2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept.

21-24, 2016, Jaipur, India

Summarization of Customer Reviews for a Product

on a website using Natural Language Processing
Akkamahadevi R Hanni#1, Mayur M Patil*2 Priyadarshini M Patil%3
#
Student, *Student (2015 Batch), Dept. of CSE, B.V.B. % Assistant Professor, Dept. of ISE, B.V.B. College of
College of Engineering and Technology, Engineering and Technology,
Hubballi - 580031, Karnataka, India Hubballi - 580031, Karnataka, India
1
[email protected], [email protected] 3
[email protected]

Abstract— In the recent past, e-commerce sites have made rapid language processing approach. We propose a dynamic system
growth. There are thousands of products and various websites for feature based comment summarization based on the
sell these products. Massive growth in the number of reviews and corresponding domains of products. In this process, the
their availability along with the advent of opinion-rich review reviews are extracted by web crawling. These compound
forums for the products sold online, choosing the right one from
a large number of products has become difficult for the users.
sentences are broken down into individual sentences and
HELP-ME- BUY APP is an android application that assists further into words by sentence-tokenization and word-
tokenization respectively [10]. The words are POS (Parts of
buyers in online shopping. It is imminent for buyers to verify for
genuineness and quality of products. What better way is there Speech Tagging) tagged [11] to help classify their position in
than to ask people who have already bought the product? This is a sentence that will help in extraction of features and opinions.
when customer reviews come into picture. The major hitch here Now, identification and extraction of the features of a product
is popular products have thousands of reviews-we do not have is done first [4,5]. Next the opinion regarding these identified
the time or patience to read all thousands of them. Hence, our features is found and their polarity (negative/positive/neutral)
application eases this task by analyzing and summarizing all is detected. Once this is done, excerpts with respect to these
reviews which will help the user decide what other buyers have
experienced on buying this product. We carry out this process by
feature-opinion pair are extracted and further used for
a number of modules that include feature extraction and opinion summarization. This summarized comment provides a
extraction which improves the process of analysis and helps in complete overview of opinions of users and also stresses on
the formation of an efficient summary. each feature of the product, making it easier for both
customers and also the producers to know the response of the
Keywords— Natural Language processing, Machine Learning, mass.
android, raw text analysis, summarization, POS tagging As the number of reviews are in terms of hundreds on
certain products and in terms of thousands on popular
I. INTRODUCTION products it is evident that the user may not read all the reviews
With the increasing use of e-commerce websites our project and might miss out on some critical reviews that concern his
improves the shopping experience for users. The domain of needs. Hence we provide a solution to summarize it based on
our project lies under Natural Language Processing (NLP) the product’s features. This saves the time and energy of the
which basically includes analysis, classification and users which would rather be well spent. The user will be able
summarization of raw text obtained from customer reviews. to decide on one look of the graphical outcome of the
The review for a product on the Amazon.in website consists summarization.
of an overall rating of the product which is obtained from a The users for this project would be all the customers who
statistics of each individual’s customer rating out of 5 stars buy the products online. Currently this project is limited to
and a customer review section where customers drop in their amazon.in website only because the web crawling and
experience on buying that particular product. While the scraping is implemented for only this site, as our major focus
overall rating gives a vague idea of the product’s genuineness, is building an analysis and summarization algorithm. The
the customer review section gives a potentially elaborated idea. feature and opinion based extraction [1] is limited only to a
Although reading through the customer reviews gives a certain variety of products such as smart phones, clothing etc.
comprehensible picture, it might be very time consuming in
some cases where the product has thousands of reviews listed.
We propose to analyse and summarize these customer reviews II. RELATED WORK
which are unstructured using sophisticated NLP toolkits [3,16]. Our work is similar to [4, 5, 9]. Minqing Hu and Bing Liu
In this project, the design of a unified opinion mining and propose a feature and opinion mining system for customer
sentiment analysis framework is presented with natural review summarization. They have adopted a 3 step process “(a)

