See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/343079524

Detection of Offensive Language in Social Media Posts.

Thesis · May 2020

DOI: 10.13140/RG.2.2.23097.80485

CITATIONS READS

0 2,697

2 authors:

Sidharth Mehra Mohammed Hasanuzzaman

Cork Institute of Technology Munster Technological university
1 PUBLICATION   0 CITATIONS    44 PUBLICATIONS   113 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Multimodal Summarization View project

Temporal Orientation View project

All content following this page was uploaded by Sidharth Mehra on 20 July 2020.

The user has requested enhancement of the downloaded file.

Detection of Offensive Language in
Social Media Posts

by
Sidharth Mehra

This thesis has been submitted in partial fulfillment for the

degree of Master of Science in MSc in Artificial Intelligence

in the
Faculty of Engineering and Science
Department of Computer Science

May 2020
 Declaration of Authorship
I, Sidharth Mehra , declare that this thesis titled, ‘Detection of Offensive Language in
Social Media Posts’ and the work presented in it are my own. I confirm that:

 This work was done wholly or mainly while in candidature for an masters degree
at Cork Institute of Technology.

 Where any part of this thesis has previously been submitted for a degree or any
other qualification at Cork Institiute of Technology or any other institution, this
has been clearly stated.

 Where I have consulted the published work of others, this is always clearly at-
tributed.

 Where I have quoted from the work of others, the source is always given. With
the exception of such quotations, this project report is entirely my own work.

 I have acknowledged all main sources of help.

 Where the thesis is based on work done by myself jointly with others, I have made
clear exactly what was done by others and what I have contributed myself.

 I understand that my project documentation may be stored in the library at CIT,

and may be referenced by others in the future.

Signed:

Date:

i
 CORK INSTITUTE OF TECHNOLOGY

Abstract
Faculty of Engineering and Science
Department of Computer Science

Master of Science

by Sidharth Mehra

Posting offensive or abusive content on the social media have been a serious concern in
the recent years. This has created a lot of problems because of the huge popularity and
usage of the social media sites like Facebook and Twitter. The main motivation lies in
the fact that our model will automate and accelerate the detection of the posted offensive
content so as to facilitate the relevant actions and moderation of these offensive posts.
We would be using the publicly available benchmark dataset OLID 2019 (Offensive
Language Identification Dataset) [1] for this research project. The scope of our work
lies in predicting whether the tweet post is offensive or not. We contributed by making
the training dataset balanced using the Random Under-sampling technique. We also
performed the thorough comparative analysis of various Feature Extraction Mechanisms
and the Model Building Algorithms. The final comparative analysis concluded that, the
best model came out to be Bidirectional Encoder Representation from Transformer
(BERT). Our results outperforms the previous work achieving the Macro F1 score of
0.82 on this OLID dataset. Finally a real time system could be deployed on various
social media platforms to detect and analyze the offensive post content and taking the
appropriate action in order to normalize the behaviour on these sites and the society.
...
 Acknowledgements
I’m grateful to Dr. Mohammed Hasanuzzaman for guiding me through out this AI
research project. His valuable suggestions and encouragement helped me a lot in the
development of this project and the thesis. Moreover I felt very fortunate enough to
work under his expertise in the field of Natural Language Processing. As a project
Supervisor, he played a crucial role in the success of this research project.

I am immensely obliged to all the lecturers, specially the Head of Artificial Intelligence
Dr.Ted Scully from the Department of Computer Science at Cork Institute of Technology
for teaching us the latest modules in the field of AI along with the best practices in
programming.

I would like to pay my sincere regards and thanks to Mr. Nitin Seth- Vice Chairman
at G.D. FOODS MFG.(I) PVT. LTD. for financially supporting my Master’s here at
Ireland.

I would also like to give my greatest thanks to my seniors of Artificial Intelligence Anirud-
dha Kalburgi and Praveen Joshi for helping me in the best way possible throughout this
Master’s and this research project.

