2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS)

Detection of Cyberbullying Using Deep Neural

Network
Vijay Banerjee Jui Telavane Pooja Gaikwad Pallavi Vartak
Computer Engineering Computer Engineering Computer Engineering Computer Engineering
Viva Institute of Technology Viva Institute of Technology Viva Institute of Technology Viva Institute of Technology
Virar, India Virar, India Virar, India Virar, India
[email protected] [email protected] [email protected] [email protected]

Abstract— Innovation is developing quickly today. This commonness, numerous cyberbullying recognition instruments
headways in innovation has changed how individuals have been presented. There are numerous exisiting framework
cooperate in an expansive way giving communication on cyberbullying location and the examination is significantly
another dimension. But despite the fact that innovation more developed yet numerous issues are not tended to yet. Most
encourages us in numerous parts of life, it accompanies of research on cyberbullying location centers around spaces, for
different effects that influence people in a few or the other example, machine learning and data mining. To the extent the
way. Cyberbullying is one of such effects. Cyberbullying is investigation none of the earlier research has tended to the issue
a wrongdoing in which a culprit focuses on an individual of digital assault identified with network, Sexism and Racism
with online provocation and loathe which has antagonistic utilizing profound neural network. In the proposed system, we
emotional, social and physical effects on the victim. So as to use convolution neural network to create a model that detect
address such issue we proposed a novel cyberbullying bully related tweets and predict the behaviour of the new data
detection method dependent on deep neural network. introduced. The proposed approach use word vectors that are
Convolution Neural Network is utilized for the better feed to the CNN for classification of tweets. The dataset of
outcomes when contrasted with the current systems. chats and tweets form various social media platforms is
collected for the evaluation.
Keywords— cyberbullying; detection; deep learning;
II. RELATED WORKS
convolutional neural network; word embedding, Glove.
An detail literature survey conducted across IEEE explorer
I. INTRODUCTION and Springer digital library. The main search strategy was the
As the Technology advances people rely on the technolgy to discovery of academic literature relavent to the theme
conductmany of the daily activities including communication. “Techniques of cyberbullying detection”. The following is the
The social media platforms are thus a great platform connecting detailed literature work.
people all over the world. But this promising shift towards Rui Zhao and Kezhi Mao [5] used a new representation
digital world comes with a pricy cost. The secrecy over the web learning method. This method is Semantic-Enhanced
has made it less demanding for individuals to post audit remarks Marginalized Denoising Auto-Encoder (smSDA) developed
on various issues without being identified, a precedent is found via semantic extension of the popular deep learning model
in the informal organization twitter. In this manner the client stacked denoising autoencoder. Elaheh Raisi and Bert Huang
can post tweets which can assault different clients, and these [6] proposed a weakly supervised machine learning technique
tweets are untraceable. This training prompts cyberbullying. for all the while surmising user rolls in provocation based
Cyberbullying is characterized as any fierce, purposeful activity bullying and new vocabulary markers of harassment. P. Zhou,
directed by people or gatherings, utilizing on the web channels et. al. [7] proposed attention based B-LSTM technique, this can
over and again against a victim who does not can possibly automatically concentrate on the words that have conclusive
react[1]. Numerous investigations have tended to cyberbullying impact on classification, to catch the most imperative semantic
with point of surveying its commonness, and results information in a sentence, without utilizing additional
demonstrated that cyberbullying is a typical issue confronting knowledge and NLP frameworks. A. Conneau, et. al. [8]
the present age and that the quantity of victims is rising[2][3]. presents a new architecture (VD-CNN) for text processing
So as to help in controlling cyberbullying and restricting its which works specifically at the character level and uses as it

604

978-1-5386-9533-3/19/$31.00 ©2019 IEEE

 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS)

