Sentence level sentiment polarity calculation for customer reviews by considering complex sentential structures

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 03 Issue: 03 | Mar-2014, Available @ http://www.ijret.org 433
SENTENCE-LEVEL SENTIMENT POLARITY CALCULATION FOR
CUSTOMER REVIEWS BY CONSIDERING COMPLEX SENTENTIAL
STRUCTURES
N.Lalithamani1
, Leela Sravanthi Thati2
, Rakesh Adhikesavan3
1
Assistant Professor (SG), Department of Computer Science & Engineering, Amrita School of Engineering, Amrita
Vishwa Vidyapeetham(University), Tamil Nadu, India
2
Student 4th
year B.Tech, Computer Science & Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham
(University), Tamil Nadu, India
3
Student 4th
year B.Tech, Computer Science & Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham
(University), Tamil Nadu, India
Abstract
People often find themselves in a situation where they need to make simple decisions, for example, what book to read, or which movie
to watch, or which smart phone to buy. In these kinds of situations, one often seeks the opinion of others in the form of IMDB
(International Movie Database) rating or magazine reviews or online product reviews etc. The general opinion about a product or a
publication or a service or a company is usually expressed in the form of customer reviews and the web has facilitated sharing and
expressing opinions in the form of free style texts which is unstructured or semi structured. In this paper, we introduce a sentence
level, sentiment polarity calculation that identifies complex sentential structures and modifies the sentence polarity accordingly. The
sentiment polarity of a text is given by a score that lies in the range [-1,1] and denotes the positivity of the tone of the text’s author.
Thus, the sentiment polarity or the sentiment score of a review can be used to perceive the opinion of the reviewer. The average
polarity of all the words in a sentence gives the polarity of a sentence. The polarity of all the common words in the English dictionary
is retrieved from CLiPS Pattern module (BSD license) for Python. After obtaining the polarity of a sentence, it is modified based on
the structure of the sentence which is identified using a set of heuristic rules that use POS (Parts of Speech) tagging to identify
relationships between words. Then the average polarity of all the sentences in a review, gives the overall polarity of the review. We
have compared the results against the inbuilt function pattern.en.sentiment (sentence) and based on empirical evidence the result are
more accurate in many cases and in other cases the results are the same.
Keywords: Sentimental analysis, Sentiment polarity, Opinion mining, and Sentiment mining
--------------------------------------------------------------------***----------------------------------------------------------------------
1. INTRODUCTION
The World Wide Web is an enormous collection of user
generated data. Social networking, blogs, forums, Twitter,
micro-blogs,etc has facilitated and encouraged people to share
their personal experiences or viewpoints / opinions about a
product, service or a company and many users of the web are
seeking these personal experiences or opinions of others
because these opinions are important due to reasons stated by
Bing Liu [1], which are:
 Opinions are key influences of our behavior
 Our perception and beliefs of reality are based on
how others see the world
 Whenever we need to make a decision we seek
opinion from others, individuals need to make
decisions to buy a product or avail a service and they
seek opinions from friends and family, while
organizations try to benchmark their products and
services and use consultants, surveys, opinion polls
etc to find consumer opinions.
With ever increasing amount of data and diverse viewpoints
that is presented on the web, it is impossible for one to discern
an overall opinion or a general viewpoint of a product, service
or a company by just reading through reviews, also it is not
humanly possible for one to read more than a particular
number of reviews, which makes automating the process of
extracting opinions from raw data a very important and
pertinent task.
In order to address this problem of Opinion Mining or
Sentiment Mining, in this paper we introduce a method to
modify prior assigned sentiment polarity by considering
complex sentential structures.

