IRJET- Classification of Business Reviews using Sentiment Analysis

International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056
Volume: 06 Issue: 06 | June 2019 www.irjet.net p-ISSN: 2395-0072
© 2018, IRJET | Impact Factor value: 7.211 | ISO 9001:2008 Certified Journal | Page 2786
Classification of Business Reviews using Sentiment Analysis
Shilpa Shendre1, Prof. Pramila Chawan2
1Mtech Student, Dept of Computer Engineering and IT, VJTI College Mumbai, Maharashtra, India
2Associate Professor, Dept of Computer and IT Engineering, VJTI College Mumbai, Maharashtra, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - A The rapid increase in mountains of unstructured
textual data accompanied by the proliferation of tools to
analyze them has opened up great opportunities and
challenges for text research. The research area of sentiment
analysis has gained popularity in the last years. Business
developers not only want to know about there product
marketing and profit based on the number of sales been done
but also want to know about the reviews and thoughts of
people using these products. The feedback they receive via
social media and other internet services becomes very
important to measure the quality of a product they are
serving. Sentiment analysis is a domain where the analysis is
focused on the extraction of feedbackandopinionsoftheusers
towards a particular topic from a structured, or unstructured
textual data. In this paper, we try to focus our effort on
sentiment analysis on restaurant review database. We
examine the sentiment expression to classifythereviewsofthe
restaurant business whether it is positive or negative and
perform the feature extraction and use these features for
updating and maintenance of the business.
Key Words: sentiment analysis; opinion mining;
classification; text reviews, Machine learning
1. INTRODUCTION
Sentiment analysis has become an important research area
for understanding people’s opinion on a matter by
differentiating a huge amount of information. The present
era of the Internet has become a huge Cyber Database which
hosts the gigantic amount of data which is created and
consumed by the users. People across the world share their
views about various services or products using social
networking sites, blogs or popular reviews sites. The
Internet is been growing at an exponential rate givingrise to
communicate across the globe in which people expresstheir
views on social media such as Facebook, Twitter, Rotten
Tomatoes and Foursquare. Opinions which are being
expressed in the form of reviews provide a platform for new
explorations to find collective reviews of people. One such
domain of reviews is the domain of business reviews which
affects business people. The feedback from the customer is
valuable for companies to analyze their customer’s
satisfaction and survey the competitors. This is also useful
for other people or consumers who want to buy a product or
a service prior to making a purchase.
In this paper, we are going to present the results of machine
algorithms for classifying reviews using semantic analysis.A
large number of customer-generated reviewsfor businesses
and service providers are classified as either positive or
negative. We propose a method to automatically classify
customer sentiments using only business text review. This
helps us to generate the result using feedback without
manual intervention. By studying only rating, it is very
difficult to judge why the user has rated the productas1or5
stars. However, the textcontentcontainsa morequantitative
value for analyzing more than rating itself.
In this paper, we are going to mention the preprocessing
steps require in order to achieve accuracy in the
classification task. Thereisnopreviousresearchavailableon
classifying sentiment of business review using the latest
reviews forms restaurant dataset. Determining the
underlying sentiment of restaurant business review is a
difficult task taking into account several factors such as the
connotation of a word depending on the context, language
used, words ambiguity when using words that don’t express
a particular sentiment or when using sarcasm.Weshowthat
a sentiment analysis algorithm built on top of machine
learning algorithms such as Naïve Bayes and Linear Support
Vector Classification(SVC)hasaccuracyabove90%business
reviews.
2. WORK REALTED
Hu et al. perform the classification of a document at the
sentence level. Instead of the whole document and feature
extract on which views have been expressed, identifying
comments words by proposing a technique that uses the
WordNet lexical database. For each feature extracted, the
related reviews sentence is stored in positive or negative
categories and computes a total count. The features are
ranked on the bases of there frequency of the appearance in
the reviews. The feature-based summary of the reviews of
the product sold online was provided by the authors.
Usually work related to sentiment analysis using machine
learning techniques in determining if the overall review is
positive or negative movie reviewsasdata.Thewriter'sused
unigram model and Navie Bayes, entropy classification, and
SVM to perform the classification and achieve accuracyupto
80%. They finally concluded that their results outperform
the method based on human tagged features.
A system was built by Blair-Goldensohn et al. which
automatically summarize sentimentfroma setofreviews for
a local service such as restaurant or hotel and combine the
review sentiment per aspect such as food, service, decor,
value etc., Basically they have implemented a custom built