Lastly, I offer my regards to all of those who supported me in any respect during the
completion of my Master’s thesis.. . .

iii
Contents

Declaration of Authorship i

Abstract ii

Acknowledgements iii

List of Figures vi

List of Tables vii

Abbreviations viii

1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Executive Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Contribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Structure of This Document . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Background 5
2.1 Thematic Area within Computer Science . . . . . . . . . . . . . . . . . . . 5
2.2 Project Scope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.1 Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2.2 Machine Learning, Deep Learning and Natural Language Processing 6
2.2.3 Moving towards Offensive Language Detection . . . . . . . . . . . 11
2.3 A Review of the Thematic Area . . . . . . . . . . . . . . . . . . . . . . . . 12
2.3.1 Motivation and Challenges . . . . . . . . . . . . . . . . . . . . . . 13
2.3.2 Datasets Involved in Previous Work . . . . . . . . . . . . . . . . . 16
2.3.3 Existing Methodology . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.3.4 Results Analysis and Evaluation Metrics . . . . . . . . . . . . . . . 21
2.4 Current State of the Art . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3 Problem - Offensive Language Detection in Social Media 25

3.1 Problem Definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2 Dataset Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3.3 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27

iv
Contents v

4 Implementation Approach 29
4.1 Experiment Design and Methodology . . . . . . . . . . . . . . . . . . . . . 29
4.1.1 Data Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.1.2 Feature Extraction Techniques . . . . . . . . . . . . . . . . . . . . 31
4.1.2.1 Feature Extraction with Statistical Models . . . . . . . . 31
4.1.2.2 Feature Extraction with Word Embeddings . . . . . . . . 33
4.1.3 Handling Imbalance . . . . . . . . . . . . . . . . . . . . . . . . . . 34
4.1.3.1 Oversampling Techniques . . . . . . . . . . . . . . . . . . 34
4.1.3.2 Undersampling Techniques . . . . . . . . . . . . . . . . . 35
4.1.4 Algorithmic Modeling . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.1.4.1 Machine Learning Based Models . . . . . . . . . . . . . . 36
4.1.4.2 Deep Learning Based Models . . . . . . . . . . . . . . . . 37
4.1.5 Evaluation Metrics . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.2 Implementation Details and Results . . . . . . . . . . . . . . . . . . . . . 40
4.2.1 Process Flow and Implementation Details . . . . . . . . . . . . . . 40
4.2.2 Various Machine Learning Models with Different Feature Extrac-
tion Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.2.2.1 Feature Extraction with BOW . . . . . . . . . . . . . . . 41
4.2.2.2 Feature Extraction with Tf-idf . . . . . . . . . . . . . . . 42
4.2.2.3 Feature Extraction with Word2Vec . . . . . . . . . . . . . 43
4.2.2.4 Feature Extraction with GloVe . . . . . . . . . . . . . . . 43
4.2.3 Results of Various Deep Learning Models . . . . . . . . . . . . . . 44
4.3 Evaluation and Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.1 Comparative Analysis of Various Feature Extraction techniques
along with Machine Learning and Deep Learning Models . . . . . 44
4.3.2 Best Model for Offensive Language Detection . . . . . . . . . . . . 45