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-2029-4/16/$31.00 @2016 IEEE 2280
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

mining product features that have been commented on by hundreds and thousands of reviews. As a result, users might
customers; (b) identifying opinion sentences in each review miss out on some critical reviews.
and deciding whether each opinion sentence is positive or
negative; (c) summarizing the results.” However, this system A. Objectives of the application
does not generate a summary of the reviews by selecting a • To build an algorithm for summarization of customer
subset of the dataset which speak of the significant features of reviews.
the product. • To extract reviews, perform analysis on them,
classify them based on polarity and produce a
In [5], the author has included a cluster based analysis and summary.
feature centric approach to solve the problem by taking into • To implement a unique ‘feature’ and ‘opinion’ based
consideration the distance of each opinion word with respect analysis to produce a more critical review summary.
to product feature and then calculating the overall opinion of • To provide a feature based rating on the respective
the sentence. The author has also adopted association rule product
mining for feature extraction and probabilistic approach for
the process of filtration to remove all the feature candidates
which are no real features, but are just frequently occurring IV. IMPLEMENTATION
nouns or noun phrases. When it comes to online shopping, for trendy products, the
number of comments/reviews i.e. the user opinion could be in
In [6], the author has implemented a linguistic-based thousands. No doubt the ratings (out of five) may give an
application that gives feedback of statistics about the positive overview of the opinions of users regarding a product but it
or negative nature of the opinions and about the semantic isn’t completely reliable. There exist certain comment
categories of the features. It is based on contextual analysis, classifiers which analyse and classify the reviews as positive
where the related data might actually point to different or negative. This isn’t an efficient approach as well. If a
contexts/aspects but only few would fit the purpose, hence comment is classified as negative, it doesn’t mean the user
extracting such features which effectively influence the results. doesn’t like any of the features. Various research efforts
attempted to mine opinions from customer reviews at different
In [7], for the process of feature extraction, the authors levels of granularity, including word, sentence, and document-
adopt stemming and assigning of syntactic roles using level. However, development of a fully automatic opinion
Stanford Dependency Parser to extract legitimate mining and sentiment analysis system is still elusive.
nouns/features. Further, they include Usefulness Pruning of In this paper, the design of a unified opinion mining and
facets technique and Compactness Pruning technique for sentiment analysis framework is presented with natural
removal of redundant facets that are not discovered in language processing approach. We propose dynamic system
association mining. for feature based comment summarization based on the
corresponding domains of products. In this process, the
In [8], the author states a comparative analysis of mining reviews are extracted by web crawling [18,19]. Let us
based summarization approaches and comprehensive review consider one of the customer reviews for a smart phone:
of opinion summarization. It provides a multi-perspective “Display is superior and of top-notch quality. The camera
classification of the approaches used and highlight some of is simply great. But the battery is extremely lousy and phone
their key weaknesses. gets heated up very often.”
These compound sentences are broken down into
In [9], the SummView is a web-based engine for individual sentences and further into words by sentence-
summarizing product reviews and customer opinions. It tokenization and word-tokenization respectively [10]. The
adopts a mining extraction approach similar to [4]. The would now generate an output as below:
application includes a user choice to select the features to be [‘Display’, ‘is’, ‘superior’, ‘and’, ‘of’, ‘top-notch’,
summarized specifically and constructs a term-sentence ‘quality’, ‘.’, ‘The’, ‘camera’, ‘is’, ‘simply’, ‘great’, ‘.’, ‘But’,
matrix which is essentially a feature-based weighted non- ‘the’, ‘battery’, ‘is’, ‘extremely’, ‘lousy’, ‘and’, ‘phone’, ‘gets’,
negative matrix factorization algorithm that groups sentences ‘heated’, ‘up’, ‘very’, ‘often’, ‘.’]
into relevant clusters. However, this application gives a
textual summary for each feature and not an overall rating. The words are POS (Parts of Speech Tagging) tagged [11]
to help classify their position in a sentence that will help in
extraction of features and opinions. The tagged output is as
III. PROPOSED SYSTEM below:
In order to know the genuine-ness and quality of the [(‘Display’, ‘NNP’), (‘is’, ‘VBZ’), (‘superior’, ‘JJ’), (‘and’,
products online, the users, as a matter of fact, tend to go ‘CC’), (‘of’, ‘IN’), (‘top-notch’, ’JJ’), (‘quality’, ‘NN’), (‘.’,
through the customer reviews and decide based on those ‘.’), (‘The’, ‘DT’), (‘camera’, ‘NN’), (‘is’, ‘VBZ’), (‘simply’,
reviews. Sometimes it is time consuming as there are ‘RB’), (‘great’, ‘JJ’), (‘.’, ‘.’), (‘But’, ‘CC’), (‘the’, ‘DT’),
(‘battery’, ‘NN’), (‘is’, ‘VBZ’), (‘extremely’, ‘’RB’), (‘lousy’,