lexicon based on WordNet and used a classifier at the
sentence level.
3. THE PROPOSED METHOD
The basic methodology to determine polarityistheonewith
a lexical approach, where we look at the words comprising
the document and apply some algorithms to quantify words
with some sentiment score and determine the collective
polarity. We have based our computational method on the
publically available library SentiWordNet
In this work for determining the polarity of the reviews , we
have focused on two areas: 1) FeatureSelectionandRanking
2) Classification using Machine Learning techniques. We use
the restaurant review dataset comprising51Mbreviews.We
tend to label the polarity as follows : 0- Strong Negative, 1-
Weak Negative, 2-Neutral, 3-week, Positive, 4-Strong
Positive. The proposed methodology can be well explained
from the below figure.
Fig1. Proposed System
Since SemEval 2016 has independent tasks for
aspect extraction and aspect categorization, we need to add
sentence preprocess before applying aspect categorization.
We split a sentence into two sentences if it has contrary
conjunction, such as “but”, “however”, etc. This preprocess
solves problem when a sentence has two or more aspects
with same category but different sentiment polarity. We
define two rules i.e.: If both clauses beforeandaftercontrary
conjunction contain extracted aspects, we split the sentence
into two sentences by using contrary conjunction as
delimiter. Example such as I like the dessert but I don’t like
the meal is splitted into first: I like the dessert but and
second: I don’t like the meal. If clause before contrary
conjunction contains aspect but clause after contrary
conjunction does not contain aspect, we also split it into two
sentences by using contrary conjunction as delimiter and
inserting aspect from the first sentence into the second
sentence. For example, “food price is expensive but
comparable” is splitted into “food price is expensive” and
“food price is comparable”.
We build multilabel binary relevance classifier with MaxEnt
algorithm for aspectcategorization.Theillustrationof binary
relevance classifier. We define four categories: food, place,
price, and service. Each category has its binary classifier so
the total classifier is same as the total of category. Each
sentence in the corpus is labeled with Booleanvaluefor each
category: true if a sentence has certain category and false
otherwise. The classifier for each category classify the
Boolean value. After that, we collect categories with true
values as multilabel output.
Fig 2.Relevance Classifier for aspect categorization
We apply the same algorithm as in aspect categorization for
sentiment classification. We define two labels of sentiment
for each category in a sentence: positive and negative. For
example, a sentence has food and place category. Each
category has its own classifier to classify sentiment of
categories in the sentence. To classifythesentiment polarity,
we use classifier for food and place category. As in aspect
categorization, we use general features from English. Bag of

