A Subjective Feature Extraction For Sentiment Analysis In Malayalam Language

Jisha P. Jayan, Deepu S. Nair, Elizabeth Sherly
Research Cell : An International Journal of Engineering Sciences, Special Issue March 2015, Vol. 14
ISSN: 2229-6913 (Print), ISSN: 2320-0332 (Online) -, Web Presence: http://www.ijoes.vidyapublications.com
© 2014 Vidya Publications. Authors are responsible for any plagiarism issues.
1
A SUBJECTIVE FEATURE EXTRACTION FOR SENTIMENT ANALYSIS
IN MALAYALAM LANGUAGE
Jisha P. Jayan1
, Deepu S. Nair2
, Elizabeth Sherly3
*1
Virtual Resource Center for Language Computing(VRCLC), Indian Institute of Information Technology and Management-
Kerala , Thiruvananthapuram
jisha.jayan@iiitmk.ac.in1 ,
deepu.s@iiitmk.ac.in2 ,
sherly@iiitmk.ac.in3
Abstract: In recent days, Sentiment Analysis has become an active research in NLP, which analyzes people's opinions, sentiments, evaluations,
attitudes, and emotions from writing language. The growing importance of sentiment analysis coincides with the growth of social media such as
reviews, forum discussions, blogs, and social network. In his paper, sentiment analysis of Malayalam film review is carried out using machine
learning techniques CRF combined with a rule based approach. The system shows 82 % accuracy.
INTRODUCTION
Sentiment Analysis (SA) is a process that helps to extract the
subjective or conceptual information from various sources. It
deals with analyzing emotions, feelings and the attitude of a
speaker or a writer from a given piece of text. In a broader
sense, SA is a cognitive process which helps computer to
understand and extract human behavior such as likes and
dislikes, feelings and emotions and many other attributes and
also to predict the behavioral aspects of human. It is also used
for opinion mining, one of the hottest topics in NLP that helps
to identify and extract subjective information in source
materials and provides valuable insights about the user’s
intentions, taste and likeliness etc.
Facts and opinions are two main types of textual information
in the world. Facts are objective expressions about entities,
events and their properties while opinions are usually
subjective expressions that describe people’s sentiments,
feelings toward entities, events and their properties. Opinion
expressions convey people’s positive or negative sentiments
and it may be a neutral comment. Present research on textual
information processing has been focused on mining and
retrieval of factual information.
The web has dramatically changed the way that people express
their views and opinions. Common men can now post reviews
of products at various online product review sites and express
their views on almost anything in Internet forums, discussion
groups, social media and blogs, which are collectively called
the user-generated content. This would lead to measurable
sources of information, which helps to improve the quality of
the product and for a better feedback and choice to the user.
Such system can be further modified to an automatic textual
analysis for sentiments, automatic survey analysis, opinion
extraction, or a recommender system. Such system typically
tries to extract the overall sentiment revealed in a sentence or
document, either positive or negative, or neutral.
Malayalam belongs to the Dravidian family, a large family of
languages of South and Central India, and SriLanka.
Malayalam exhibits heavy amount of agglutination. Due to the
agglutination and rich morphology of words along with high
ambiguity of Malayalam language, research in NLP for
Malayalam is always challenging and is same for sentiment
analysis because of its high dependence on words that are used
for expressing the feelings or other sentiments.
The sentence-level and document-level review has been
considered. Focus on the sentence-level sentiment extraction is
significant because in most of the websites, user comments are
just a single sentence. Document -level provides the semantics
of the entire document, but often fails to detect sentiment
about individual aspects of the topic. A statistical approach
using simple co-occurrence that commonly used machine
learning techniques is a trivial approach, but fail to provide a
better result, especially in cases where both negative and
positive comes in two differ sentences in a document. In order
to resolve such shortcoming, we propose a hybrid statistical
model using rule based and extracting the grammatical
features.
The paper is organized into different sections. First section
dealt with the introduction about SA and the objective of the
paper. The second section exposed the states of the art that
provides some of the major work carried out in this area. The
third section reveals the proposed work and the
methodologies. The fourth section includes the
implementation and the result obtained. The fifth section
concludes the paper.
STATES OF THE ART - SENTIMENT ANALYSIS
There has been a wide range of work carried out on this topic.
The main research carried out in the area of sentiment analysis
is in the document and sentence level. Document and sentence
level classification methods are usually based on the
classification of review context or words. Most of the work
done is by using either of these three methods, Semantic
Orientation method, Machine Learning method or Rule Based
approach.
One of the first attempts in this field was done by Alekh
Agarwal and Pushpak Bhattacharyya [1] for English. In this
paper they made an attempt to determine the overall polarity
of a document, such as identifying for the appreciation or
criticism of a movie. They presented machine learning based
approach to solve the problem of determining the sentiments
similar to text categorization. The movie review was selected
for their experiments. Their paper concluded with an accuracy
of over 90% for the first time.
Another work on the sentiment extraction of movie was done
by Pang [2]. The ultimate aim of that work was to find the best
way to classify the sentiment from text, either standard
machine learning techniques or human-produced baseline.
Three different machine learning techniques explained were
mainly Maximum Entropy, Support Vector Machine, and
Naive Bayes. In their experiment, they tried different
variations of n-gram approach like unigrams presence,

