NOVATEUR PUBLICATIONS

International Journal of Research Publications in Engineering and Technology [IJRPET]

ISSN: 2454-7875
VOLUME 3, ISSUE 6, Jun.-2017
SENTIMENT ANALYSIS OF MARATHI LANGUAGE
SUJATA DESHMUKH
Department of Information Technology, FRCRCE, Affiliated to Mumbai University, Mumbai, India.
sujata.p.deshmukh@gmail.com

NILEEMA PATIL
Dept. of Physics, LTCE, Affiliated to Mumbai University, Mumbai, India. patil. nileema@gmail.com

SURABHI ROTIWAR
Department of Information Technology, FRCRCE, Affiliated to Mumbai University, Mumbai, India.
surabhi.rotiwar@gmail.com

JASON NUNES,
Department of Information Technology, FRCRCE, Affiliated to Mumbai University, Mumbai, India. jason.nunes@gmail.com

ABSTRACT: is the field of study that analyzes people’s opinions,

Sentiment analysis offers many benefits and sentiments, evaluations, appraisals, attitudes, and
opportunities from business, government and emotions towards entities such as products, services,
consumer perspective in this digital data explosive organizations, individuals, issues, events, topics, and their
age. According to Google, in partnership with KPMG attributes. It represents a large problem space [2, 3].The
India report (April-2017), titled 'Indian Languages - analysis of sentiments may be document based, paragraph
Defining India's Internet', currently India today has based or sentence based. In document approach the
234 million Indian Language users who are online, sentiment in the entire document is summarized as
compared to 175 million English web users and positive, negative or objective. In sentence based where
expected 536 million Indians to use regional languages individual sentences, bearing sentiments, in the text are
while online by 2021. However Marathi users are classified. Sentiment Analysis task is considered a
expected to make significant contribution to define sentiment classification problem. Following are the general
India Internet volume. By considering aspects of steps–
Marathi language and benefits of sentiment analysis,  To select text features i.e. obtaining a dataset.
this paper presents a approach to overcome the  Identifying Parts of speech (POS)– finding adjectives, as
barriers and difficulties being faced for analyzing text they are important indicators of opinions.
in Marathi language. The proposed system detects  Identifying Opinion words and phrases– these are
hidden sentiments in text of Marathi language. The words commonly used to express opinions including
system uses sentiment analysis methodology in order good or bad, like or hate.
to achieve desired functionality. In this system, a  Negation handling–The appearance of negative words
corpus based approach is proposed, i.e the creation of may change he opinion orientation like not good is
a diverse up to date corpus of Marathi keywords, along equivalent to bad.
with their individual polarities, with respect to the Calculate sentiment score and summarize as positive,
Word Net, which is consider as a corpus. The algorithm negative or objective [2, 3].
is used to calculate the cumulative polarity of the text
and rank the sentence as positive, negative or neutral II. LITERATURE REVIEW:
on a set scale standard. The following literature review provides, issues,
KEYWORDS: Sentiment Analysis of Marathi Language, challenges, approaches to understand the system. Walaa
sentiment analysis, corpus based approach Medhat et al.(2014), highlighted that there is still a lack of
resources and researches concerning languages other than
I. INTRODUCTION: English sentiment analysis [4]. Authors reviewed Fifty-
Sentiment Analysis (SA) or Opinion Mining (OM) is the four of the recently published and cited articles, and were
computational study of people’s opinions, attitudes and categorized and summarized. This paper provides basic
emotions toward an entity. The entity can represent understanding and helped to understand corpus and
individuals, events or topics. Sentiment Analysis identifies dictionary based approaches. Soha Ahmed et.al (2013),
the sentiment expressed in a text then analyzes it. presented the challenges faced by researchers when
Therefore, the target of SA is to find opinions, identify the conducting Social SA (SSA) on Arabic text [6]. Further
sentiments they express, and then classify their polarity. It
93 | P a g e
 NOVATEUR PUBLICATIONS
International Journal of Research Publications in Engineering and Technology [IJRPET]
ISSN: 2454-7875
VOLUME 3, ISSUE 6, Jun.-2017
authors suggested some solutions inspired from the Key Issues in This paper Paper presented Same
Conducting highlights key simple pre-processing approach can
literature, and discussed the difficulties faced when Sentiment issues researchers steps and showed how be applied to
performing SSA on Arabic social media text. A pre- Analysis on are facing and these steps improve Indian
Arabic Social innovative classification accuracy. languages.
processing phase to sentiment analysis is proposed and Media Text approaches that
showed the noticeably improvement to the results of Soha Ahmed, have been
Michael developed when