5 Conclusions and Future Work 46

5.1 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

Bibliography 49
List of Figures

2.1 Different Areas Under Artificial Intelligence . . . . . . . . . . . . . . . . . 7

2.2 Difference between Traditional Programming and Machine Learning . . . 8

2.3 Process Flow of Machine Learning . . . . . . . . . . . . . . . . . . . . . . 8

2.4 Categorization of Machine Learning . . . . . . . . . . . . . . . . . . . . . 9

2.5 Artificial Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

2.6 Categorization of Deep Neural Networks . . . . . . . . . . . . . . . . . . . 11

2.7 Composition of Natural Language Processing . . . . . . . . . . . . . . . . 11

2.8 Relationship between AI, ML, DL and NLP . . . . . . . . . . . . . . . . . 12

3.1 Distribution of 3 levels of annotation in OLID . . . . . . . . . . . . . . . . 27

3.2 WordCloud for Non-Offensive and Offensive Tweets . . . . . . . . . . . . 28

4.1 Feature Extraction through BOW . . . . . . . . . . . . . . . . . . . . . . 32

4.2 Training set distribution of class labels . . . . . . . . . . . . . . . . . . . . 34

4.3 Test set distribution of class labels . . . . . . . . . . . . . . . . . . . . . . 38

4.4 Representation of Confusion Matrix . . . . . . . . . . . . . . . . . . . . . 39

vi
List of Tables

4.1 Evaluation Metric for Feature Extraction with BOW (Uni-gram) . . . . . 41

4.2 Evaluation Metric for Feature Extraction with BOW (Uni-gram + Bi-gram) 41

4.3 Evaluation Metric for Feature Extraction with BOW (Uni-gram + Bi-
gram + Tri-gram) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

4.4 Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram) . . . . . 42

4.5 Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram + Bi-gram) 42

4.6 Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram + Bi-
gram + Tri-gram) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.7 Evaluation Metric for Feature Extraction with Word2Vec . . . . . . . . . 43

4.8 Evaluation Metric for Feature Extraction with GloVe . . . . . . . . . . . . 43

4.9 Evaluation Metric for Deep Learning Models . . . . . . . . . . . . . . . . 44

4.10 Comparative Analysis of Various Feature Extraction, Machine Learning

and Deep Learning Techniques . . . . . . . . . . . . . . . . . . . . . . . . 44

vii
Abbreviations

AI Artificial Intelligence
ML Machine Learning
DL Deep Learning
NLP Natural Language Processing
SVM Support Vector Machine
CNN Convolutional Neural Network
BiLSTM Bidirectional Long Short Term Memory networks
GRU Gated Recurrent Unit
BERT Bidirectional Encoder Representations from Transformers
OLID Offensive Language Identification Dataset
BOW Bag Of Words
SMOTE Synthetic Minority Oversampling Technique
TF-IDF Term Frequency Inverse Document Frequency

viii
Dedicated to my Mother Mrs.Anita Mehra and Father Mr.Rakesh
Kumar Mehra for providing the constant support and
encouragement throughout this Master’s degree.. . .

ix
Chapter 1

Introduction

This introduction chapter mainly focuses on four sections. Firstly we will discuss about
the motivation behind choosing this particular topic for our AI research project. Here
we briefly explains the importance and the major societal impact of choosing this topic
and making the contribution in the field. Secondly we will give a brief executive sum-
mary about our project describing about the main features along with the functional
requirements. Next we will highlight our main contributions achieved in this particu-
lar problem area by explaining about the research objectives. Lastly, we present the
structure of this thesis report representing what content is there in which chapter and
corresponding sections.

1.1 Motivation

Increase in the usage of social media sites likes Facebook and Twitter have given the
crowd a great platform to express their opinions/feelings for the individual, groups or
events happening around them or in society. This digital media has become a great
resource to share the information and also gives the full freedom of speech to everyone
on the platform.

With the gaining popularity of these platforms; there also comes the negative part
along with its benefits. This feature of the social media to express something openly to
the world have created the major problems for these online businesses and negatively
impacted the well being of the societal decorum. There are increasing cases of the abuse
or offense on the social media like Hate speech, Cyber-bullying, Aggression or general
Profanity. It is very much important to understand that this behaviour can not only
immensely affect the life of an individual or a group but could be suicidal in some cases;
adversely hampering the mental health of the victim/s.
1
Introduction 2

This increasing negative situation on the internet has created a huge demand for these
social media platforms to undertake the task of detecting the objectionable content and
taking the appropriate action which can prevent the situation becoming more worse.
This task of detecting the offensive content can be performed my human moderators
manually, but it is both practically infeasible as well as time consuming because of the
amount of the data generated on these social media platforms, therefore there is a need
to fill this gap. Numerous studies in the past as evident by the following chapter 2 tries
to tackle this problem and gap by leveraging the technologies like Artificial Intelligence,
Machine Learning and Natural Language Processing. They are able to build the systems
that can efficiently detect these types of offensive content so that the appropriate action
could be taken as quickly as possible.

