international journal of pure and applied mathematics volume … › hub › 2018-118-22 ›...
TRANSCRIPT
DATA ANALYSIS USING DEEP NEURAL NETWORKS
1Kushal Chakraborty,
2D. Malathi
1,2Dept. of Computer Science and Engineering
SRM Institute of Science and Technology, Chennai, India [email protected], [email protected]
Abstract: Social media has become very popular
especially in this modern age of Internet. It is a large
reserve of opinionated data. Nowadays person not only
use social media to gain information but also to provide their opinions, reviews about a multitude variety of
topics. Data Analysis is defined as the systematic,
objective and exhaustive search for the study of the
data and facts relevant to any real world problems. It is
the process which is used to inspect, clean, transform
and model data with the purpose of finding useful
information, suggesting conclusions, and supporting
decision making. It has various techniques and
approaches. Sentiment analysis and opinion mining is
one of the approaches that enables us to determine the
overall view or opinion that is held or expressed by the
people regarding any product, movie or any other topic.
Data Analysis also includes analysis of various other aspects of data like the locations in the world where the
people have talked about the topic, number of people
who have been and are still talking about the topic
(which indicates the topic's popularity) and many other.
The utility of data analysis is innumerable. It enables
the companies to perform research on the market, sales,
products, enables to find out the responses of the public regarding particular policies implemented by
government. In this paper we have performed a detailed
analysis of the tweets relevant to a particular topic of
interest using Deep Neural Networks and provide a
comprehensive analytical solution about the entity.
Keywords: Data Analysis; Sentiment; Opinion; Recursive Neural Network; Emoticon; Hashtag; Entity;
Locations.
1. Introduction
Sentiment Analysis is the field of study that employs
comprehensive text analysis, computational linguistics
and accepted natural language processing techniques to
analyze people’s opinions, sentiments, evaluations,
attitudes and emotions and to identify, quantify and
study affective states and subjective information. It is
applied to the voice of the customer materials towards
entities such as products, services, organizations, individuals, issues, events and their attributes. It
emphasizes on the statement "What is the psychology of
the people? “. In this paper, we have implemented Deep
Learning Neural Networks to solve the following
problem: Given by the user query about:
• The quality of a particular product
• The rating about a particular movie
• The analysis of the upcoming elections
• The analysis of new policies of the government
The objectives of our research work are:
• Is to find the sentiment about the reviews of the
public regarding a particular entity of interest i.e.
whether it is very negative, negative, neutral, positive,
and very positive.
• Is to find the locations in the world where the topic
has been mostly spoken about.
• Is to find all the relevant hashtags.
• Is to find all the topics relevant to our topic of
interest.
• Is to find the number of people who have spoken
about the topic at a particular point of time which indicates the popularity of the topic.
There are various terminologies associated with the
sentiment analysis which has very subtle differences:
• Opinion: Judgement formed about something
• Sentiment: A view or opinion about something
• Evaluation: Assessment of an entity’s worth or
significance
• Appraisal: An act of assessing something or
someone
• Attitude: A way of thinking or feeling about
something
• Emotions: A feeling derived from one’s
circumstances
2. Various Techniques of Sentiment Analysis
Various techniques of sentiment analysis found in the
literature survey are given in Table 1. Many authors
have implemented various techniques at sentence level.
The accuracy obtained and limitations are also given in
the table. The authors in [3] have used the existing
linguistic features as well as resources to find out the
information from the informal languages used in
International Journal of Pure and Applied MathematicsVolume 118 No. 22 2018, 177-187ISSN: 1314-3395 (on-line version)url: http://acadpubl.eu/hubSpecial Issue ijpam.eu
177
Twitter. They have used the supervised machine
learning approaches to find the solution to the problem.
Apoorv Agarwal et al. [5] have used tree kernel and
feature based models to implement sentiment analysis.
They have used a supervised machine learning
technique to solve the task. They have classified the
sentiment into three classes: positive, negative and
neutral. The authors in [6] have provided a technique for
automatically classifying the sentiment of the Twitter
messages based on a keyword. They have classified the sentiments into two classes: positive and negative.