sentiment extraction from Arabic social media data. This
Pasquier, performing SSA on
paper provided the improvements by pre-processing the Ghassan Qadah, Arabic text in
text of SA. Hatem Ghorben and David Jacot ( 2012), showed 2013 [6] general and Arabic
social media text in
that the combination of lexical, morpho-syntactic and particular.
semantic features achieves relatively good performance in Sentiment In order to extract Although translation The concept
Analysis of the semantic does not necessarily of
classifying French movie reviews according to their French Movie orientation of preserve the semantic SentiWordNet
sentiment polarity (positive, negative) [7]. In order to Reviews Hatem words from orientation of words can be applied
Ghorben, David SentiWordNe, a due to the variation of to Indian
extract the semantic orientation of words from
Jacot, 2012 [7] standard word- language common languages.
SentiWordNet, a standard word translation process is translation process usage, in spite of all its
used. This paper provided to use SentiWordNet for corpus was found to be side it has been argued
used. that dictionary-based
based approach for proposed system. Further it argued approach could
that dictionary-based approach could contribute better achieve better results.
A Practical An unsupervised Words like, Lexicon
results. Yakshi Sharma et. Al (2015), proposed an Approach to lexicon based ”NAHI” can invert the coverage can
approach Sentiment Analysis of Hindi Tweets using Sentiment approach for SA is polarity of the be increased
Analysis of Hindi discussed sentence. So, these that present
SentiWordNet with Subjective lexicon method [8]. Each
Tweets, Yakshi words are also one which is
entry in lexicon is categorized into Verb, Noun, Adjective, sharma, Veenu considered in finding limited.
and Adverb. The results indicate that proposed algorithm mangat, Polarity of text.
Mandeep Kaur,
gave better accuracy than unigram presence method. In 2015 [8]
Hindi language, lexicon coverage can be increased as it has Using Document is By means of Performance
SentiWordNet translated into SentiWordNet, scores of system is
limited coverage now. Further authors suggested to do for Multilingual English using for positivity and depending on
research on subjective lexicon which can be extended to Sentiment standard negativity are translation
machine learning, n-grams or both combined. This paper Analysis Translation determined for these software.
Kerstin Denecke, software. Then, the words. An
provides motivation to use SentiWordNet and lexicon 2008 [9] translated interpretation of the
method. Namrata Godbole et. Al (2012), showed a system document is scores then leads to
classified according the document polarity.
that consists of a sentiment identification phase, which to its sentiment
associates expressed opinions with each relevant entity, into one of the
classes “positive”
and a sentiment aggregation and scoring phase, which and
scores each entity relative to others in the same class [5]. “Negative”.
Finally, authors evaluated the significance of scoring Sentiment The various Corpus based Various other
analysis sentiment analysis approach was found to techniques
techniques over large corpus of news and blogs. The above algorithms and algorithms were be the best approach are also
literature summarized in following Table 1. applications: A noted and why one for this system. available and
survey is better than the a combination
TABLE I. LITERATURE REVIEW TO IDENTIFY GAP Walaa Medhat, other is found. of techniques
Title of paper Conclusion Relevance Research Ahmed Hassan, can also be
and authors discussed Gap Hoda Korashy favourable.
Sentiment The various Corpus based Various other 2014[4]
analysis sentiment analysis approach was found to techniques
algorithms and algorithms were be the best approach are also
applications: A noted and why one for our project. available and III. PROBLEM STATEMENT:
survey is better than the a combination
The aim of this system is to address the overall problem
Walaa Medhat, other is found. of techniques
Ahmed Hassan, can also be of sentiment analysis being faced in the Marathi language.
Hoda Korashy favourable. The idea behind the development of this system is twofold;
2014[4]
Large-Scale Authors concluded Mapping should be In large scale Firstly, to create a diverse up to date corpus consisting of
Sentiment that sentiment can done to find the best data it is Marathi keywords like adverbs, adjectives etc. along with
Analysis for vary by opinion while using difficult to
News and Blogs demographic corpus based map the
their individual polarities. This in keeping with other
Namrata group, news source techniques opinions and sentiment analysis literature can be called Marathi Word
Godbole, or geographic also due to
Net. Secondly, aim of this system is to carry out advanced
Manjunath location. the various
Srinivasiah, combinations sentiment analysis using optimal algorithm to obtain the
Steven Skiena, of the words. cumulative polarity of text. Thus, this twofold problem
2012 [5]
statement is converted into following objectives:

94 | P a g e
 NOVATEUR PUBLICATIONS
International Journal of Research Publications in Engineering and Technology [IJRPET]
ISSN: 2454-7875
VOLUME 3, ISSUE 6, Jun.-2017
• To create an system that that uses a corpus based be redundant while calculating the overall polarity of the
approach and can carry out advanced Sentiment Analysis sentence. Hence they are known as stop words.
for a vernacular language i.e. Marathi that detects hidden • Obtaining the polarity of relevant keywords from the
sentiments in text and analyzes the content accordingly. corpus. Here, it is +0.75 and +0.5.
• To create an system that that uses a corpus based • Calculating cumulative polarity.
approach and can carry out advanced Sentiment Analysis The cumulative polarity P is calculated as P=(a+b+c+….)/n
for a vernacular language i.e. Marathi that detects hidden where; a,b,c are the individual polarities mapped from the
sentiments in text and analyzes the content accordingly. corpus, n= No of words.

IV. PROPOSED SYSTEM: V. RESULTS AND DISCUSSION:

The proposed sentiment analysis approach is a This system is implemented in Java. Following screen
Corpus based approach. It is twofold and involves two shot shows the output of the system.
modules. First module provides foundation to the
proposed system by creating of feasible corpus for Marathi
language from English SentiwordNet and second module
presents mapping of keywords to analysis sentiments. The
main tasks of creation of feasible corpus from English
SentiwordNet are Categorization of keywords with respect
to parts of speech (POS) and allocating polarity to
keywords. The main steps for Mapping of keywords to
analysis of sentiments are accept Marathi content as input
for analysis and use the devised algorithm to find polarity
and degree of given input.
A. CORPUS GENERATION FROM ENGLISH SENTI WORD
NET:
This system proposed the creation of a Marathi Figure 1: Highly positive output
corpus. In order to generate the corpus, steps are as
follows:
1. Collection of data online: Find text files which contained
lists of various Marathi keywords and their meanings etc.
2. Segregation and lemmatization: It pre-processed the
data to categorize words as PoS i.e parts of speech, ie
Adjective etc and another corresponding lemma file.
3. Conversion to csv format: After pre-processing, it
converted the text files into 11 csv format in order to be
able to index the keywords properly. Every keyword was
attributed a unique identification number. Then Data File
is created, which has the same corresponding UIDs and the
Figure 2 Neutral polarity output
meaning of the words, followed by its synonyms to
facilitate mapping.
4. Attributing polarity to each keyword: polarity of each
word is calculated for word using English sentiwordnet.
Then, for the purpose of easy indexing, the text files are
converted to csv format.

B. ALGORITHM
The algorithm first find individual polarity of each word in
the sentence and then find the cumulative polarity to
determine if the sentiment is positive, negative or neutral
and to what degree. Sentence in Marathi is a input for the
algorithm. The steps are follows:
• Elimination of stop words. Any words that do not
attribute any specific polarity to the sentence are found to Figure3: Negative polarity output