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
2281
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

‘JJ’), (‘and’, ‘CC’), (‘phone’, ‘NN’), (‘gets’, ‘VBZ’), (‘heated’, The paragraph summary is generated from these lists by
‘VBN’), (‘up’, ‘’RP’), (‘very’, ‘RB’), (‘often’, ‘RB’), (‘.’, ‘.’)] considering the ratio with respect to the overall rating.
Whereas the feature based rating is generated from the
Where the tags have a specific meaning, NNP- analysis of feature-opinion pairs along with the overall
proper noun, NN- singular or mass, VBZ- verb-3rd person- statement score. Refer Fig. 1 which depicts the
singular, VBN- verb-past-participle, JJ- adjective, RB- adverb, implementation in modular flow diagram.
CC- coordinating conjunction, DT- determiner, IN-
preposition etc.
Now, keeping in mind that nouns are name of things and
adjective describes the noun, identification and extraction of
the features of a product is done first, which essentially is
depicted by a noun in a sentence. Next the opinion regarding
these identified features are found which is depicted by
adjectives. Once this is done, excerpts with respect to these
feature-opinion pair are extracted and run through a frequency
distributor [12] to filter out only those which have a high
frequency and are most related to the product in consideration.
The sentences are now analysed [14, 15, 16] and classified
into positive, neutral and negative statements with a score(x)
given to each statement between -1<=x<=1. The above
example is analysed as below:
“Display is superior and of top-notch quality.”- “0.5423”
“The camera is simply great.”- “0.6249”
“But the battery is extremely lousy and phone gets heated
up very often.”- “-0.5849”
We can see that the first two statements get a positive score
while the 3rd statement gets a negative score. If there are any
neutral statements, the score is zero.
After the analysis, the feature based summary is generated
[17] and represented by a rating bar on a scale of 10 and a
paragraph summary which includes statements that speak of Fig. 1 The Modular flow of the Summarization Process
the features in their review. This provides a complete
overview of opinions of users and also stresses on each feature
of the product, making it easier for both customers and also B. Process Details
the producers to know the response of the mass.
A. Modular flow of the application • Tokenizing [10]: It is essential to break down the
data that is web crawled for simplicity of processing,
When the user is looking at a particular product and wants
analysis and classification. Each customer review is
to find the summarization of the customer reviews, they have
obtained as a paragraph. This is broken down into
to enter the URL which points to the product in the
individual sentences and stored in lists. Further it is
application and click on Get Summary which validates the
necessary to break each sentence into words in order
given URL and then for the validated link, the customer
to infer its meaning, perhaps, also to extract the
review information is fetched.
features of the product that are spoken of.
The fetched information is then pushed to the local storage
• POS Tagging [11]: Parts of speech tagging is the
and simultaneously undergoes sentence tokenizing [10], word
process of tagging or labelling each word in a
tokenizing and POS Tagging [11] which refers the Word
sentence with its part of speech as in a verb, noun,
Corpus for processing.
adjective, adverb etc. i.e. it tags each word with the
The tagged POS words are then used to extract the Feature
part of speech it belongs to.
and Opinion words and undergo filtration in order to identify
the suitable words(features) that are more specifically related • Word Corpus: This essentially refers to the nltk
to the product. From this we get the Feature and Opinion word corpus package which defines a collection of corpus
list. reader classes, which can be used to access the
The customer reviews stored on the local storage and the contents of a diverse set of corpora. This is used by
Feature and Opinion word list are used for analysis with the the POS tagger.
help of Sentiment analysis training set [14, 15, 16]. After this • Frequency distribution [12]: The ‘FreqDist’ class is
process we have three categorized lists: positive, neutral and used to encode the count of number of times that
negative list. each of an experiment occurs. The issue we face is
that not always the reviews submitted by the