In paper [9], the author has proposed a technique to
provide a synopsis of the comments of some products
based on the votes given by the customer. It provides
the ratings of the essential aspects of so that the
customer can have different viewpoints of the target
product. The authors in [10] have developed an Android
app for polarity analysis of the reviews and comments.
They have used different steps to perform this task like:
acquisition of reviews, polarity and feature
identification, parsing. The detection of sarcastic
sentences is discussed in [11]. For sentiment analysis it
is very important to find out the sarcasm in sentences if
present. The authors have discussed about the various approaches for sarcasm detection like: Rule based
approaches, Deep learning techniques.
Table 1. Various Techniques of Sentiment Analysis
Author Technique/
Approach
Dataset Accuracy/Drawback
Khan et al [17] Rule based Method/
Sentence level
1000 reviews each on
movies, airlines and
2600 reviews on
hotels
91% at document level and 86%
at sentence level/Based on
WordNet Database.
Ana C.E.S Lima
et al [1]
Emoticon, Word,
Hybrid based approaches/
Sentence level
Tweets related
Brazilian TV shows
90% average/
Criteria used to surmise the sentiments are static in nature.
Samaneh
Moghaddam et al
[20]
ILDA/
Document level
Various reviewing
websites that use
rating
73%/
Correspondence between
identified clusters and ratings is not explicit
Jorge Carrilo et al
[19]
Machine learning/
Document and sentence
level
25 reviews/ hotel from
60 different hotels
from booking.com
71.7% for 3 categories and
46.9% for 5 categories/
Not applicable on reviews written
in languages other than English
Samaneh
Moghaddam et al
[18]
Opinion Digger/
Sentence level
Reviews from rating
websites like
Amazon.com
Ranking loss of 0.49/
Requires guidelines and known
aspects to work on and based on
WordNet Database
Tai et al [23] Dependency tree –
Long Short Term
Memory/
Sentence level
Stanford Sentiment
Treebank dataset
48.4%/
N/A
Ouyang et al [16] Convolution Neural
Network/
Sentence level
Movie reviews from
rottentomatoes.org
45.4%/
N/A
In paper [12], the authors have done a detailed
analysis on hashtags and memes which have become
very important components of tweets nowadays. They
have revealed some very interesting facts about some
expected and non-expected hashtags. In the paper [13], the authors have discussed about the approaches of
using sentiment analysis to determine whether the end
user is a bot or a human. They have used SentiBot
framework for this purpose and have used the India
Election Dataset for this purpose.
International Journal of Pure and Applied Mathematics Special Issue
178
The authors in [14] have used the sentiment
predictions from some websites as noisy labels in order
to train a model. They have used 1000 tweets, which
were manually labelled, for tuning. They have used
another 1000 manually labelled tweets for testing
purpose.
A technique for aspect level sentiment analysis of
different entities is proposed in [2]. The authors have
also discussed about the various levels of sentiment
analysis which are being used nowadays. They have also discussed about the various challenges which are
faced by researchers in the field of sentiment analysis.
3. Aspect Level Sentiment Analysis Algorithm
We have proposed a system that performs sentence level
sentiment analysis on tweets and categorizes the tweets into five categories based on the score:
• 0 -Very Negative or Negative
• 1 – Somewhat Negative
• 2 – Neutral
• 3 – Somewhat Positive
• 4 – Positive
A. Twitter
Twitter is a social networking service and
microblogging platform that allows end users to post
real-time brief concise messages called tweets. There
are various characteristics of tweets:
• The maximum length of the twitter message is 140
characters, although experiments are still carried out by
Twitter to increase the length to 280 characters.
• The magnitude of data that is available in twitter is
vast. Using Twitter API and twitter4j it is very easy to
collect millions of tweets for our experiment.
• Twitter users post messages from different types of
devices. The number of errors, misspellings, non-
English words and slangs used are huge which makes it
absolutely necessary to carry out preprocessing tasks.
• Twitter users use brief, concise messages about a
variety of topics which are freely available in Twitter.