2
unigrams with frequency, unigrams with bigrams, bigrams,
unigrams with POS, adjectives, most frequent unigrams,
unigrams with positions They concluded that machine learning
techniques are quite good in comparison to the human
generated baseline. The paper also remarked that the Naïve
Bayes approach tend to do the worst while SVM performs the
best.
Manurung, and Ruli [3] work was carried out in 2008 for
Indonesian Language using machine learning method. In this
work, he initially translated English movie review into
Indonesian language and then applied to the machine learning
approach such as Naive Bayes, SVM, and maximum Entropy
method to perform the sentiment classification. He reached at
the conclusion that SVM is the best classification method
giving 80.09% accuracy.
Saggion and Funk [4] used senti-wordnet to perform opinion
classification. They calculated positive and negative score for
a review and based on the maximum score, the polarity of the
review was assigned. They also extracted features and used
machine learning algorithms to perform classification of the
sentiments from the text.
Turney [5] also worked on part of speech (POS) information.
He used tag patterns with a window of maximum three words
using trigrams. In his experiment, he considered JJ, RB, NN,
NNS POS-tags with some set of rules for classification of
product reviews. He used adjectives and adverbs for
performing opinion classification on reviews. PMI-IR
algorithm is used to estimate the semantic orientation of the
sentiment phrase. He achieved an average accuracy of 74% on
410 reviews of different domains collected from opinion.
Barbosa [6] designed a 2-step automatic sentiment analysis
method for classifying tweets. They used a noisy training set
to reduce the labelling effort in developing classifiers. First,
they classified tweets into subjective and objective tweets.
Then subjective tweets are classified as positive and negative
tweets. Celikyilmaz [7] design a pronunciation based word
clustering method for tweet normalization. In pronunciation
based word clustering, words having similar pronunciation are
clustered and assigned common tokens. They also used text
processing techniques like assigning similar tokens for
numbers, html links, user identifiers, and target organization
names for normalization. After doing normalization, they used
probabilistic models to identify polarity lexicons. They
performed classification using the BoosTexter classifier with
these polarity lexicons as features and obtained a reduced error
rate.
In Malayalam, the works on sentiment analysis is in its infant
stage. Geethu Mohandas [8], had proposed a semantic
orientation method for extraction of the mood from any
sentence. In their study, they have applied the semantic
orientation method using an unsupervised learning technology
for classifying the input text for classification. They used a tag
set which includes the tags sorrow, joy, anger and neutral for
tagging the manually created corpus and then calculated the
semantic orientation by semantic association using SO-PMI
(Semantic Orientation from Point wise Mutual Information).
They concluded their paper with a conclusion that the SO-PMI
method gives about 63% accuracy.
PROPOSED WORK
The proposed work concentrates on sentiment analysis to find
the positive, negative or neutral opinions from the user’s
writings at the document level. The polarity of sentence and
rating of individual category, such as film, direction, acting,
song, script etc. is individually computed. The different
suggestions, opinion and feedback about the film by
considering different factors improve the overall ranking in a
more meaningful manner and also item wise scoring. This
work has been implemented on a hybrid approach combining
the machine learning technique with rules. Since Malayalam is
a highly agglutinative language with rich morphology, and
also of free order, it has a wide range of fluctuated words with
the same meaning. Also such reviews, there are a number of
colloquial usages, short forms and broken sentences. Here we
used Conditional Random Field (CRF) techniques for proper
extraction and classification of sentiments.
Conditional Random Fields (CRFs) is a probabilistic
framework for labeling and segmenting structured data, such
as sequences, trees and lattices. The underlying idea is that of
defining a conditional probability distribution over label
sequences, given a particular observation sequence, rather than
a joint distribution over both label and observation sequences.
The primary advantage of CRFs over Hidden Markov Models
is their conditional nature, resulting in the relaxation of the
independence assumptions required by HMMs in order to
ensure tractable inference. Additionally, CRFs avoid the label
bias problem, a weakness exhibited by Maximum Entropy
Markov models (MEMMs) and other conditional Markov
models based on directed graphical models.
IMPLEMENTATION
A document level feature extraction and analysis is carried out
for finding the polarity and rating of the film reviews. The
polarity indicates positiveness, negativeness or neutrality of
the document and the rating gives the rate in each category
separately. The categories mainly dealt with our song, acting,
direction, script and film. POS tagging is performed to the
sentence, but for many of the attributes in sentiments requires
additional tagsets, that is being included in the proposed
method for better analysis and prediction. The training and
testing process using CRF is depicted in Figure 1and 2.
Figure1: Training Phase Figure 2: Testing Phase
Corpus Collection and
refinement
Learning using CRF
Tagset Definition
Tokenization
Manual Tagging
Input Text
Polarity Analysis
Tokenization
Tagging using CRF
Implementing Rules
Rating/Score Analysis