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
2282
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

customers are simple sentences like ‘camera is o They use feature selection techniques such
simply great’, the data consists of random junk as Document Frequency Thresholding
sentences that cannot be used for analysis. After (DFT), Information Gain (IG) and Mutual
gathering all the nouns that may include key-features Information (MI).
of the product and definitely includes junk objects, o The drawback being there is no application
we filter out the words that occur less frequently and on customer reviews
are less likely to depict as a ‘feature’ of an object. • Mining and summarization of customer reviews [1,4]
• To improve the results, we may compare our results o Mine the features of the product on which
with the actual feature list of the product which are the customers have expressed their
extracted by crawling. But this cannot be accurate as opinions
feature names used may be different. For example, a o The drawback being there is no ready
product xyz’s feature list may include ‘dimensions’ application present for use of customers.
where as in the reviews the customer may refer to it • Opinion Observer: analysing and comparing opinions
as ‘the size of xyz’. Though they mean the same, they on the web
cannot be directly compared. We may try to solve o Provides a comparative analysis on the
this by adopting stemming or lemmatization [13] that strengths and weaknesses of products in
reduce inflectional forms and sometimes terms of various product features. Uses
derivationally related forms of a word to a common machine learning algorithms.
base form. o The drawback being that it is a comparison
• Now, for the above set of nouns, we identify and methodology that compares products
extract the corresponding adjectives in the sentences. across the web which is not the objective
These adjectives indicate the opinions of the of our project.
customers regarding those specific features and
sometimes the overall product.
• Sentiment Analysis [14,15,16]: It is the process of VI. THE APPLICATION IN PICTURES
computationally identifying and categorizing Fig. 2 shows the application displaying out the results for a
opinions expressed in a piece of text, in order to product searched which is a camera. The underlying features
determine whether the writer’s attitude towards a for example the picture quality, the battery life and their
particular topic, product etc. is positive, negative or respective ratings based on the reviews are displayed.
neutral.
• The analysis can be improved by adopting a machine
learning approach (naïve Bayes classifier) [20] where
the algorithm in trained on a labelled data set.
• Summary generation [17]: The cumulative score of
the opinion based analysis is reduced to a scale of 10
and represented with a rating bar. For the paragraph
summary the overall score of the reviews is
considered and sentences are picked based on the
overall polarity ratio. i.e. the number of positive or
negative sentences picked are proportional to the
overall positive and negative scores respectively.

V. BRIEF STUDY OF OTHER SUMMARIZATION SYSTEMS

We look at the other summarization systems that are
available today in the market with their disadvantages.
• Automatic summarization systems
o Automatically summarizes a single text file
into a paragraph by picking up random Fig. 2 The summarization screen for a camera
sentences.
o The drawbacks of such applications are Fig. 3 shows the application displaying out the results for a
there is no analysis is performed on the product searched which is a Television. The underlying
content and there is no application on features for example the picture quality, the sound quality and
Customer reviews. their respective ratings based on the reviews are displayed.
• Summarization systems that use feature selection
methods

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
2283
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