__________________________________________________________________________________________
Generally, as observed by Polanyi and Zaenen [6] and Wilson
et al. [7], there exist complex relationships between words in a
sentence that influences its sentiment polarity. For example,
consider negation (eg. “not” in the sentence “The screen
resolution is not good” or “never” in the sentence-“The
products of this series has never performed well”), intensifier
terms (eg. “very” intensifies the sentiment of its adjacent word
in “The performance is very good” ), Un certainty (eg. “The
mileage has to be good”) etc.
In our approach, we use prior assigned sentiment scores of
words in the English dictionary from CLiPS pattern module
for Python [2] and calculate sentence polarity and then
identify such complex relationships between words with the
use of heuristic rules and then identify its influence on the
sentiment polarity and then modify the polarity accordingly.
2. RELATED WORK
Previous works have studied sentiment analysis at document
level or sentence level to predict the overall sentiment polarity
of the document. Hu and Liu [4] extracted and summarized
customer reviews of Electronic goods such as smart phones,
televisions etc. They extracted features of entities (products)
such as resolution, battery life etc and predicted whether the
opinion of the feature or aspect is positive or negative by
checking for the presence of positive or negative opinion
words. Wilson et al. [7] used a machine learning approach to
predict sentiment polarity in the phrase level, while Shaikh et
al [3] and Moilanen and Pulman [5] analyzed the words
present within phrases to detect sentiment polarity.
Kim and Hovy [15] tried an alternate way of summarizing
reviews by extracting information on why a reviewer liked or
disliked a particular product, additional problem is use of
comparative sentences [16, 17] like “ Brand X is better than
brand Y because brand Y has poor performance” and are
important source of information that influence the sentiment
of the review.
3. PROPOSED MODEL
Part-of-speech information is very often exploited in sentiment
analysis because POS tagging can be used for word sense
disambiguation [9]. Researchers suggest that verbs (eg. “like”)
can be strong influencers of sentiment. There have been many
comparisons of the effectiveness of adverbs [14], verbs,
adjectives in influencing sentiment polarity and sub
categorization often plays a role too [10,11,12].
In our approach, sentence structures and complex relationships
between words are identified using POS tagging which is
achieved by using the tag( ) function that is imported from
pattern.en package.
Some of the common tags are:
Table -1: POS tags
Tag Description Example
CC conjunction, coordinating and, or, but
CD cardinal number five, three, 13%
DT determiner the, a, these
EX existential there there were six boys
FW foreign word mais
IN conjunction,
subordinating or
preposition
of, on, before, unless
JJ adjective nice, easy
JJR adjective, comparative nicer, easier
JJS adjective, superlative nicest, easiest
NN noun, singular or mass tiger, chair, laughter
NNS noun, plural tigers, chairs, insects
NNP noun, proper singular Germany, God, Alice
NNPS noun, proper plural we met
two Christmases ago
POS possessive ending 's
PRP pronoun, personal me, you, it
PRP$ pronoun, possessive my, your, our
RB adverb extremely, loudly, hard
RBR adverb, comparative better
RBS adverb, superlative best
RP adverb, particle about, off, up
SYM symbol %
VB verb, base form think
VBZ verb, 3rd person singular
present
she thinks

__________________________________________________________________________________________
VBP verb, non-3rd person I think
VBD verb, past tense they thought
VBN verb, past participle a sunken ship
VBG verb, gerund or present
participle
thinking is fun
Fig - 1: System Design
Fig -1 is a block diagram that shows the basic design of the
system and gives a crude idea of how the sentiment polarity is
obtained; the input is read in the form of raw text which is
converted into an array of sentences using NLTK (Natural
language Toolkit) sentence tokenizer.
(http://www.nltk.org/api/nltk.tokenize.html)
The polarity of words is retrieved from the package pattern
and the sentence polarity is calculated using: Sum of polarity
of all the words in a sentence divided by the total number of
words in the sentence.
Polarity(word) – polarity value returned by
pattern.en.polarity(word)
N(words) – total number of words in the sentence
A set of heuristic rules is now used to modify the sentence
polarity by identifying complex phrases and relationships
between words. Each sentence is checked against each rule
and if there is a match the existing sentence polarity is
modified based on the rule. The more rules that can be
identified and added to the system, more complex
relationships between words or phrases can be identified and
the accuracy of the end resultant polarity value of the sentence
will be more accurate.
Fig - 2: System Design with defined Rules
Fig – 2 expands the fourth step of Fig – 1 (System Design)
which is “Modify Sentence Polarity by identifying sentence
structure” and displays the set of heuristic rules that are
explained below. The dotted blocks imply that more rules can
be added to the design to improve accuracy of the system. The
briefings of rules are as follows:
3.1 Rule 1: Negation
Polarity of the word that follows the word “not” or “never” or
a word ending with “n‟t” is negated.
Consider the following example sentences from product
reviews: (Note: All the following examples are sentences
extracted from reviews that were taken from the web and the
product name and model have not been mentioned)
 The battery life is not good.
 The lamp isn’t good.
 Negate polarity of “good”
 The models of this series have never performed
great.
 Negate polarity of “great”