3
TAGSET DEFINITION
The additional tagset definition for sentiments is an important
task in this work. There is no exact and standard tagset for the
sentiments in Malayalam presently, without that proper
tagging is difficult. We have defined additional 10 tags for
sentiment analysis in our study. Tagging not only depends on
word but also the context of that particular document. The
same words have the different tags in different contexts.
Table I. Sentimental Tags
MACHINE LEARNING USING CRF
The collected and refined corpus has been tagged manually for
training the engine. Here the engine has the capability of
recalling the previous experience, there by learns for better
classification. About 30000 tokens and its tags were used for
training. The rules were also implemented appropriately with
respects to the various semantics.
Algorithm
Step 1: Take Input
Step 2: Classification Using CRF (Tagging)
Step 3: Analyze the tagged output
Step 4: Apply 9 rules for finding the polarity of the sentence or
document (Positive, Negative, Neutral)
Step 5: Find the rating of the individual category (Excellent,
good, not bad, bad, not good, worst)
Step 6: Results
Step 7: Exit
RESULT
The system has been analyzing the sentiments from
Malayalam film review at document level. Find out the
polarity and rating of individual category. The categories
which are Film, Direction/Script, Song/Acting, and other
factors and attained an overall accuracy of 82 %.
Eg: Input text: കുറ്റകൃത്യവും ദുരൂഹത്യും അന്വേഷണവും
സിവിമയില്‍
‍ മാവയമായി പറയാവറിയാവന്ന ആളാണ് ത്ാനവന്ന്
ആദ്യചിത്രമായ ‘ ഡിറ്റക്ടീവി ’ലുനെയും വാലാമനെ ചിത്രമായ
‘ നമമറീസി ’ലൂനെയും നത്ളിയിച്ച സുംവിധായകവാണ് ജീത്തു
ന്ജാസഫ് . എന്നാല്‍
‍ മലയാളികള്ക് അത്ര പരിചിത്മല്ലാെ
കുടുംബത്രില്ലര്‍
‍ ഗണെിലാണ് പുത്ിയ ചിത്രമായ ‘ ദൃശയും ’ ജീത്തു
ഒരുകിയിരിക്കുന്നത്് . മലന്യാരഗ്രാമെിനല സാധാരണ
കുടുംബെിലുണ്ടാകുന്ന ഗൗരവകരമായ പ്രത്ിസന്ധി ത്ന്ത്രപരമായി
കകകാരയും നചയ്യുന്നനത്ങ്ങനവനയന്ന് ത്രില്ലെിപ്പിക്കുും വിധും
പറഞ്ഞാണ് ‘ ദൃശയ ’നെ സുംവിധായകവ്‍
‍ സമ്പന്നമാക്കുന്നത്് .
കൂട്ടിവ് ന്മാഹലവ്‍
ലാലിനവ അഭിവയവഴകവും. ഇവ രണ്ടുമാകുന്മ്പാള്‍
‍
‘ ദൃശയും ’ ദൃശയാനുഭവമാകുന്നു . മൂഡ് അനുസരിച്ചുള്ള ഗാവങ്ങളും
പശ്ചാെലസുംഗീത്വും ചിത്രനെ ഉന്േഷമുള്ളത്ാക്കുന്നു .
ന്മാഹവ്‍
ലാലിനെ അവായാസമായ അഭിവയ മികവ് ത്നന്നയാണ്
കര്‍