This is very effective for our data analysis project.
Following is a brief terminology about the various
components of tweets:
• Handle: A twitter handle is a username preceded by
“@” used to refer users or other users on the blog which
alerts them. It must contain less than 15 characters.
Each handle has a distinct URL with the handle
concatenated after twitter.com. These are also
sometimes referred to as the “target”.
• Emoticon: These are the pictorial facial expressions
which are generally represented by the concatenation of
punctuation and letters which expresses the user’s
sentiment or mood. These are very essential from the
sentiment analysis point of view.
• Hashtag: It is used to mark or refer to a topic,
keyword or phrase preceded by “#” symbol. It is used to
categorize messages and find out the relevant topics on
Twitter.
• Timeline: It actually shows a list of tweets which
are updated dynamically in such a way that the most
recent tweet is displayed at the top.
• Retweet: It is a common activity in Twitter which
shows the extent of popularity of tweet. In this case the tweets are generally forwarded or resent by someone to
the followers or others, although the tweet was
originally written by someone else.
Example: “PM Shri @narendramodi congratulates
@isro team for successfully launching of 100 satellites
in a single mission. #National Youth Day”.
This is a tweet taken from the Timeline of BJP in Twitter. Here there are two handles
“narendramodi”,”isro”. There is a hashtag “National
Youth Day”. So the synopsis is that here the sentiment
is positive. The topic the tweet is referring to is about
the National Youth Day. Here the entity is “Satellites”.
B. Deep Neural Network
The deep neural network which we have used in this
project is the Recursive Neural Network. In this we
recursively apply the same set of weights over an input
that is structured, to provide a prediction that is
structured, over an input that is variable-sized by using
topological order for its traversal. In its most simple
form the Recursive Neural Network can be represented
as follows:
As shown in Fig. 1, let v1 and v2 be the vectors with
n-dimensions each and W be the weight matrix which is
nx2n. Here v1,and v2 are the child nodes. The parent
vector pv can be calculated as: pv = func (W [v1 v2]
T) (1)
Here pv is also a vector of n-dimensions.
score = Ws T
.pv (2)
Here Ws € R1xn
.
Figure 1. A Simple Architecture of Recursive
Neural Network
International Journal of Pure and Applied Mathematics Special Issue
179
Figure 2. Architecture of Recursive
Neural Network
Fig. 2 shows the architecture of Recursive Neural
Network. There are 140 nodes in the input layer and 8
hidden layers and finally the output layer which will
provide the final sentiment of the sentence. The process
used for calculating the sentiment is given in the next
section.
C. Sentiment Analysis
In this paper, we have used Recursive Neural Network
for sentiment analysis. Here we have used the concept
of semantic vector space and compositionality.
Semantic vector spaces consist of the technique of
converting each unique word into vectors where each
element in the vector is used to capture the various
contexts in which a particular word has been used in the
corpus. This can be achieved using Word2vec or Glove
tools. Although it is able to detect those words that
share common contexts, but it cannot capture the
semantics of longer phrases or sentences. Example:
Suppose we have two sentences:
• The country of my birth
• The place where I was born
Now if we closely observe the two sentences we
will see that both the sentences convey almost the same meaning, although the words used in both the sentences
are different. If we take the word vectors of each word
of the two sentences we will not be able to deduce the
meaning of the sentences. This is the reason why we
move to another technique called compositionality.
Compositionality is a technique to deduce the meaning
of longer phrases. And it is shown in Fig. 3. It works in
the following way:
• Each sentence is given to compositional model.
• It is then represented as a binary tree
• By using different types of compositionality
function Recursive Neural Network will compute the
parent vectors in a bottom up approach.
• Now we give these parent vectors as features to
our classifier for the sentiment classification process.
• Softmax classifier is used for sentiment
classification i.e. giving a sentiment tag to a vector.
• Tanh and sigmoid functions are used as activation
functions to compute the parent vectors.
Figure 3. Computing parent vector and sentiment
of a trigram func(b,c), func(a,p1) are compositionality
function. pa1,pa2 are the parent vectors. Sentiment
score obtained at each step by using softmax classifier.