__________________________________________________________________________________________
However, one difficulty with handling negation is that it can
often be expressed in subtle ways. Irony and sarcasm is very
hard to detect. Wilson et al. [7] discusses some of the complex
negation effects. Na et al. [13] attempted to model negation
more accurately and observed a 3% increase in accuracy.
3.2 Rule 2: Symbols
The presence of emoticons or smiles and exclamation marks
influences the sentiment polarity. Consider the following
sentences:
 This Ultra Book is great!
 Intensify sentence polarity
 The maps is very unreliable 
 Intensify sentence polarity
3.3 Rule 3: Adjective Intensifiers
There are some words that can be used to intensify adjectives
such as, „very‟ „utterly‟ „totally‟ „completely‟ „absolutely‟
„entirely‟ „really‟.
Consider the following sentence:
 The design is very beautiful
 Intensify polarity of beautiful
3.4 Rule 4: Uncertainty or Suggestion
Many a times, reviewers make a suggestion on how a
particular feature of the entity or product or service has to be
improved, the polarity of such sentences should rather be
neutral than positive or negative. Consider the following
examples:
 The picture clarity has to be made better
 Neutralize polarity of “better”
 The graphic processor has to be excellent for Hi Def
movies to play smoothly
 Neutralize polarity of “excellent”
Sometimes, reviewer is uncertain about something, for
example
 The mileage ought to be good
 Neutralize polarity of “good”
 The product might be expensive
 Neutralize polarity of “expensive”
This is achieved by looking for a comparative adjective that
follows the word “be”, In the case of superlative adjective the
polarity is not modified, consider the example
 This has to be the best phone I have ever seen
 Sentence Polarity is not modified
POS tags are identified using POS tagger that is imported
from pattern.en package. (Refer to Table-1 for the POS tags).
3.5 Rule 5: Identify PROS and CONS
Some reviews have a PROS and CONS list, consider the
example:
PROS
Battery life
Tons of apps
All slimmed up
It‟s fast
Dual Cameras
CONS
No standard slots or connectors
No Flash
Wi-Fi issues
Most of the words in the above list have zero polarity, but a
PROS and CONS list definitely influences the sentiment of the
review. So we could compare the number of items in each list
and modify polarity according to which list has higher number
of items.
3.6 Rule 6: Identify “however” and “although”
This rule has not been implemented and tested yet and is a
mere suggestion or a proposal. Consider the following
sentences:
 Although the phone is very expensive it‟s a great buy
because of the excellent aesthetics
 The overall sentiment is positive
 Although the printer is efficient it is very expensive
 The overall sentiment is negative
 The boot space is good however the leg space in the
front seat is very less
 The overall sentiment is negative
 The CFL is expensive however it is very energy
efficient
 The overall sentiment is positive
In such cases the polarity can be modified by identifying the
words “however” or “although” and identifying towards which
side the sentiment is leaning towards.
4. RESULTS
A tool (named Sentinel) was built using Python programming
language that follows the above proposed approach to modify
sentence polarity based on the sentential structure.
Sentinel reads a text file as input and prints the Sentiment
Score of the text. The polarity of all the sentences is obtained
and modified based on the heuristic rules and the average
polarity of all the sentences in the text file gives the Sentiment
Score of the text which denotes the positivity of the tone of the
text‟s author.