മ്മന്യാദ്ധായനെ സവിന്ശഷത്നയന്ന് വിസ്സുംശയും പറയാും .
ഡയറക്ഷനെ ന്പാരായ്മ ചിത്രനെ ശരിക്കുും ബാധിച്ചു .
ഓര്‍
നെടെ് പറയാവന്ന സന്ദര്‍
ഭങ്ങളും സുംഭാഷണങ്ങളും
ചിലതുണ്ട് ചിത്രെില്‍
‍ . ന്മാഹവ്‍
ലാലിനെ അവായാസമായ
അഭിവയ മികവ് ത്നന്നയാണ് കര്‍
മ്മന്യാദ്ധായനെ
സവിന്ശഷത്നയന്ന് വിസ്സുംശയും പറയാും . അത്യാവശയെിവ്
ന്പ്രക്ഷക നവറുപ്പ് ന്വൊവ് ആ വില്ലവ് കഥാപാത്രെിവ് കഴിഞ്ഞു
എങ്കിൽ അത്് മുരളി ശര്‍
മ്മയനെ വിജയും . ജവപ്രിയമായ ഒരു
നെലിവിഷവ്‍
‍ന്കാമഡി ന്ഷായിനല കഥാപാത്രങ്ങനളനയാനക ത്നെ
സിവിമയില്‍
‍ഉള്‍
നപ്പടെിയിട്ടുണ്ട് സത്യവ്‍
അന്തികാെ് . സിവിമയനെ
ഒഴുകിനവ വല്ലാനത് ബാധിക്കുന്നുണ്ട് ഈ മാധയമങ്ങളനെ ഇെനപെല്‍
‍.
വിരസത് കൂൊനത് സിവിമ ത്ീര്‍
കാവ് സിദ്ധിഖിവ് കഴിഞ്ഞു എങ്കിലുും
രണ്ടാും പകുത്ിയില് ഒനട്ടാന്ന് ദ്ിശാന്ബാധും വഷ്‍
െനപ്പട്ടുന്വാ എന്ന്
ന്ത്ാന്നിപ്പിക്കുന്നുണ്ട്‍
‍ഈ നജെില്‍
മാവ്‍
‍. ആദ്യപകുത്ിയനെ ആന്വശും
ഇെയിനലവിനെന്യാ നകട്ടുന്പാവന്നു . രസകരമായ കുന്റന്യനറ
വിമിഷങ്ങളും ഹൃദ്യസ്പര്‍
ശിയായ സുംഭാഷണ ശകലങ്ങളും
അസോദ്യകരമായ വര്‍
മ്മങ്ങളും ഇമ്പമാര്‍
ന്ന ഗാവങ്ങളമായി
സിദ്ധിഖിനെ സ്ത്രീകളും മാവയവായ മനുഷയനുും വിഷുകാല
ആന്ഘാഷെിവ് ത്ിെന്മ്പറ്റും . സിവിമയനെ മിഴിവിനെ മികവിനു
നത്ളിവായി സത്ീഷ് കുറുപ്പിനെ ഛായാഗ്രഹണും . നക.ആർ.ഗൗരി
ശങ്കര്‍
‍വിദ്ഗ്ദമായി എഡിറ്റിുംഗ് വിര്‍
വ്വഹിച്ചിരിക്കുന്നു . ഇമ്മാനുവലായി
മമ്മൂട്ടി കാണികനള മുഷിപ്പിച്ച് വിയര്‍
പ്പിക്കുന്നുമില്ല . റഫീഖ്അഹമ്മദ്്
എഴുത്ി അഫ്സലൂസഫ് സുംഗീത്വിര്‍
വ്വഹണും നചയ്ത പാട്ടുകള്‍
‍
ഇമ്മാനുവല്‍
‍ എന്ന സിവിമയ്ക്ക്
ആവശയന്മയില്ലായിരുന്നു .സുവില്‍
സുഖദ്യും സുകുമാരിയും മുക്തയും
ത്ങ്ങളനെ നചറുന്വഷങ്ങള്‍
‍ മന്വാഹരമാകി . ന്കാര്‍
പ്പന്ററ്റ്
കപെത്കളനെ വടവിലുും വേ വിറഞ്ഞ ചിന്തകന്ളാനെ ഒരുവനു
വാഴാും എന്ന ശുഭസൂചകമായ ഒരു പാഠും ഇമ്മാനുവല്‍
‍ വനമ്മ
ഓര്‍
മ്മനപ്പടത്തുന്നു എന്ന ന്പാസീറ്റീവ് ചിന്തന്യാനെ വമുക് പിരിയാും .
പരിപൂര്‍
ണ്ണത് എന്ന അവസ്ഥയ്ക്ക് ഒരു ദൃശയ , ശ്രവയ രൂപമുനണ്ടങ്കില്‍
‍
അത്ാണ് ആന്മവ്‍
‍ ഒരു സിവിമനയ ഓന്രാ ന്പ്രക്ഷകനുും
വയത്യസ്‍
ത്മായ ഭാവത്ലങ്ങളില്‍
‍ വിന്നാണ് കനണ്ടടക്കുന്നത്് .
ചിലര്‍
ക‍
‍ സിവിമ നവറുനമാരു കാഴ്‍
ചയാവാും . ആന്മവ്‍
പരിപൂര്‍
ണ്ണത്നയ സ്പര്‍
ശിക്കുന്നു എന്നതുത്നന്ന .
Sl
No.
Tags Example Description
1 CC_CCD പനക്ഷ Conjunction
2 DIR സുംവിധാവും , സ്ക്രിപ്റ്റ് Direction /Script
3 INEG എന്നല്ല Inverse Negative
4 INTF വളനര Intensifier
5 NEG ന്മാശമാണ് Negation
6 NEU ഒരുകമായി Neutral
7 POS മികച്ചത്ാണ് Positive
8 RD_PUNC . Sentence Ending
9 SPCL
സിവിമയാണ്‍,
ചിത്രെിവ് Film
10 TST കഥാപാത്രും Acting / Song