So in the last, considering the importance and the sensitivity of this particular topic in
the today’s digital world there is, still a lot of emerging further scope in tackling and
improvise on the previous work done in the field with the help of these new age AI and
NLP techniques. Therefore we aim to contribute in this particular direction.

1.2 Executive Summary

In this section of executive summary, we will try to summarize the work carried out in
all the chapters of our project in some details. We will start off with some Introduction
about the project area that will describe the motivation to take this particular topic.
Then we will briefly describe what we have done in the Background chapter to discuss
about the extensive review and analysis of the previous work related to our topic. Then
we focus on the architecture and implementation approach that we proposed in this
thesis. Finally we conclude this summary by the contributions achieved in this problem
area of Offensive Language Detection.

The task that we have taken for this research project is to detect whether a given tweet
is offensive or not with the help of NLP techniques along with ML and DL. The main
motivation for choosing this topic is to develop an efficient system to detect this offensive
content on the social media platform which can be removed afterwards. This will prevent
the unease and negativity that may spread in society, people or any community group’s
mind.

The background chapter contains all the technical concepts related to AI, ML, DL and
NLP which we have used for this project. We have included these details so that it
becomes relatively easy for a person who has limited knowledge about the subject to
progress towards the later chapters and sections of this document. In this chapter we
Introduction 3

critically analysed about 15 academic research papers that revolve around our project
topic. This study helped us to form an appropriate research direction and identify the
important challenges in the field. This chapter also helped us to come up with the
relevant dataset for the problem that is OLID 2019 dataset.

Further, we focussed more on the problem description and also mentioned the important
features and the distribution of our OLID dataset. As this dataset is the recent one and
not much work has been done on it, we laid down some objectives that we can achieve
in order to make the significant contribution after the baseline results on this dataset.
We divided our proposed solution methodology into mainly five steps. These were Text
Pre-processing, Feature Extraction, Handling Imbalance, Model Building and Model
Evaluation. In this chapter, we presented our results for the top performing techniques
in each of the steps of this NLP pipeline.

As our main contribution to this field, we contributed by making the training dataset
balanced using the Random Under-sampling technique that worked best for our problem.
Moreover from our comparative analysis of various Machine learning and Deep learning
algorithms, BERT performed the best and achieved the Macro F1 score of 0.82 on this
OLID dataset. Therefore we made the small improvement in the field when compared
with the previous work on this dataset and opened up a new pace to work more on this
novel dataset.

1.3 Contribution

1. For this research project we researched on the recent OLID 2019 dataset which
considered the problem as the whole by taking all the offense types into the account.
As the dataset is new and not much work has been done on this, therefore to
carry out our research to its full extent and explore new insights; we performed
a thorough comparative analysis of various Feature Extraction Mechanisms and
Model Building Algorithms.

2. We also contributed towards mitigating the imbalance in the training twitter data
with the help of Random Under-sampling Technique.

3. Finally we developed an efficient system based on state of the art deep learning
algorithm BERT to detect whether a given tweet is offensive or not. We achieved
the Macro F1 score of 0.82 which clearly outperformed the performance of previous
work [1] on the dataset.
Introduction 4

1.4 Structure of This Document

Whole of this thesis document is mainly divided into 5 chapters.

Chapter 1- Introduction

This chapter comprises of four sections. First section talks about the motivation for
this particular topic. Second section gives a brief executive summary about the project
chapters. Third section focus on the main contributions achieved in this problem area.
Finally we close our chapter with the last section which present the structure of this
thesis document.