95 | P a g e
 NOVATEUR PUBLICATIONS
International Journal of Research Publications in Engineering and Technology [IJRPET]
ISSN: 2454-7875
VOLUME 3, ISSUE 6, Jun.-2017
Following Table 2 provides different test cases its better services and facilities in terms of sentiment analysis.
polarity. The scarcity of resources is one of the biggest challenges
Table 2: Test cases with Result while dealing with sentiment analysis for Marathi
Input Output language. This system focused on resource creation which
मला वाटते ती, अततशय सुंदर आहे 0.84, Highly Positive
includes building of an up to date corpus for Marathi
तिती सुंदर आहे ती! 0.75 Highly Positive
ती आहे म्हणून द: खी? 0.0 Neutral Polarity language. The algorithm being proposed is used to give
मी खूप आनुंदी आहे 0.14 Positive Polarity cumulative polarity using the Wordnet. This sentiment
मी आनुंदी नाही -0.12 Negative Polarity analysis model proposes a novel and effective approach to
मी नाही धन्य तिुंवा सखी -0.125, Negative polarity
achieve desired functionality for Marathi language. This
तो नाही आहे िी, मी आनुंदी नाही 0.125, Positive Polarity
system mainly focused on creation of a vast and diverse
"तम्ही िसे आहात?" 0.0 Neutral Polarity
“"मी द: खी आहे " 0.0 Neutral Polarity up-to-date corpus, efficient mapping of data and
After careful analysis of the test cases, the few generation of accurate sentiments for the data to erase the
discrepancies in the obtained results were mainly due to language barriers faced in the field of Sentiment analysis
the following 3 factors: for Marathi.
1) Corpus word limit and Translation accuracy : If
words are not present in corpus built, then online Yandex VII. FUTURE SCOPE:
translator is used and sentiment is analyzed with English The scope of this system is limited for sentence level
SentiWordNet. The freely available Yandex translator that which can further be increased so as to analyze the
system is using provides an accuracy in the range of 60- sentiment for paragraphs and even larger text documents.
70%.Thus, especially for Asian, vernacular languages, the This can be done, by firstly finding the individual polarity
translation is only partially accurate. Thus, these of the sentences using our suggested algorithm and then
inaccurate translations affect the overall polarity of the finding the net cumulative polarity of all sentences to
sentences which can result in minor discrepencies in the obtain the overall polarity of the paragraph or document.
overall output. However, for smaller and simple sentences, There is a need for an optimized and accurate algorithm to
the translation accuracy is comparatively higher than that do the same. Further limited size of corpus can be
of complex compound sentences. It was an observation increased by not only considering adjectives but also verbs
that the framing of the sentences makes a difference in the and nouns which are already present in the English Senti
sentiment analysis. Word Net. This while evaluating the sentiment, better
2) Limited scope of English SentiWordNet: The English accuracy and results can be obtained. Also, increase the
SentiwordNet 3.0 which system is using to obtain the dictionary size by regularly updating it with new words.
polarities of the respective words, also needs a lot of Further real time update of dictionary can be future
optimization. Many words which are actually of neutral research direction in the field of sentiment analysis of
polarity, are misclassified as adjectives having higher or Marathi language.
lower polarity and hence either provide wrong polarities
or tamper with the algorithm. REFERENCES:
3) Non acceptance of special characters: This system 1) Report 'Indian Languages - Defining India's Internet', a
perfectly analyzes and parses basic punctuation like study by KPMG in India and Google April 2017.
commas, question marks, exclamation marks etc., but a 2) Sentiment analysis, https:// en.wikipedia.org/ wiki/
special character like “” etc. is parsed along with the word, Sentiment_ analysis, accessed in Aug. 2016.
the system fails to recognize and process it and doesn’t 3) Alessia D’Andrea, Fernando Ferri, Patrizia Grifoni,
give any output. Hence there is a need to include a separate “Approaches, Tools and Applications for Sentiment
exception class to eliminate these special characters. Analysis Implementation”, International Journal of
Computer Applications (0975 – 8887) Volume 125 –
VI. CONCLUSION: No.3, September 2015
Sentiment Analysis has been quite popular and has lead 4) H. K. Walaa Medhat, Ahmed Hassan, “Sentiment
to building of better products, understanding user’s analysis algorithms and applications: A survey,” Ain
opinion, executing and managing of business decisions. Shams Engineering Journal (2014) 5, 1093–111, 2014.
With rapidly increasing technology, the early approach of 5) S. S. Namrata Godbole, Manjunath Srinivasaiah, “Large-
word-of-mouth has been shifted towards the mass opinion scale sentiment analysis for news and blogs,”
what the people like and appreciate in majority. The rise in ICWSM’2007, Boulder, Colorado,USA, 2007.
user-generated content for Marathi language across 6) G. Q. Soha Ahmed, Michael Pasquier, “Key issues in
various genres- news, culture, arts, sports etc has opened conducting sentiment analysis on Arabic social media
the data to be explored and mined effectively, to provide text,” IIT`13, 2013.

96 | P a g e
 NOVATEUR PUBLICATIONS
International Journal of Research Publications in Engineering and Technology [IJRPET]
ISSN: 2454-7875
VOLUME 3, ISSUE 6, Jun.-2017
7) D. J. Hatem Ghorben, “Sentiment analysis of french
movie reviews,” Proceedings of the 7th Atlantic Web
Intelligence Conference, AWIC 2011, pg. no. 19-28,
2011.
8) M. K. Yakshi sharma, Veenu mangat, “A practical
approach to sentiment analysis of Hindi tweets,” 1st
International Conference on Next Generation
Computing Technologies (NGCT), pg. no. 677-680,
2015.
9) Denecke, Kerstin. "Using sentiwordnet for multilingual
sentiment analysis." Data Engineering Workshop, 2008.
ICDEW 2008. IEEE 24th International Conference on.
IEEE, 2008.

97 | P a g e