Figure 4. Comprehensive diagram to show the process
of computing parent vectors using Recursive Neural
Network
The Comprehensive diagram to show the process of computing parent vectors using Recursive Neural
Network is shown in Fig 2. The recursive formula for
computing the parent vector can be written as
ht = tanh(wh.ht-1 + wx.x + b) (3)
It can also be written as
ht = tanh([wh wx] [ht-1 x]T
+ b) (4)
where wh is the weight matrix associated with the
phrase, wx is the weight matrix associated with a
word, ht-1 is the vector associated with the previously
computed phrase, ht is the vector associated with
currently computed phrase, b is the bias.
International Journal of Pure and Applied Mathematics Special Issue
180
This process can be repeated again and again by
using Recursive Neural Network to find out the
sentiment of the entire sentence.
4. Data Analysis System Design
In this section we have provided the flow diagram of the
system which we have used for data analytics in Fig 5.
We have also provided the sequence of execution of our
proposed model in Fig. 6.
A. Flow Diagram of the system
Figure 5. Twitter Data Analysis Project Block Diagram
B. Steps of execution of the system
In Fig. 5 we have shown the block diagram of our
project which we have used for data analysis. It consists
of mainly two modules: Webapp and Data fetch API.
The Webapp constitutes the front-end of the project and
Data fetch API constitutes the backend of the project.
The proposed system executes in the following steps:
• User logs into our webapp using the credentials of
twitter. It is mandatory that the person who wants to use
our app for data analytics purpose needs to have an
account in twitter.
• After logging into his account successfully, the
user needs to create an entity with a particular name.
This is the place where all the analytical solution about
a particular product will be stored.
• Now the user needs to provide a keyword or
keywords about which he/she needs to find the
solution.
• Now all the tweets relevant to the particular
keyword will be taken from Twitter Stream API and
will be enqueued in RabbitMQ.
• After this all the tweets will be given to the
analytics module where the real analysis takes place.
• The result of the analysis will be segregated and
the required data will be stored in database in their
respective tables.
• A comprehensive analytical solution of our topic
of interest will then be displayed to the user.
5. Experiments and Result Discussion
A. Dataset
Large volume of datasets of tweets is not freely
available. Therefore we train our neural network using
the public dataset available at
http://nlp.stanford.edu/sentiment. This dataset contains
corpus of movie review excerpts from
rottentomatoes.com website. It contains 10662
sentences of movie reviews out of which one half was
positive and the other half was negative in the original
dataset. The label can give us the net sentiment of the
long movie review. R.Socher in 2013 wanted to achieve
5-class classification of the sentiment i.e. somewhat
negative, negative, neutral, somewhat positive, positive.
He had used the Amazon Mechanical Turk to relabel the
sentiments in the original dataset. The following are the
details of the dataset:
• Number of classification – 5
• Maximum sentence length – 53
• Number of sentences in the dataset used – 11855
• Size of the Vocabulary – 17833
• Number of words present in the Google News word
vector – 16262
• Number of sentences in the test set – 2210
B. Data Collection and Preprocessing
Tweets are collected using Twitter Stream API. We
have used Twitter4J in our project. It is an unofficial
library written in java for the Twitter API. We have
integrated our Web application with the Twitter service
using Twitter4J. We have integrated it into our project using maven build tool.
The twitter data after being collected, has been
preprocessed using Stanford Core Nlp toolkit[7]. It
contains various annotators which performs the
preprocessing tasks. Some of the important annotators
which we have used in our project are:
• tokenize: tokenizes the text into sequences of
smaller units called tokens. For English it uses PTB-
style tokenizer.
• ssplit: it is used to split the sequence of tokens into
sentences.
• pos: it is used to provide the part-of-speech of the
tokens.
• ner: it is used to recognize various types of entities
like : person, location, organization, money, date etc.