Chapter 2- Background

This chapter comprises of four sections. We start of with the first section of the overview
of the Thematic area within Computer Science. Second section lays focus on the Project
Scope. Third section is all about the Review of the Thematic Area. Lastly, we close this
chapter by the Current State of the Art section.

Chapter 3- Problem (Offensive Language Detection in Social Media)

There are a total of three sections in this chapter. First section provides us the descrip-
tion of the problem topic. Second section gives the insights about the dataset we are
going to use for this thesis work.Third and the last section focus on enumerating all the
Research Objectives that we need to achieve for the meaningful contribution.

Chapter 4- Implementation Approach

This chapter is broadly divided into three sections. First section is the experiment
design and methodology further dividing into the detailed subsections of the phases of
our proposed approach. The second section lays focus on the implementation details
and the results achieved. Finally we close this chapter with the evaluation and analysis
of the results achieved in the previous section.

Chapter 5- Conclusions and Future Work

This chapter has only two sections. First section discusses about the Conclusion reached
for this research project. Second section is about the possible Future Work in the
problem domain.
Chapter 2

Background

This chapter starts with the thematic area within Computer Science which explains
the concrete topic of the project along with the concrete areas and the main area of
Computer Science under which our project falls. Next section deals with the project
scope which explains all the basics of the topics and the areas discussed in the previous
section. This is done to give a short technical overview of the domain knowledge to a
person who knows nothing or little about the field. Further, we move on to the review
of the thematic area which explains the previous work done related to our project topic
in terms of motivation and challenges, datasets involved in the previous work, existing
methodology, results analysis and evaluation metrics. Next section of current state of
the art focus on critically analyzing the previous work to find the major similarities and
differences between our project and each of these works. This will form a strong basis
to contribute to this particular problem area.

2.1 Thematic Area within Computer Science

1. This research project is about the detection of offensive language on the social me-
dia platforms like Twitter or Facebook. This problem is very serious and increasing
largely on these sites, therefore we aim to provide a solution for normalising these
derogatory/abusive posts using algorithmic techniques in Machine Learning. This
can effectively minimize the negative impact of these posts on the individual or a
group.

2. The area under which our project falls is Natural Language Processing which
is the art of understanding (Natural Language Understanding) and generating
(Natural Language Generation) the human language such as text or speech by the
machines in a similar manner like human beings. Here particularly we focus on

5
Background 6

Natural language understanding which gives machines, the ability to understand

the human text (which is in our case are offensive posts) and automate the process
of removing its presence on the social media without the human intervention.

3. The main area of Computer Science under which our project falls is Artificial
Intelligence which aims to inculcate the human-like behaviour into the machines
and make them capable to think, act and interact as similar to human beings.
Our project combines the techniques from Machine Learning, Deep Learning and
Natural Language Processing (all falls under the umbrella of Artificial Intelligence)
to accomplish the main objective.

2.2 Project Scope

2.2.1 Artificial Intelligence

“Artificial intelligence (AI) is an area of computer science that emphasizes the creation
of intelligent machines that work and reacts like humans.”(Andrew Ng, 2015)

Artificial Intelligence is the branch under Computer Science that deals with the develop-
ment of the machines that mimic the human behavior and intelligence. The main goal
of an AI is to work and perform similar to the cognitive functions of a human brain.
These functions are mainly decision making, problem solving and learning through the
environment. There is an enormous progression in the generation of the data in the
today’s digital world. This is all facilitated through the web-based life, sensors, gadgets,
online businesses and many more factors giving rise to the huge amount of structured,
un-structured and semi-structured data. This explosion of data is giving rise to the
new economy called data economy. New oil is the data which is precious but useful
only when cleaned and processed. This data is the only driving source for the machines
with which they learn and then perform similar to humans and derive the meaningful
insights. There are many sub-fields under the hood of Artificial Intelligence which are
depicted in the below figure 2.1.