TABLE I. RESULTS OF SUMMARIZATION

Summarization Results
Sl. No Website URL and Name Feature Based Review
of the Product Rating Summary
http://www.amazon.in/ Very soft
Lifestyle-You-Unisex- on heel.
Height-Increasing- Comforta
Insoles/product- ble to
reviews/B01EG1RDE walk.
Comfortable - 4 /10
E/ref=dpx_acr_txt?sho Height
price - 6 /10
1 wViewpoints=1 increasing
quality - 5 /10
. Poor
functionality - 4 /10
Lifestyle-You Unisex quality.
6 Cm Height Only half
Increasing Shoes part of
Insoles 3 Layers sole can
be used.
http://www.amazon.in/
Formals-Koolpals-
Cotton-Blend-Vertical-
Formal
Stripes/dp/B017FK6G
Shirt at
Fig. 3 The summarization screen for a Television 46/ref=sr_1_14?s=app
cheap
arel&ie=UTF8&qid=1 Comfortable - 6 /10
price.
2 470547277&sr=1- Price - 6 /10
Poor
14&keywords=shirt Quality - 6 /10
quality.
Fig. 4 shows the application displaying out the results for a Smelling
Formals by Koolpals-
product searched which is a phone. The underlying features Cotton Blend Shirt
bad
for example the camera, the display quality and their White Vertical Stripes
respective ratings based on the reviews are displayed. on Dark Blue
http://www.amazon.in/
Zacharias-Men-brown-
leather-
Fitting is
Belt/dp/B01GTQXYU
not
G/ref=sr_1_1?s=appar Size & fitting - 2
perfect.
3 el&ie=UTF8&qid=147 /10
Duplicate
0547478&sr=1- quality - 2/10
product.
1&keywords=belt
Cheap
Men brown leather
Belt
http://www.amazon.in/
Laurels-Black-Wallet-
Combo- Good
Analog/dp/B0154GX3 product as
FI/ref=sr_1_4?s=appar shown.
el&ie=UTF8&qid=147 Belt and
Material - 5 /10
0547539&sr=1- watch is
4 Quality - 6 /10
4&keywords=watch ok. Best
Price - 5 /10
Complime
Laurels Mens Black ntary gift.
Wallet & Belt Combo Watch not
With Black Dial working.
Analog Watch - (Polo-
802-Ast-01)
Fig. 4 The summarization screen for a phone Super nice
quality.
http://www.amazon.in/ Good
AUXTER- value for
AUXTER_bb013456- money.
VII. EXPERIMENTAL RESULTS Auxter-Gym- Long rope
The following table “TABLE I – Results of Duffle/dp/B01BHQCB Material - 7 /10 to make
Summarization” shows the results obtained from the U2/ref=lp_291743103 Roomy - 6 /10 bag as
5
1_1_10?s=luggage&ie Dimensions - 7/ 10 sling.
processing of the listed products on the amazon.in website. It =UTF8&qid=1470548 Price - 8/10 Nice gym
includes the link of the product, name of the product, overall 018&sr=1-10 bag.
feature based rating and review summary. Difficult
Auxter Gym Duffle to carry
Bag when
riding a
bike.

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
2284
 2016 Intl. Conference on Advances in Computing, Communications and Informatics (ICACCI), Sept. 21-24, 2016, Jaipur, India

Summarization Results National University of Singapore Computing 1, 13 Computing

Sl. No Website URL and Name Feature Based Review Drive Singapore 117417. [Online]. Available:
of the Product Rating Summary https://www.comp.nus.edu.sg/~sugiyama/papers/p311.pdf
Budget
bag. [8] Comprehensive Review of Opinion Summarization HYUN DUK
Quality of KIM University of Illinois at Urbana-Champaign KAVITA
stitching GANESAN University of Illinois at Urbana-Champaign
is poor. PARIKSHIT SONDHI University of Illinois at Urbana-
http://www.amazon.in/ Champaign and CHENGXIANG ZHAI University of Illinois at
Levis-Jeans- Urbana-Champaign. [Online]. Available: https://s3-us-west-
6901462423809_1835 2.amazonaws.com/mlsurveys/134.pdf
Really
8-0021_32W-
sober and
34L_Dark- [9] SumView: A Web-based engine for summarizing product reviews
comfortab
Indigo/dp/B0122ZCM and customer opinions Dingding Wang a, Shenghuo Zhu b, Tao Li
Fitting - 8 /10 le.
6 0G/ref=lp_842072403 a֛Ǥ[Online] Available:
Price - 7 /10 Attractive
1_1_14?s=apparel&ie http://users.cis.fiu.edu/~taoli/pub/Sumview.pdf
colour.
=UTF8&qid=1470548
Excellent
340&sr=1-14
product.
[10] The Text Mining Online website. [Online]. Available:
Levi's Men's 511 Slim http://textminingonline.com/dive-into-nltk-part-ii-sentence-
Fit Jeans tokenize-and-word-tokenize