4
അഭിവന്ദവെിനുന്മല്‍
‍ എത്ര അഭിവന്ദവങ്ങള്‍
‍ നചാരിഞ്ഞാലുും
മത്ിയാകില്ല .
Table II. Result of Individual Category
Category Polarity Rating
Film POSITIVE GOOD
Direction/Script NEGATIVE BAD
Song/Acting POSITIVE EXCELLENT
Others POSITIVE EXCELLENT
Overall POSITIVE GOOD
Over all Result: POSITIVE
Over all Rating: GOOD
CONCLUSION AND FUTURE WORK
The sentiment analysis is the part of cognitive science that
gives the artificial intelligence power to the machine. This
work proposes a method of extracting the sentiments from the
Malayalam film review. We have been implementing a hybrid
approach for finding the sentiment from given sentence or
document. This work would help to assign the rank and
popularity of the new arrival film and also to the users for
expressing their feelings after watching new films. Also help
to find the rating and the score of the film. The polarity and
rating of the individual categories like song, acting, direction,
and script are also done. Presently, the sentiments can be
extracted only from the movie reviews. This work can be
enhanced for extracting the emotions from other areas like
story, novels, product reviews and so on. The other machine
learning approaches can also be used in this study.
REFERENCES
[1] Alekh Agarwal, and Pushpak Bhattacharyya. 2005 . Sentiment
analysis: A new approach for effective use of linguistic
knowledge and exploiting similarities in a set of documents to be
classified, Proceedings of the International Conference on
Natural Language Processing (ICON).
[2] Pang, Bo, Lillian Lee, and Shivakumar Vaithyanathan. 2002.
Thumbs up?: sentiment classification using machine learning
techniques", Proceedings of the ACL-02 conference on
Empirical methods in natural language processing-Volume 10.
Association for Computational Linguistics.
[3] Manurung, and Ruli, 2008 Machine Learning-based Sentiment
Analysis of Automatic Indonesian Translations of English
Movie Reviews, In Proceedings of the International Conference
on Advanced Computational Intelligence and Its Applications
[4] H. Saggion and A. Funk. 2010. Interpreting sentiwordnet for
opinion classification. In LREC.
[5] Turney, Peter D. 2002. Thumbs up or thumbs down?: semantic
orientation applied to unsupervised classification of
reviews, Proceedings of the 40th annual meeting on association
for computational linguistics. Association for Computational
Linguistics.
[6] Barbosa, Luciano, and Junlan Feng.2010. Robust sentiment
detection on twitter from biased and noisy data, Proceedings of
the 23rd International Conference on Computational Linguistics:
Posters. Association for Computational Linguistics.
[7] Celikyilmaz, Asli, Dilek Hakkani-Tur, and Junlan Feng.2010.
Probabilistic model-based sentiment analysis of twitter
messages”, Spoken Language Technology Workshop (SLT).
[8] Mohandas, Neethu, Janardhanan P.S. Nair, and V.
Govindaru.2012. Domain Specific Sentence Level Mood
Extraction from Malayalam Text, Advances in Computing and
Communications (ICACC), 2012 International Conference on.
IEEE.

A Subjective Feature Extraction For Sentiment Analysis In Malayalam Language

Recommended

More Related Content

Similar to A Subjective Feature Extraction For Sentiment Analysis In Malayalam Language (20)

More from Jeff Nelson (20)

Recently uploaded (20)

A Subjective Feature Extraction For Sentiment Analysis In Malayalam Language