2.2.2 Machine Learning, Deep Learning and Natural Language Pro-

cessing

Our project combines the three areas of AI which are Machine Learning, Deep Learning
and Natural Language Processing. Here we will look at the basics of each of the field to
get the better sense of the project topic.
Background 7

Figure 2.1: Different Areas Under Artificial Intelligence

[2]

Overview of Machine Learning

“Machine learning is the science of getting computers to learn and act like human do,
and improve their learning over time in autonomous fashion, by feeding them data and
information in the form of observations and real-word interactions”(Arthur Samuel,
2016).

Machine Learning is one of the most important application areas of Artificial Intelligence.
Machines are developed in such a way that they learn from the enormous amount of
data and get trained. Once our ML model is trained on this training data, it can make
the predictions on the data examples which it has never seen. According to Stanford
university definition which is “Machine learning is the science of getting computers to act
without getting explicitly programmed”.The widely known formal, social and scientific
definition of machine learning was given by Mitchell (1997) is ”A computer program is
said to gain for a fact E regarding some class of errands T and execution measure P, if
its execution at assignments in T, as estimated by P, improves with experience E”. The
task defines the objective of the problem. For example classification, regression, anomaly
detection, etc. The sole base for Machine learning is the availability of the data and
does not rely on any rule based programming. These algorithms by generalizing from
examples can perform important tasks.There is the difference between Machine learning
and Traditional programming. In traditional programming we give data and program as
inputs to the machine and it will provide the output whereas on the other side if we talk
about machine learning, the input to the system is data and output and the output of the
machine is the program that have been learnt to make predictions on unknown examples
points. Figure 2.2 illustrates the major difference between traditional approach and the
approach followed by Machine Learning.
Background 8

Figure 2.2: Difference between Traditional Programming and Machine Learning

[3]

Basic Process Flow of Machine Learning

Below Figure 2.3 is the basic process that is carried out in most of the Machine Learning
problem statements. We train our Machine learning model on the training data using the
various algorithms available. The most common types of tasks that are associated with
machine learning are Regression and Classification. Then our trained model makes the
prediction on the unseen test examples and finally we evaluate the model performance.

Figure 2.3: Process Flow of Machine Learning

[4]

Types of Machine Learning Techniques

• Supervised Learning: These types of learning algorithms are most common

in Machine learning space. In these type of techniques, we are given with the
labelled outputs of each of the input training examples so that the model can
learn the patterns and relationship between the inputs and those outputs. Some of
Background 9

the examples of supervised learning algorithms are K-Nearest Neighbours, Linear

Regression, Support Vector Machines, Logistic Regression and many more.

• Unsupervised Learning: In this type of learning technique, we are not given

any input to output mappings in our dataset. The model will analyse the hidden
patterns and groupings in the data to find the similarities and differences between
each of the data points.

• Reinforcement Learning: This type of learning mechanism is based on the

agent and environment. The program learns to make a sequence of decisions
based on the reward and penalty model and try to maximize the reward.

Below figure 2.4 shows the three different types of Machine learning techniques that are
widely used.

Figure 2.4: Categorization of Machine Learning

[5]

Overview of Deep Learning

Deep Learning is the field of Artificial Intelligence that comes under the domain of
Machine Learning which we have seen above in good details. Deep Learning can be
defined as the sub-set of Machine learning that is inspired by the biological structure
and function of human brain containing over billions of neurons. These are also known
as Deep Neural Networks as they contain many layers of neurons which are capable to
recognize patterns from the raw input data and perform the decision making in the same
way as the human mind performs. Every Deep Neural Network learns the patterns from
the data through the technique called Back-propagation. We have mainly three types of
layers in an Artificial Neural Network which are Input Layer, Hidden Layers and Output
Layer, each of them are briefly introduced below.
Background 10

• Input Layer: All the neurons in the input layer correspond to each of the feature
in the dataset. The values of each instance are the inputs to this layer.