[11] The Text Mining Online website. [Online]. Available:

VIII. CONCLUSIONS http://textminingonline.com/dive-into-nltk-part-iii-part-of-speech-
tagging-and-pos-tagger
After studying the existing systems, we conclude that our
solution provides a more realistic and efficient summarization [12] The NLTK website. [Online]. Available:
of user opinions. Also building an application that uses this http://www.nltk.org/_modules/nltk/probability.html
solution has brought about its use in an apt manner. This can [13] The Text Mining Online website. [Online]. Available:
be scaled into other domains data analytics applications that http://textminingonline.com/dive-into-nltk-part-iv-stemming-and-
concern analysis of raw text data and summarization. lemmatization

[14] The NLTK website. [Online]. Available:

http://www.nltk.org/_modules/nltk/sentiment/vader.html

REFERENCES [15] The NLTK website. [Online]. Available:

http://www.nltk.org/howto/sentiment.html
[1] The University of Illinois Chicago website. [Online]. Available:
https://www.cs.uic.edu/~liub/publications/aaai04- [16] Stanford CoreNLP- a suit of core NLP tools. [Online]. Available:
featureExtract.pdf http://stanfordnlp.github.io/CoreNLP/

[2] The usenix website. [Online]. Available: [17] The Glowing in python website. [Online]. Available:
https://www.usenix.org/legacy/event/wosn10/tech/full_papers/Zha http://glowingpython.blogspot.in/2014/09/text-summarization-
ng.pdf with-nltk.html
[3] The NLTK website. [Online]. Available: http://www.nltk.org/
[18] Web crawling in python. [Online]. Available: http://www-
[4] Mining and Summarizing Customer Reviews. By: Minqing Hu rohan.sdsu.edu/~gawron/python_for_ss/course_core/book_draft/w
and Bing Liu. Department of Computer Science. 851 South eb/web_intro.html
Morgan Street Chicago, IL 60607-7053 {mhu1, liub}@cs.uic.edu
University of Illinois at Chicago. [Online]. Available: [19] An open source and collaborative framework for extracting data.
https://www.cs.uic.edu/~liub/publications/kdd04-revSummary.pdf [Online]. Available: http://scrapy.org/

[5] Feature Based Summarization of Customers' Reviews of Online [20] supervised learning algorithms based on applying Bayes’ theorem
Products. By: Kushal Bafna, Durga Toshniwal M. Tech. Computer with the “naive” assumption of independence between every pair
Science, Electronics and Computer Engineering Dept., IIT of features. [Online]. Available:
Roorkee-247667, India Associate Professor, Electronics and http://scikit-learn.org/stable/modules/naive_bayes.html
Computer Engineering Dept., IIT Roorkee-247667, India. [Online].
Available:
http://www.sciencedirect.com/science/article/pii/S1877050913008
831

[6] Feature-Based Customer Review Summarization. By: Alessandro

Maisto, Serena Pelosi. [Online]. Available:
http://link.springer.com/chapter/10.1007%2F978-3-662-45550-
0_30

[7] Product Review Summarization from a Deeper Perspective. By:

Duy Khang Ly, Kazunari Sugiyama, Ziheng Lin, Min-Yen Kan.

Authorized licensed use limited to: PES University Bengaluru. Downloaded on February 10,2024 at 05:55:47 UTC from IEEE Xplore. Restrictions apply.
2285