International Journal of Pure and Applied Mathematics Special Issue
181
C. Word2vec
Word2vec model was first given by Mokolov et al and a
group of scientists at Google. It is an open source tool
which is used to learn the represent the words in vector
form. It strictly adheres to Apache License 2.0 open
source license. This method of representation is also
called “word embeddings”. In terms of distribution
representation it is written as: “Any word wi in the
corpus is given a distributional representation by an embedding wi € Rd i.e. a d-dimensional vector which is
usually learnt”. It consists of a shallow neural network
which is two-layered. These neural networks are trained
so that it can reconstruct the lexical contexts of words.
A large corpus of text is given as input to Word2vec. As
a result of which a vector space is produced as output.
This vector space is generally of very high dimensions, such that each unique word in the corpus is assigned a
vector in the space. When these word vectors are
positioned in the vector space, it is found that words that
have similar contexts in the corpus are located very
close to each other in space. Internally it uses mainly
two algorithms or models to produce word embeddings.
They are: Continuous Bag of Words Model, Skip Gram
Model. An important advantage of Word2vec is that it
takes less time to execute even on large datasets.
In this paper, we use the pre-trained vectors from
the Google News dataset which consists of about 100
billion words. The vectors are freely available and can
be easily downloaded from https://code.google.com/p/word2vec/. It is contained in
the file GoogleNewsvectors-negative300.bin. It consists
of 300 dimensional vectors for 3 million words and
phrases. It is very difficult to work with such a high
dimensional vector, so we implement a famous
technique of dimensionality reduction, mostly used in
Data Science, called t-Distributed Stochastic Neighbour Embedding (t—SNE).
D. t-Distributed Stochastic Neighbour Embedding
It is a prize-winning machine learning technique for
which is used to reduce the dimensions. The main
problem with Word2vec is that for each unique word it produces a vector of very high dimensions. Since it is
not possible to visualize and work with the vectors in
ridiculously high dimensional space, we use t-SNE for
visualizing the high dimensional vectors on a two or
three dimensional space. It was developed by Geoffrey
Hinton and Laurens van der Maaten[8].
It consists of mainly two steps:
1) It creates a probabilistic distribution over pairs of
words which are of very high dimensions. Here high
dimensional words actually refer to high dimensional
vectors that represent each word. It is done in such a
way that words with similar vectors have high
probability of being selected, while words with dissimilar vectors have an extremely small probability
of getting selected. For the similarity metric, the
Euclidean distance is used.
2) It defines a probabilistic distribution over the
points in map which is of low dimensions. Its main
objective is to minimize the Kullback-Leibler
divergence between the two distributions with respect
to the location of the points in the map.
E. RabbitMQ It is an open source message broker. It accepts and
forwards messages. The main data structure used inside
RabbitMQ is queue. It can be said to be a collection of
software programs which provides various
functionalities required to access a queue. The program
that sends the messages to the queue is called the
producer. The program which removes the messages from queue, forwards the messages is called the
consumer. The addition and removal of the messages
from the queue follows the FIFO principle. The features
of RabbitMQ are:
• It is highly reliable
• It provides routing which is highly flexible.
• It provides clustering.
• It provides highly available queues.
In the propose model, we have used RabbitMQ
technology to store all the tweets relevant to our entity of interest and later the analysis module will use the
tweets from RabbitMQ to provide analytical solution
about our entity of interest.
F. Screenshots of the system
Figure 6. Entity Creation
International Journal of Pure and Applied Mathematics Special Issue
182
Figure 7. Tweet Collection through RabbitMQ
Figure 8. Most Retweeted Tweets
Figure 9. Relevant Topics
International Journal of Pure and Applied Mathematics Special Issue
183
Figure 10. Relevant Hashtags
Figure 11. X-axis: Date, Y-axis: Number of
people talking
Figure 12. X-axis: Date, Y-axis: Sentiment score
Figure 13. Green spots represent hotspots which
indicates the locations where people have talked about
the keyword.
In this section we have showed the screenshots of
our project. In Fig. 6 we have created an entity named
BJP. This entity actually refers to the place where the
statistics related to the data analysis will be stored. Here
we also mention the keywords about which we want a
comprehensive data analysis. The keywords mentioned
here are: elections and modi. If we want we can mention
the handles which will increase the chances of receiving
related tweets faster.