__________________________________________________________________________________________
Sample input and output:(Note: Inputs are sentences extracted
from product reviews obtained from cnet.reviews.com)
Input: Battery pack case plus extra flash storage: too good to
be true? The Mophie Space Pack might be one of the best
IPhone accessories ever seen
Output: 0.65
Input: Although the battery life is poor, the phone is a great
buy because it is affordable
Output: 0.6
Input: The performance is good
Output: 0.7
Input: The performance is really good
Output: 1
Input: The case isn‟t attractive
Output: - 0.8
5. COMPARISON WITH BENCHMARK
CLiPS (Computational Linguistics & Psycholinguistics
Research Center) patter.en module contains a fast part-of-
speech tagger for English (identifies nouns, adjectives, verbs,
etc. in a sentence), sentiment analysis, tools for English verb
conjugation and noun singularization & pluralization, and a
WordNet interface. Interested readers refer to [2] for more
details on CLiPS pattern module.
Sentinel‟s results have been tested against the inbuilt function
pattern.en.sentiment(sentence) that returns (polarity,
subjectivity) tuple, here we consider the sentiment polarity
alone.
In: patten.en.sentiment(“The screen ought to be good”)[0]
Out: 0.7
Setinel Output: 0
In: patten.en.sentiment(“The maps has never been good”)[0]
Out: 0.7
Sentinel Output: - 0.7
In: patten.en.sentiment(“The graphic processor has to be
excellent for Hi Def movies to play smoothly”)[0]
Out: 0.5
Sentinel Output: 0.1
6. SUMMARY
To review, determining the overall Sentiment score of
user/customer reviews of a product, publication, service or a
company is an important research topic, it is especially useful
for services that have a large number of reviews, to help an
opinion seeker discern the public view point quickly and
accurately. However I would like to add some cautionary
notes, human interpretation is still required to understand
sarcasm, satire, etc. and a machine can easily misread the
above, also the redundancy issue as mentioned by Tsaparas,
Ntoulas and Terzi [8] is also a valid concern, In my opinion
the distribution (in terms of negative and positive points) is
important in determining the quality of a review and users
tend to feel a review is helpful or not based on whether the
review appears to be genuine and whether it expresses opinion
about many features of the product or not. So developing the
program to determine the genuineness and helpfulness of a
review could be a useful and important addition.
7. CONCLUSIONS
Through this paper we have covered some very interesting
techniques and approaches to calculate sentiment polarity and
have proposed a way to modify pre assigned polarity values to
increase accuracy of sentiment polarity calculation, and have
explained the pertinent and intriguing task of sentimental
analysis and we encourage the readers to take up some of the
many open challenges that still remain.
ACKNOWLEDGEMENTS
We would like to express our sincere thanks and gratitude to
Ms.N.Lalithamani, Assistant Professor(SG), Dept. of CSE,
Amrita Vishwa Vidyapeetham for her guidance and we would
also like to express our greatest gratitude to Mr.Sai
Sundarakrishna,Research Scientist, R&D Dept., Caterpillar
India Pvt. Ltd for his motivation and mentoring.
Rakesh & Sravanthi
REFERENCES
[1]. Bing Liu, “Sentiment Analysis and Opinion Mining”,
Morgan and Claypool Publishers, May 2012.
[2]. Tom De Smedt “Pattern For Python”, International
Journal of Machine Learning Research 13, pp. 2063 – 2067,
2012.
[3]. Shaikh, M.A.M., Prendinger, H., Ishizuka, M.: “Sentiment
Assessment of Text by Analyzing Linguistic Features and
Contextual Valence Assignment”, Applied Artificial
Intelligence, Vol. 22, No.6, pp. 558–601, 2008.
[4]. Hu M, Liu B. “Mining and Summarizing Customer
reviews”. In Proceedings of the 10th Special Interest Group on
Knowledge Discovery and Data Mining (SGKDD), Seattle,
WA, USA, pp 168-177, August 2004.
[5]. Moilanen, K., Pulman, S.: “Sentiment Composition”,
“International Conference on Recent Advances in Natural
Language Processing”, pp. 378–382, 2007
[6]. Polanyi, L., Zaenen, A. “Preface: Contextual Valence
Shifters”, Computing Attitude and Affect in Text: Theory and
Applications, Springer Publications, Vol. 20, pp. 1–10, 2006
[7]. Wilson, T., Wiebe, J., Hoffmann, P.,“Recognizing
contextual polarity: An exploration of features for phrase-level
sentiment analysis”, Journal of Computational Linguistics,
Vol. 35, No.3, pp. 399–433, 2009

