classified as a label. P (features) specifies the prior probability
that a given feature set has occurred in the process. On basis of Naive assumption i.e. all features
are independent; the equation can be rewritten as:
P(labelfeatures)=(P(label)*P(f1label)*.*P(fnlabel))/(P(features))
2.Bayesian Network (BN)
Bayesian Network model is a form of directed acyclic graph in which nodes represent

random variables and edges represent conditional dependencies in the graph. Complete
joint probability distribution (JPD) is specified for the model as it is reckoned as complete
model for the variables and their relationships.BN is not frequently used for text mining as
computation complexity is very expensive.

20

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

3.Maximum Entropy Classifier (ME)

The MaxentClassifier also known as a conditional exponential classifier uses encoding to create
encoded vectors by converting labelled feature sets. This encoded vector are used to calculate
weights associated with each feature which can then be combined to determine the most likely
label for a feature set. Main parameters for this classifier is a set of X {weights}, which is used to
merge the joint features which are generated from a feature-set by an X {encoding}. Every C
{(featureset, label)} pair is mapped to a vector through encoding.
B.Linear classifiers
Probability of a particular classification is based on a linear combination of features and their
weights. A linear classifier determines which class an object belongs by making a classification
decision based on the value of a linear combination of the characteristics of the objects. Feature
vector are used to represent object characteristics. Different linear classifiers are available
1.Support Vector Machines Classifiers (SVM)
The main principle of SVMs is to determine linear separators in the search space which can best
separate the different classes. Due to sparse nature of text the text data are well suited for SVM
classification. In text data some of the features are irrelevant, but most of them seem to be
correlated with one another and can be easily organized into linearly separable categories. The
basic SVM has no dependency on probability as it predicts which two possible class forms an
output for a given input.
2.Neural Network (NN)
Neural Network simply consists of neurons which are arranged in layers and convert a given
input vector into meaningful output. Each neuron processes an input by applying nonlinear
function to it and the output is passed to the next layer for further process. Mostly neural
networks are designed as feed-forward network .Signals passing from one neuron to another are
assigned different weights and during the training phase these weights are adjusted so that neural
network adapts to a particular problem to solve. Multi-layer neural networks are used for nonlinear boundaries. These multiple layers are used to create multiple piecewise linear boundaries,
which are mainly used to approximate enclosed regions which belongs to a particular class of
output. Here training process is more difficult as there is a need of back-propagation of errors
over several layers.
C.Decision tree classifier
In decision tree classifier training data set ate hierarchical decomposed where to divide the data
the condition on the attribute value is used. The condition or predicate of the attribute signifies the
presence or absence of one or more words associated with it. Data space is divide recursively
until certain minimum numbers of records which are used for the purpose of classification are
available at leaf node.
D.Rule based classifiers
The data space is modeled with a set of different rules in rule based classifiers where the left side
signifies a condition on the feature set expressed in disjunctive normal form while the right hand
side represents the class label. During training phase all the rules are constructed based on
different criteria used to generate rules. Support and confidence are commonly used criteria
21

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

where support signifies absolute number of instances present in the training data set which are
relevant to the rule specified and Confidence refers to the conditional probability that the right
hand side of the rule is satisfied when the left-hand side is satisfied for a given input.
4.1.2.Unsupervised learning
Unsupervised learning is deals with finding hidden structure in unlabeled data set. There is no
error or reward signal to evaluate a potential solution as examples given to the learner are
unlabeled. Unsupervised learning methods are useful when there are documents to classify which
are unlabeled. Nearest neighbor (KNN) is unsupervised machine learning algorithm in which
objects are classified based on the majority of its nearest neighbor of the object. The class which
is assigned to the object is based among its most k nearest neighbours object. Objects are
classified based on their similarities to objects in the training data in this algorithm. Selection
process is based on either majority voting or distance weighted voting.