clusters is obtained from CBOW, LDA, andGloVe.Weuse 2to
5 words skips for Skip-bigram. N value of 1 and2isalsoused
for bag of N-gram feature.
4. EXPERIMENT, EVALUATION, AND DISCUSSION
In order to use supervised learning and train a classifier, we
usually require a predefined training data, but taking into
account the large range of restaurant businesses and the
large number of reviews, it would be very difficult to
manually annotate the data to train a sentimentclassifier for
reviews.
4.1 FEATURE EXTRACTION
Mostly the researchers apply standard feature selection in
there approach to improve performance with few using
more practical approaches. We are focusing completely on
feature election to improve sentiment analysis are few. One
of them is the famous Pang & Lee, who removed objective
sentences on a testbed consisting of objectiveand subjective
text trained on SVM. Initially, they found that sentiment
classification result is actuallyslowandmoderate. Theythen
concluded it was more likely that sentences adjacent to
discarded sentencesimprovedclassificationresultovertheir
baseline.
For opinion structure generation, we employ CBOW model
to find similarity between extracted aspect with seed words
for each category. Category that has maximum similarity
score will be paired with the extractedaspect.Forexample,a
sentence has food and place category and we want to pair
“cake” with a category from the sentence. First, we find
similarity score for each seed words for food and place
category. The maximum similarity score for food and place
category are 0.2 and 0.1 respectively. After that, we can pair
“cake” with food category because similarity score for food
category is higher than place category. The generated
opinion structures are used to calculate rating for each
category with equation as in.
Feature Definition
Bag of N-grams
The occurrence of a N-gram in the
context
window
Bag of Head
Words
Bag of word that determines the
syntactic
category of that word from the
dependency tree
Bag of Clusters
The occurrence of word’s clusterinthe
context
window
Bag of k-skip-
bigram
Bag of N-gram which has skipped over
gaps.
Table 1. Features For Sentiment Classification
TABLE 1 show token distribution and example of
training data for aspect extraction step. For clusters and
clusters bigram feature, there are four scenarios to
experiment with the number of clusters. The best
combination for this step is bag of N-gram, bag of POS N-
gram, clusters with 5000 clusters, and clusters bigram with
100 clusters CBOW.
4.2 FEATURE CATEGORIZATION
The label distribution in training data and example of
training data for aspect categorization . For bag of clusters
feature, we also use four scenarios and additional six
scenarios to experiment with number of topic for LDA. The
best feature for this step is bag of clusters using CBOW
model with 1000 clusters.
This is the highest accuracy obtained usingthismethod.Also
it’s worth noting that giving equal importance to all factors
i.e. giving each a value of 0.165 has resulted in a lower
accuracy of 78.268% than the highest accuracy obtained by
unequal distribution of factors. Thus by changing the
importance of that aspect, we can see its effect in the
accuracy of the overall classification of the review.
Thus we can interpret from the results that in the reviews
used from the dataset, the user has given more importance
to these factors while writing the review.Italsomeansthatif
the user tends to give a positive review towards these
aspects then, due to their increased importance, the overall
review will tend to be positive even if the user gives a
negative feedback towards the other aspects. Giving more
importance to certain factors also has anaddedadvantage,it
tends to suppress the users opinion about other factors.
Suppose we have a reviewed 'X' and it contains user’s
opinion about two factors F1 and F2. Also the overall
orientation of the review is positive in nature. The user has
given a positive review about F1 and a negative about F2.
Also the amount of text in the review for F1 aspect is less as
compared to the F2 aspect. If we use any non-aspect based
sentiment analysis method then since text size of F2 is
greater than text size of F1 and also since F2 is negative in
orientation, the overall review score will tend to reduce and
skew towards.
The various performance measures used were:
Accuracy = (Total correctly classified word / Total
number of words)
Precision = tp / (tp +fp)
Specificity = (tn / Total number of negatively oriented
review in the dataset)
Recall = (tp / Total number of positively oriented
review in the dataset)
Where tp, fp and tn are the true positives, false positivesand
true negatives obtained during the classification. On the