In Fig. 7 the screenshot shows that the tweets are collected through RabbitMQ. In the first graph, the X-
axis represents the time and the Y-axis represents the
number of tweets which have been collected. The
second graph shows the rate at which the tweets are
collected. In Fig. 8 the screenshot shows the most
retweeted tweets along with the number of times the
tweets have been retweeted. This actually indicates the popularity of the topic. In Fig. 9 the screenshot shows
the topics which are relevant or related to the keywords
which we have mentioned while creating the entity.
In Fig. 10 the screenshot shows the relevant
hashtags which we have obtained from the tweets. In
Fig. 11 the screenshot shows the graph which tells us
the number of people who have talked about this topic
at a particular date. The X-axis represents the Date and
the Y-axis represents the number of people. This will
also indicate the popularity of the topic. In Fig. 12 the
screenshot shows the graph which provides the
sentiment score at a particular date. The X-axis
represents the Date and the Y-axis represents the
sentiment score of the tweets. In Fig. 13 the screenshot
shows the geographical locations where people have
talked about the keywords. The green dots represent the
hotspots showing the locations in the world where
people have tweeted about the keyword.
6. Conclusion
Big data and Data Analytics has become one of the
important subjects of Computer Science. It has become
one of the fields in the corporate world. One of the most
International Journal of Pure and Applied Mathematics Special Issue
184
challenging subfields of data analysis is prediction of
trends and sentiment analysis. One of the greatest
challenges in sentiment analysis is detection of sarcasm.
It is really hard to detect the sentiment of a sarcastic
sentence. In future, we will focus on mainly two
aspects: increasing the accuracy of the sentiment
analysis task and detection of sarcastic sentences which
will in turn enable us to provide a greater analytical
solution in our project.
References
[1] Ana c E S Lima and Leandro N de
Castro,”Automatic sentiment Analysis of Twitter
messages”, fourth IEEE conference on CASoN, 21-23
Nov 2012, Sao Carlos.
[2] Kiruthika M, Sanjana Woona, Priyanka Giri,”Sentiment Analysis Of Twitter Data”,
International Journal of Innovations in Engineering and
Technology, Volume 6 Issue 4, April 2016,
ISSN:2319-1058.
[3] Efthymios Kouloumpis, Theresa Wilson,
Johanna Moore,”Twitter Sentiment Analysis: The
Good the Bad and the OMG!”, Proceedings of the Fifth
International AAAI Conference on Weblogs and Social Media, 2011.
[4] Alec Go, Richa Bhayani, Lei Huang,”Twitter
Sentiment Classification using Distant Supervision”,
CS224N Project Report, pp. 1-12, 2009.
[5] Apoorv Agarwal, Boyi Xie, llia Vovsha, Owen
Rambow, Rebecca Passonneau,”Sentiment Analysis of
Twitter Data”, Proceedings of the Workshop on
Language in Social Media (LSM 2011), pages 30-38, Portland, Oregon, 23 June 2011, Association for
Computational Linguistics.
[6] David M. Blei, Andrew Y. Ng, Michael I.
Jordan,”Latent Dirichlet Allocation”, Journal of
Machine Learning Research, Volume 3, 2003.
[7] Cristopher D. Manning, Mihai Surdeanu, John
Bauer, Jenny Finkel, Steven J. Bethard and David
McClosky 2014,”The StanfordCoreNlp: Natural
Language Processing Toolkit”, Association for
Computational Linguistics (ACL) System
Demonstrations, 2014.
[8] L.J.P van der Maaten and G.E. Hinton, “
Visualizing High-Dimensional Data Using t-SNE”,
Journal of Machine Learning Research, Volume 9, Nov
2008,2579-2605.
[9] Yue Lu, ChengXiang Zhai, Neel
Sundaresan,”Rated aspect summarization of short
comments”, In www 2009,pp 131-140.