4.2.Lexicon-based approach
The lexicon-based approach involves calculating sentiment polarity for a review using the
semantic orientation of words or sentences in the review. The semantic orientation is a measure of
subjectivity and opinion in text. Sentiment lexicon contains lists of words and expressions used to
express peoples subjective feelings and opinions. For example, start with positive and negative
word lexicons, analyze the document for which sentiment need to find. Then if the document has
more positive word lexicons, it is positive, otherwise it is negative. The lexicon based techniques
to Sentiment analysis is unsupervised learning [13] because it does not require prior training in
order to classify the data.
Manual construction, corpus-based methods and dictionary-based methods are the methods
through which sentiment lexicon are constructed. The manual construction of sentiment lexicon is
a difficult as it involves humans to manually assign polarities to sentimental words and its a
time-consuming task. Dictionary based method is an iterative technique which is initially
constructed manually by selecting small set of sentimental word and this set then iteratively
grows by adding the synonyms and antonyms from the WordNet. This iterative process continues
till no new words are reaming to be added to the seed list. The dictionary based approach have a
limitation is that it cant find opinion words with domain specific orientations. Corpus based
techniques rely on syntactic patterns in large corpora. Corpus-based methods can produce opinion
words with relatively high accuracy. Most of these corpus based methods need very large labelled
training data but it helps to easily find domain specific opinion words and orientations of this
words towards a context.
Following tables presents a comparison of sentiment classification techniques cited in Indian
regional languages.

22

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015
Table 1. Sentiment Analysis techniques used in Hindi text.

Type
Corpus

Hindi [2]
250 Hindi Movie Reviews and English movie reviews
In-Language using SVM

Machine
Resource-based using
Translation
SentiWord list
Based using
SVM

Classificat
ion
Technique

Features

Term
Frequency

Term
presenc
e

TF-IDF

TF-IDF

Most Common All sense

Sense

74.57%

72.57%

78.14%

65.96%

56.35%

60.31%

Accuracy

Benefit

MT-based systems give superior classification performance as compared to

majority-based systems based on lexical resources.

Limit

The error of the machine translation system affects the performance of MT-based
SA.

Table 2. Sentiment Analysis techniques used in Hindi text.

Hindi

Type
Corpus
Classificatio
n Technique

[1]
Hindi Movie Reviews
Semantic approach using HindiSentiWordNet (HSWN)

Features

Improved
HSWN

Improved
negation

Accuracy

69.78%

78.39%

Benefit
Limit

HSWN

+ Improved
Discosure

HSWN

negation+

80.21%

Increases the coverage of HindiSentiWordNet (HSWN)

Low accuracy for words which have dual nature.

23

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015
Table 3. Sentiment Analysis techniques used in Manipuri text.

Type
Corpus

Manipuri [9]
Manipuri news paper, total 2,75000 words

Classification Conditional Random Field (CRF).

Technique
Features

Part Of Speech(POS)

Accuracy

Recall:72.10, Precision:78.14, F-Score:75.0

Benefit
Limit

Model is easy to interpret

More methods and algorithms are to be search and implemented in order to
improve the accuracy
Table 4. Sentiment Analysis techniques used in Punjabi text.

Type
Corpus

Punjabi [8]
Documents written in Punjabi language

Classification Subjective lexicon

Technique
Features

Part Of Speech(POS)

Accuracy

Recall:70, Precision:78,F-Score:67

Benefit
Limit

Better accuracy
Performance is low.
Lexicon developed for Hindi language has limited coverage.

Table 5. Sentiment Analysis techniques used in Tamil text.

Type
Corpus

Tamil [7]
Tamil News Text

Classification Neural network

Technique
Features

Part Of Speech(POS)

Accuracy

Precision : 60

Benefit
Limit

The emotion is identified by assigning weights for features based on their

affective influence.
Local minima and over fitting

24

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015
Table 6. Sentiment Analysis techniques used in Urdu text.

Type
Corpus

Urdu[6]
Movie

Product

Classification Lexicon based using SentiUnits

Technique
Features

Part Of Speech(POS)

Accuracy

72%

Benefit
Limit

78%

Achieving better results by using SentiUnits.

The SentiUnits with adjectives made by postposition combined with nouns, cause
errors and hence, an improved algorithm is required.

5.CONCLUSIONS
Sentiment analysis has lead to determine the attitude or inclination of a communicator through the
contextual polarity of their speaking or writing. Sentiments can be mined from texts, tweets,
blogs, social media, news articles, comments or from any source of information.
Sentiment Analysis has been quite popular and has lead to building of better products,
understanding users opinion, executing and managing of business decisions. People rely and
make decisions based on reviews and opinions. This research area has provided more importance
to the mass opinion instead of word-of-mouth.
Large amount of work in sentiment analysis has been done in English language, as English is a
global language, but there is a need to perform sentiment analysis in other languages also. Large
amount of other languages contents are available on the Web which needs to be mined to
determine the sentiment.