were small convolutions and pooling operations. A. System Flow

S. Bhoir, et. al. [9] presented a similar investigation of For a proposed solution the flow diagram is as follows:
various word embedding techniques based on various
parameters. classifier used and effect of changes in
dimensionality. E. Raisi, et. al. [10] presented the participant
vocabulary consistent model a weakly supervised methodology
for concomitantly learning the roles of online social users in the
badgering type of cyberbullying. H. Zeng, et. al. [11] utilized a
representation procedure that has 4 connected view that helps
to investigate learning parameters. It is important to get the
knowledge of how the model parameters advance from lower fig 1. 1 : System flow
to higher precision so we can improve the training procedure.
B. Word Embedding
V. N. Kumar, et. al. [12] This paper uses naïve Bayes as the
classifier for the content classification in email application it Starting word embeddings choose information portrayal for
deals with the classification of spam words when massage is Neural Network models. In any case, amid the training, Neural
received and it is processed using feature set extraction method Network model adjust these underlying word embeddings to
Andrew M. Dal and Quoc V. Le [13] proposed sequence learn assignment specific word embeddings. For our proposed
learning supervised model using CNN and LSTM.K. Duan, et. system we experimented with different word embedding
al. [14] explains SoftMax combination for multicategory methods and used Glove for our project.
classification both one-versus-all and one-versus-one classifier. C. Architecture
This paper explains how to efficiently extend binary
The very next layer from the embedding layer (if there
classification method for multi-category classification. Q. Li,
should arise an occurrence) is the convolutional layer. This is
et. al. [15] proposed a new tweet sentiment classification
the main core layer of the whole convolution neural network. In
approach using SSWE and WTFM produce classes based on the
our project we have used different layers of convolution as
weighting scheme and text negation and a new text
follows:
classification method. I. Raid [16] carried out research using
data mining technique here there are several stages as data • Input
collection, preprocessing, TF-IDF, weighting, data validation • Word Embedding layer
and classification using naive Bayes classifier. K. Sahay, et. al. • Dropout layer(0.25)
[17] explains the training in machine learning model using • Convolution layer
supervised learning. • Max Pooling
III. PROPOSED METHODOLOGY • Convolution-layer
• Max Pooling
CNN is a Type of Deep Neural Network Model that
• Convolution layer
comprises of neurons. Neurons work on biases and learnable
• Max Pooling
weights. The architecture of CNN comprises of an input, output
• Dropout layer(0.5)
layer and a few of hidden layers.The input layer includes
succession of vectors. It is examined utilizing fixed size of • Fully connected layer
filter. The filter shifts or strides only one row or one column on • Softmax
the matrix. Each filter distinguishes different features in the • Classification layer
content so as to portray it into the feature map the next layer is Input layer:
maxpooling layer. The maxpooling layer minimizes the in this layer data from the tweeter dataset is given to the
features in the feature map. It selects the highest value in the embedding model.
feature map to capture most important feature. Due to that, it Embedding layer:
decrease the calculation in the propelled layers the dropout this word embedding is done for words into vector
strategy is connected to diminish overfitting with dropout rate conversion. Neural Network only takes numerical value as
is 0 5. The last layer of the model is Dense(FullyConnected) input so, tokens that is words are converted into vectors thus it
layer and it is used for classification purpose. This layer each represents text as a row of vectors Each word in the text,
classifies the text based on the classes specified. which is one token is embedded into a vector. This step is a

605
 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS)

matrix of size m x n, where n is the length of the vector and m Fully Connected Layer
is the number of tokens in the texts. The last layer of the model is Dense (FullyConnected) layer
Dropout layer: and it is utilized for classification reason. This layer groups the
Dropout is a regularization technique that approximates content dependent on the classes determined.
training an expansive number of neural networks with various
IV. IMPLEMENTATION
structures in parallel. In proposed arrangement dropout layers
are utilized subsequent to embedding and before fully The proposed system is implemented in Python and
connected layer for regularization. tensorflow. Tensorflow is a high-performance computing
Convolutional Layer framework which is widely used in research, development and
Each input contains an arrangement of vectors. It is checked analysis in the fields of data science and deeplearning. The
utilizing fixed size of filter. The filter moves or strides just a Twitter dataset used consists of 69874 tweets, which are
single row or one column over the matrix. Each filter identifies converted to vectors using open source wordembedding Glove.
various features in the content so as to speak to in the element These messages were sorted and labels were generated.
map. Neural Network model revealed here were implemented
Rectified Linear Unit (ReLU) utilizing Keras on top of tensorflow. We preprocess the data,
Layer applies a max function f(x)=max(x,0) to the matrix of exposing it to standard tasks of expulsion of stop words,
the convolved after convolution. It sets all the negative qualities accentuation marks and lowercasing, before clarifying it to
in the dot products of the matrix to 0. Every other esteem are allocating individual labels to each remark.
unaltered. It builds the speed of training the network by
IV. RESULT AND ANALYSIS
evacuating negative activation in the slope, consequently
staying away from complex negative computations. The below Table 1, shows the results of the various existing
Max-Pooling Layer cyberbullying detection systems based on data mining, machine
learning and RNN based deep learning system. The table shows
After convolutional layer, the following layer is Max-
that the proposed system derives better results than all the
Pooling layer. The Max-Pooling layer limits the features in the systems referred so far.
Feature map. It chooses the most noteworthy incentive feature
in the feature map to catch most critical feature.
Title of paper Techniques used Accuracy
The proposed system Architecture is as follows:
Optimized Twitter
Cyberbullying
RNN, GloVe 81.60%
Detection based on
Deep Learning [4]

Detecting cyberbullying Natural Language

and aggression in social Processing and 75%
commentary [17] Machine Learning

Detection of
cyberbullying on social
Data Mining 75%
media using data
mining techniques [16]
Detection of
Cyberbullying using CNN Testing
deep neural (Convolution Accuracy-
network(Proposed Neural Network) 93.97%
System)