__________________________________________________________________________________________
[8]. P. Tsaparas, A. Ntoulas,E. Terzi, “Selecting a
Comprehensive Set of Reviews”, Proceedings of International
Conference on Knowledge Discovery and Data Mining
(KDD), August 2011
[9]. Yorick Wilks and Mark Stevenson. “The grammar of
sense: Using part-of-speech tags as a first step in semantic
disambiguation”, Journal of Natural Language Engineering,
Vol. 4, No.2, pp. 135–144, 1998.
[10]. Farah Benamara, Carmine Cesarano, Antonio Picariello,
Diego Reforgiato, and V. S. Subrahmanian.
“Sentiment analysis: Adjectives and adverbs are better than
adjectives alone”, Short Paper: Proceedings of the
International Conference on Weblogs and Social Media
(ICWSM), 2007.
[11]. J. Kamps and M. Marx and R.J. Mokken and M. De
Rijke, “Using WordNet to measure semantic orientation of
Adjectives”, Proceedings of the International Conference on
Language Resources and Evauation (LREC-04), Vol. 4, 2004,
pp. 1115 - 1118, Lisbon, Portugal.
[12]. Janyce M. Wiebe, Theresa Wilson, Rebecca Bruce,
Matthew Bell, and Melanie Martin. “Learning
subjective language”, Journal of Computational Linguistics,
Vol.30, No.3, pp. 277–308, September 2004.
[13]. Jin-Cheon Na, Haiyang Sui, Christopher Khoo, Syin
Chan, and Yunyun Zhou. “Effectiveness of simple
linguistic processing in automatic sentiment classiﬁcation of
product reviews”, Proceedings of the International
Conference of the International Society for Knowledge
Organization (ISKO)”, pp. 49–54, 2004.
[14]. Vasileios Hatzivassiloglou and Kathleen McKeown.
“Predicting the semantic orientation of adjectives”,
Proceedings of the International Joint Conference on
ACL/EACL (Association for Computational Linguistics/
European Association for Computational Linguistics) ”,
pp.174–181, 1997.
[15]. Soo-Min Kim and Eduard Hovy. Crystal: “Analyzing
predictive opinions on the web”, Proceedings of the
International Joint Conference on Empirical Methods in
Natural Language Processing and Computational Natural
Language Learning (EMNLP-CoNLL), 2007.
[16]. Nitin Jindal and Bing Liu. “Identifying comparative
sentences in text documents”, Proceedings of the ACM
Special Interest Group on Information Retrieval (SIGIR)”,
2006.
[17]. Nitin Jindal and Bing Liu. “Mining comparative
sentences and relations”, Proceedings of the Interanational
Conference on Association for the Advancement of Artificial
Intelligence (AAAI)”, 2006.
BIOGRAPHIES
Ms.N.Lalithamani was born in 1975. She
received her Master‟s degree in 1998. She
holds a University rank as a part of her post
graduate programme. Her fields of interests
include (active) databases, automated
biometrics, Information retrieval etc. She is currently an
Assistant Professor at Amrita School of Engineering, Amrita
Vishwa Vidyapeetham. In the past she has worked on several
projects concerning E-learning.
Leela Sravanthi That is the student of Amrita
School of Engineering, Coimbatore, India. She
was awarded Prathibha Award in her 10th
Standard from Andhra Pradesh State
government and she also stood among the top
qualified students in International Maths Olympiad in the year
2009-2010.
Rakesh Adhikesavan, born in 1992 is now
pursuing his B.Tech degree in Computer
Science and Engineering from Amrita
Vishwa Vidyapeetham and is currently
participating in a Global Study Program at
University of California Davis. His interests
in the field of Computer Science include Information Retrieval
and Information Visualization.

Sentence level sentiment polarity calculation for customer reviews by considering complex sentential structures

Recommended

More Related Content

What's hot (19)

Viewers also liked (20)

Similar to Sentence level sentiment polarity calculation for customer reviews by considering complex sentential structures (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

Sentence level sentiment polarity calculation for customer reviews by considering complex sentential structures