ACKNOWLEDGEMENTS
I am using this opportunity to express my gratitude to thank all the people who contributed in
some way to the work described in this paper. My deepest thanks to my project guide for giving
timely inputs and giving me intellectual freedom of work. I express my thanks to the head of
computer department and to the principal of Pillai Institute of Information Technology, New
Panvel for extending his support.

25

International Journal on Computational Sciences & Applications (IJCSA) Vol.5, No.2,April 2015

REFERENCES
[1] Mittal, Namita, et al. "Sentiment Analysis of Hindi Review based on Negation and Discourse
Relation," 11th Workshop on Asian Language Resources (ALR), In Conjunction with IJCNLP. 2013.
[2] Joshi, Aditya, A. R. Balamurali, and Pushpak Bhattacharyya, "A fall-back strategy for sentiment
analysis in Hindi: a case study," Proceedings of the 8th ICON (2010).
[3] Balamurali, A. R., Aditya Joshi, and Pushpak Bhattacharyya, Robust sense-based sentiment
classification," Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and
Sentiment Analysis, Association for Computational Linguistics, 2011.
[4] Das, Amitava, and Sivaji Bandyopadhyay. "SentiWordNet for Bangla," Knowledge Sharing Event-4:
Task 2 (2010).
[5] Das, Amitava, and Sivaji Bandyopadhyay, "SentiWordNet for Indian languages," The 8th Workshop
on Asian Language Resources. 2010.
[6] Syed, Afraz Z., Muhammad Aslam, and Ana Maria Martinez-Enriquez. "Lexicon based sentiment
analysis of Urdu text using SentiUnits," Advances in Artificial Intelligence, Springer Berlin
Heidelberg, 2010.
[7] Giruba Beulah and Madhan Karky, On Emotion Detection from Tamil Text.
[8] Kaur, Amandeep, and Vishal Gupta. ,"Proposed Algorithm of Sentiment Analysis for Punjabi Text,"
Journal of Emerging Technologies in Web Intelligence 6.2 (2014): 180-183.
[9] Nongmeikapam, Kishorjit, et al. "Verb Based Manipuri Sentiment Aanalysis,"2014.
[10] Balamurali, A. R. "Cross-lingual sentiment analysis for Indian languages using linked wordnets."
(2012).
[11] Pang, Bo, and Lillian Lee. "Opinion mining and sentiment analysis," Foundations and trends in
information retrieval 2.1-2 (2008): 1-135.
[12] Strapparava, Carlo, Alessandro Valitutti, and Oliviero Stock. "The affective weight of lexicon,"
Proceedings of the Fifth International Conference on Language Resources and Evaluation. 2006.
[13] VOHRA, MRSM, and JB TERAIYA, "A Comparative Study of Sentiment Analysis Techniques,"
JIKRCE, 2013.
[14] Medhat, Walaa, Ahmed Hassan, and Hoda Korashy. "Sentiment analysis algorithms and applications:
A survey." Ain Shams Engineering Journal (2014).
[15] Remus, Robert, Uwe Quasthoff, and Gerhard Heyer, "SentiWS-A Publicly Available Germanlanguage Resource for Sentiment Analysis," LREC, 2010.
[16] Esuli, Andrea, and Fabrizio Sebastiani, "SentiWordNet: A publicly available lexical resource for
opinion mining," Proceedings of LREC, 2006.
[17] http://www.cfilt.iitb.ac.in/resources/surveys/SA-Literature%20Survey-2012-Akshat.pdf.
[18] https://class.coursera.org/nlp/lecture/145
[19] http://wsd.nlm.nih.gov
Authors
Pooja Pandey is currently a graduate student pursuing masters in Computer
Engineering at PIIT, New Panvel, and University of Mumbai, India. She has received
her B.E in Computer Engineering from University of RGPV. She is having 2 years of
past experience in the field of teaching. Her areas of interest are Natural Language
processing, theory of computation and ethical hacking.
Sharvari Govilkar is Associate professor in Computer Engineering Department, at
PIIT, New Panvel, and University of Mumbai, India. She has received her M.E in
Computer Engineering from University of Mumbai. Currently she is pursuing her
PhD in Information Technology from University of Mumbai. She is having 17 years
of experience in teaching. Her areas of interest are text mining, Natural language
processing, Compiler Design & Information retrieval etc.

26