other hand if driving factors are used and F1 is given more
importance the review score will better reflect the positivity
of the review. Since each aspect of a restaurant is analysed
separately in this method, we cantrack theeffecteachaspect
has towards the overall score of the review. This individual
aspect based tracking can be used in a fined grained aspect
based recommendation system, which recommends
restaurants based on its various aspects instead of the
overall rating of the restaurant. Also this method can be
applied on a product review dataset thus enabling us to see
what opinion each user has on the various aspects of the
product, thus helping in the development of proper product
placement strategy. It is very difficult to acquire such in-
depth knowledge from the dataset using non-aspect based
methods.
Label Total Sentence
Food 503
Service 97
Price 125
Place 440
TABLE 2. DISTRIBUTION OF ASPECT CATEGORIZATION
IN DATA
Some misclassifications occur when the sentence has word
“restaurant” or restaurant name. For example, sentence“we
come to Atmosphere cafe for celebrating our anniversary
with high expectation” is classified as place category while
the sentence actually does not have any category. The
sentence that have word “atmosphere” only appear once in
the training data and it is labeled as place category so the
sentence misclassified as place category. Another
misclassification happens in sentence “this is an old
restaurant that still exists until now”. Thesentencehasword
“restaurant” so it is classified as place categorybythemodel.
But word “restaurant” does not have any sentiment so the
sentence is not labeled as place category even though it has
word “restaurant".
Misclassification can happen if words never co-occur in the
training data as in aspect extraction. For example, sentence
“but the speciality is the environment” is classified as food
category while the sentence actually has place category.
Most sentences have word “special” labeledasfoodcategory
in the training data because it co-occurswithwordrelatedto
food and it never cooccurs with word “environment”.
Besides that, the word "environment" never appears in
training data. Because of that,thesentenceismisclassified as
food category.
Aspect Category Example Seeds
Food Food, beverage, dessert, meal, taste
Service Service, waiter, waitress
Price Price
Place place, atmosphere,
TABLE 2. ASCPECT CATEGORIES
After we have all aspect categories and its aspect, we will
calculate the rating for each aspect categories. The rating
calculation will follow the equation 1.
Rating = (p / (p+n) *4) + 1
Variable P/N is the total of positive/negative opinion in the
aspect category. The rating is scaled in 1 – 5.
Example: The place was comfortable, the view was nice, and
the price was affordable. In my opinion, the foodwasgood, but
the cocktail was not too good, the bartender still has a lot to
learn.
Experiments are conducted for various aspect level
sentiment classification with feature selection methods and
different feature set size. The aim of this analysis is to see 1)
if machine learning algorithms for aspect level sentiment
classification work 2) if the size of the feature set influence
the performance of classification.
Category Sentiment Aspect Rating
Food Positive food 5.00
Negative -
Place Positive - 0.00
Negative -
Price Positive price 5.00
Negative -
Service Positive
place
view
5.00
Negative -
TABLE 4. EXAMPLE OF GENERATED OPINION
STRUCTURES AND RATING
4.3 Aspect and Sentiment Extraction
For token classification, the accuracy of the label is quite
high with 88.48. The results also show that F1-Measure for
OP_NEG_I is quite low while for OTHER is high. Many
misclassifications occurred and the tokens are mostly
classified as OTHER class. This is because of the use of
infrequent words to describe the aspects and the opinions.
Those infrequent words are then classified as OTHER.
TABLE 4. EVALUATION RESULT ON TOKEN
CLASIFICATION
Label Precision Recall F1
ASPECT-B 0.7104 0.7455 0.7275
ASPECT-I 0.4929 0.5475 0.5188
OP_POS-B 0.7524 0.8505 0.7985
OP_POS-I 0.6885 0.8235 0.75
OP_NEG-B 0.6923 0.5373 0.605
OP_NEG-I 0.5926 0.4444 0.5079
OTHER 0.943 0.9243 0.9336
Accuracy 0.8848