TABLE 1. RESULT ANALYSIS

VI. CONCLUSION AND FUTURE SCOPE

Rapid growth of technology is affecting the way we
fig 1. 2 Model Architecture communicate on the social media platforms resulting in

606
 2019 5th International Conference on Advanced Computing & Communication Systems (ICACCS)

cyberbullying and many such issues. although many 9. MS.Snehal Bhoir, Tushar Ghorpade, Vanita Mane
researches addressed cyberbullying in SMP(social medai “Comparative Analysis of Different Word Embedding
platform), The techniques used for the detection proves Models” IEEE,2017.
inefficient in classification. In the proposed system, we 10. Elaheh Raisis,Bert Huang “Cyberbullying Detection with
represent a new approach for the detection of cyberbullying. Weakly Supervised Machine Learning” International
This system uses convolution neural network algorithm which Conference on Advances in Social Networks Analysis
operates through many layers and gives accurate classification. andMining IEEE/ACM,2017.
Thus a more intelligent way, compared to the traditional 11. Haipeng Zeng, Hammad Haleem, Xavier Plantaz, NanCao
classification algorithms is designed. and Huamin Qu “CNN Comparator: Comparative
Future scope includes regressive training of system so as to Analytics of CNN” arXiv,15 Oct,2017.
detect cyberbullying in real time chats and also the detection of 12. Vandana NandaKumar,Binsu C,Kovoor,Sreeja M.U
cyberbullying in chats containing Hinglish(hindi and english) “Cyber-Bullying Revelation in Twitter Data using Naive-
code mix languag. Bayes Classifier Algorithm” International Journal of
Advanced Research in Computer Science. Volume 9, No.
REFERENCES
Jan-Feb 2018.
13. Andrew M.Dal,Quoc V.Le “Semi-Supervised Sequence
1. R. Shetgiri, "Bullying and Victimization Among Learning “ arXiv,4 Nov 2015.
Children", Advances in Pediatrics, vol.60, no. 1, pp. 33-51, 14. Kaiob Duan, S.Sathiya Keerthi,Wei Chu, Shirish Krishnaj
2013. Shevade and Anu Neow Poo“Multi-Category
2. Brown, E. Clery and C. Ferguson, "Estimating the Classification by Softmax Combination of Binary
prevalence of young people absent from school due to Classifiers” Department of Computer Science and
bullying," National Center for Social Research, 2011. Automation, Bangalore.
3. Van Royen, K. Poels, W. Daelemans and H. Vandebosch, 15. Quanzhi Li, Sameena Shah, Rui Fang, Armineh
"Automatic monitoring of cyberbullying on social Nourbakhsh, Xiaomo Liu “Tweet Sentiment Analysis by
networking sites: From technological feasibility to Incorporating Sentiment Specific Word Embedding and
desirability" Telematics and Informatics, vol. 32, no.1, Weighted Text Features” International Conference on Web
pp.8997, 2015. Intelligence IEEE/WIC/ACM,2016.
4. Monirah A. Al-Ajlan, Mourad Ykhlef, “Optimized Twitter 16. Hariani,Imam Raid “Detection of Cyberbullying on Social
Cyberbullying Detection based on Deep Learning”, 978-1- Media using Data Mining Techniques” International
5386-4110-1, IEEE-2018. Journal of Computer Science and Information Security,
5. Rui Zhao,Kezhi Mao “CyberBullying Detection based on Vol.15, No.3, March 2017.
Semantic-Enhance Marginalize Denoising Auto- 17. Kahitiz Sahay, Harsimran Singh Khaira,Prince Kukreja,
encoders”IEEE Transaction on Affective Computing, Nishchay Shukla “Detecting Cyberbullying and
2015. Aggression in Social Commentary using NLP and Machine
6. Elaheh Raisi, Bert Huang “Weakly Supervised Learning” International Journal of Engineering
Cyberbullying Detection with Participant Vocabulary Technology Science and Research, ISSN-2394 3386,
Consistency” Social Network Analysis and Mining, May Volume5, Issue1, January 2018
24,2018.
7. Peng Zhou, Wei Shi, Jun Tian, Zhenyu Qi, Bingchen Li,
Houng Wei, Hao,Bo Xu“Attention-based Bi-directional
Long Short Term Memory Network for Relation
Classification” proceedings of the 54th Annual Meeting of
the Association for Computational Linguistics,pages 207-
212,August 12,2016.
8. Alexis Conneau, Holger Schwenk, Yann Le cun “Very
Deep CNN for Text Classification”Association for
Computational Linguistics, Volume1, pages 1107-1116,7
April 2017.

607