[10] Shital S. Dabhade, Prof. Sonal S. Honale,”An
Application for Sentiment Analysis Based on
Expressive Feature in the Sentence”, International
Journal of Advance Research in Computer Science and
Management Studies, Volume 3, Issue 5, May 2015,
ISSN:2321-7782.
[11] Aditya Joshi, Pushpak Bhattacharyya, and
Mark J Carman. 2017. “Automatic Sarcasm Detection”
A Survey, ACM Comput. Surv. 0, 0, Article 1000
(2017), 22 pages, DOI:00.00.
[12] Dimitros Kotsakos, Panos Sakkos, Ioannis
Katakis, Dimitrios Guanopulos,”Hashtag : Meme or
Event ?”, IEEE/ACM conference on ASONAM, 17-20 August 2014, Beijing, China.
[13] John P Dickerson, Vadim Kagan, V S
Subramanian,”Using sentiment to detect Bots on
Twitter: Are Humans more opinionated than Bots ?”,
IEEE/ACM conference on ASONAM, 17-20 August
2014, Beijing, China.
[14] Luciano Barbosa and Junlan Feng
2010,”Robust sentiment detection on twitter from
biased and noisy data”, Proceedings of the 23rd
International Conference on Computational Linguistics:
Posters, pages 36-44.
[15] Xi Ouyang, Pan Zhou, Cheng Hua Li, Lijun
Liu,” Sentiment Analysis Using Convolutional Neural
Networks”, 2015 IEEE International Conference on
Computer and Information Technology; Ubiquituous
Computing and Communications; Dependable,
Autonomic and Secure Computing; Pervasive
Intelligence and Computing.
[16] Aurangzeb Khan, Baharum Baharudin, and
Khairullah Khan,”Sentiment classification from online
customer reviews using lexical contextual sentence
structure.”, In Software Engineering and Computer
Systems, ICSECS International Conference on
Software Engineering and Computer Systems, pages
317-331, Springer, 2011.
[17] Samaneh Moghaddam and Martin
Ester,”Opinion digger: an unsupervised opinion miner
from unstructured product reviews”, In Proceedings of
the 19th ACM International Conference on Information
and knowledge management, pages 1825-1828, ACM ,
2010.
[18] Jorge Carrillo de Albornoz, Laura Plaza, Pablo
Gervas and Alberto Diaz,”A joint model of feature
mining and sentiment analysis for product review
rating”, In Advances in Information Retrieval, pages
55-66, Springer, 2011.
[19] Samaneh Moghaddam and Martin Ester,”Ilda:
interdependent lda model for learning latent aspects and their ratings from online product reviews”, In
Proceedings of the 34th International ACM SIGIR
conference on Research and development in
Information Retieval, pages 665-674, ACM, 2011.
International Journal of Pure and Applied Mathematics Special Issue
185
[20] Simon O. Haykin,”Neural Networks and
Learning Machines”, Pearson Education India, Third
Edition, 1 April 2016, 944 pages.
[21] Daniel Jurafsky, Jmaes H. Martin,”Speech and
Language Processing: An Introduction to Natural
Language Processing, Computational Linguistics and
Speech Recognition”, Second Edition, 2013 , 940
pages.
[22] Kai Sheng Tai, Richard Socher, Cristopher D.
Manning,”Improved Sentiment Representations From
Tree-Structured Long Short-Term Memory Networks, In the Proceedings of the 53rd Annual Meeting of the
Association for Computational Linguistics and the 7th
International Joint Conference on natural Language
Processing, pages 1556-1566, Beijing, China, July 26-
31, 2015.
[23] S.V.Manikanthan and T.Padmapriya “Recent
Trends In M2m Communications In 4g Networks And
Evolution Towards 5g”, International Journal of Pure
and Applied Mathematics, ISSN NO:1314-3395, Vol-
115, Issue -8, Sep 2017.
[24] T.Padmapriya and V.Saminadan, “Utility based
Vertical Handoff Decision Model for LTE-A
networks”, International Journal of Computer Science and Information Security, ISSN 1947-5500, vol.14,
no.11, November 2016.
International Journal of Pure and Applied Mathematics Special Issue
186
187
188