One of the example of class ASPECT classified as OTHERisin
the sentence “order grilled carp, the crap slightlyoverburn”.
In that sentence, the aspect is “the crap” however the model
cannot detect the
aspect. It is caused by that word never appeared in training
data. The opinion in that sentence is “slightly over burn”.
Those words also never appearinthedata andalsoclassified
as OTHER.
In contrast, all tokens in the sentence “the place is nice, the
atmosphere is also comfortable, and a lot of variation in the
menu” are correctly classified. The reason behind this is all
of the tokens have occurred in the training data and the
model can easily recognize the pattern in the sentence.
5. CONCLUSION
Sentiment analysis is vast research area and it has wide
variety of issues to bediscussedwithseveral challenges. This
paper explains the aspect basedfeatureselectionmethodsin
combination with other machine learning algorithms. The
results of experiment explain that features or aspects are
selected and iterative classifier using machine learning
technique is proposed. Aspect-based sentimentanalysishas
six steps i.e. preprocess, aspect extraction, aspect
categorization, sentiment classification, opinion structure
generation, and rating calculation. The experiment includes
review data sets which includes positive and negative
aspects. Our method gives best results for precision, recall
and accuracy compared to SVM and naives bayes method.
The proposed method iteratively runs while processing the
data and analyses based on previous experience. The
accuracy increases up to 83.5% the accuracy scaled up to a
great extent. The naïves method scaled up to 78.44% and
SVM scaled up to 80.34 percent. We identify and calculate
the precision, recall and accuracy for the models .It shows
that the proposed method gives better results. Future work
would be to combine different feature selection schemes for
analyzing the accuracy of the review data sets.
References
[1] Kudakwashe Z, Oludayo O., et al. “A Framework for
Sentiment Analysis with Opinion Mining of Hotel
Reviews.2018”. ICTAS
[2] Kulvinder S.,& Sanjeev D., &Pratibha "Real-time Data
Elicitation from Twitter: Evaluation
and Depiction Strategies of Tweets ConcernedtotheBlazing
Issues Through Twitter Application " LREC. Vol. 10. 2010.
[3] Rincy J.,& Varghese S. "Prediction of Election Result by
Enhanced Sentiment Analysis on Data using Classifier
Approach " Computational linguistics 267-307.
[4] Xiaobo Z., &Qingsong Y., "Hotel Reviews Sentiment
Analysis Based on Word Vector Clustering " 2017 2nd IEEE
International Conference on Computational Intelligenceand
Applications.
[5] Wang Z., Qin S., A Sentiment Analysis MethodofChinese
Specialized Field Short Commentary” 2017 3rd IEEE
International Conferenceon ComputerandCommunications
[6] Harpreet K., Veenu M., Nidhi., “A Survey of Sentiment
Analysis techniques International conference on I-SMAC
(IoT in Social, Mobile, Analytics and Cloud) (I-SMAC 2017)
Mining Workshops. IEEE, 2011.
[7] Jie Li; Lirong Qiu, “A Sentiment Analysis Method of
Short Texts in Microblog” in 2017 IEEE International
ConferenceonComputational ScienceandEngineering(CSE)
[8] Kudakwashe Z., Oludayo O., “A Framework for
Sentiment Analysis with OpinionMiningofHotel Reviews”in
2018 (ICTAS)
[9] V. S. Jagtap, Karishma Pawar, “Analysis of different
approaches to Sentence-Level Sentiment”, International
Journal of Scientific Engineering and Technology Volume 2
Issue 3, PP (ISSN : 2277) (2013).
[10] Nidhi Mishra, “Classification of Opinion Mining
Techniques”,IEEE Trans.Syst., International Journal of
Computer Applications (0975 – 8887) Volume 56– No.13,
October 2012.
[11] Aansi A. Kothari and Warish D. Patel,” A Novel
Approach towards Context Sensitive Recommendations
Based On Machine Learning Methodology” 2015 Fifth
International Conference on Communication Systems and
Network Technologies.
[12] A. Mudinas et al, 2012, Combining lexiconandlearning
based approaches for concept-level sentiment analysis,
Proceedings of the First International WorkshoponIssuesof
Sentiment
[13] Discovery and Opinion Mining, ACM, New York, NY,
USA, Article 5, 1-8.
[14] M. RushdiSaleh et al, 2011, Experiments with SVM to
classifyopinions in different domains, Expert Systems with
Applications 38.
[15] Emma Haddi, Xiaohui Liu, Yong Shi, "The Role of Text
PreProcessing In Sentiment Analysis", Information
Technology and Quantitative Management (Itqm2013),
1877-0509, Elsevier © 2013.
[16] T. K. Das, D. P. Acharjya and M. R. Patra: "Business
Intelligence from Online Product Review-A RoughSetBased
Rule Induction
Approach", International Conference on Contemporary
Computing and Informatics (IC3I-2014), November 03 – 05,
2014, Mysore, India, IEEE Xplore, pp. 800 – 803, 2014.

[17] T. K. Das, D. P. Acharjya and M. R. Patra: "Opinion
Mining about a Product by Analyzing Public Tweets in
Twitter", International
Conference on Computer Communication and Informatics
(ICCCI - 2014), January 03 – 05, 2014, Coimbatore, India,
IEEE Xplore, pp. 1- 4, 2014.
[18] D. P. Acharjya and L. Ezhilarasi: "A Knowledge Mining
Model for Ranking Institutions usingRoughComputing with
Ordering Rules
and Formal Concept Analysis"; International Journal of
Computer Science Issues; Vol. 8 (2), pp. 417 – 425, 2011.
[19] T. K. Das and D. P. Acharjya: "A Decision Making Model
using Soft Set and Rough Set on Fuzzy Approximation
Spaces", International
Journal of Intelligent Systems Technologies and
Applications, Vol. 13 (3), pp. 170-186, 2014.
[20] T. K. Das, D. P. Acharjya and M. R. Patra:"MultiCriterion
Decision Making using Intuitionistic Fuzzy Rough Set on
Two Universal Sets", International Journal of Intelligent
Systems and Applications, Vol. 7 (4), pp. 26-33, 2015.

IRJET- Classification of Business Reviews using Sentiment Analysis

More Related Content

What's hot (19)

Similar to IRJET- Classification of Business Reviews using Sentiment Analysis (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Classification of Business Reviews using Sentiment Analysis