• Hidden Layers: The values from the input layer multiplied by the weights are
the input to the hidden layer neurons to get the pre-activation outputs. Then we
pass these pre-activation outputs to an activation function to get the final outputs
from the hidden layer.

• Output Layer: The activated outputs from the hidden layers acts as the input
to the neurons in the output layer which are multiplied by the weights to the
get the pre-activation outputs. Then we pass these pre-activation outputs to the
activation function to get the final outputs from the output layer.

Same as Machine learning, Deep learning can be also classified into three categories
Supervised learning, Unsupervised learning and Reinforcement learning. Below figure
2.5 shows the structure of an Artificial Neural Network with one input layer, two hidden
layers and one output layer.

Figure 2.5: Artificial Neural Network

Different types of Deep Neural Networks

Depending on the use-case of the problem that we intend to solve with Deep Learning
and the type of available data, there are various kinds of Deep Neural Networks that
are meant to tackle that particular use-case. For example an Artificial Neural Network
may work well on the numerical kind of data but on the other hand a Convolutional
Neural Network may perform better on the images. Below figure 2.6 shows a whole host
of Neural Networks that exists for solving a particular type of problem.
Background 11

Figure 2.6: Categorization of Deep Neural Networks

Overview of Natural Language Processing

Natural Language Processing is the sub-field that takes the inspiration from the areas
of Artificial Intelligence and Linguistics. It enables the computers/machines to process
and analyze the large amount of human language data such as speech or text. We can
think NLP as the computer programs that can understand and generate the human
understandable language to derive meaningful insights. NLP can be further classified
into Natural Language Understanding and Natural Language Generation which can be
shown by the below figure 2.7

Figure 2.7: Composition of Natural Language Processing

[6]

Below figure 2.8 is the relationship between AI, ML, DL and NLP. It is quite evident
from the Venn Diagram that all ML, DL and NLP comes under the hood of AI and also
DL is the sub-set of ML.

2.2.3 Moving towards Offensive Language Detection

Now we have seen the basic concepts and ideas revolving Artificial Intelligence, Machine
Learning, Deep Learning and Natural Language Processing; we will now focus on un-
derstanding the specifics of our project. The topic of our project is the Detection of
Offensive Language in Social Media which encompass these fields AI, ML, DL and NLP.
Background 12

Figure 2.8: Relationship between AI, ML, DL and NLP

[7]

The problem can be simplified as to detect whether the posts on the twitter are offensive
or not and if they are then they can be removed from the platform to minimize the social
instability. We as humans are able to detect whether the post is offensive or not and
here we are fusing this capability into machines so it relates well with AI. As it deals
with understanding of the human written text by machines then this makes it a NLP
problem too. Lastly we need to understand how the ML and DL applies to our problem
statement. A system or a computer program will be trained on the various offensive
and non-offensive tweets so that it recognizes the patterns from them to distinguish
between the two. Now in future our Machine Learning or Deep Learning model can
accurately classify the unseen tweet as offensive/non-offensive and further action could
be automated.

2.3 A Review of the Thematic Area

In this section we would be critically analysing the few academic papers that revolve
around the detection of offensive language in social media using machine learning. We
would be extracting the summaries of these papers along with their main characteristics
which are defined with the proper structure described in the later part of this chapter.
We will also determine the common and different points between these papers which
will help us to extract the structure and the context of the problem solution. After
analysing these papers that are formulated according to the below structure, we will
develop a similar structure to illustrate the context of our problem solution. We will
also compare the existing methods and try to identify that is there any clear best solution
or are there any trade-offs. The main characteristics of these papers are structured as
the individual sub-sections of this section following a chronological order.
Background 13

2.3.1 Motivation and Challenges

The motivation, challenges and the direction of mitigation are mentioned below for each
of the related works

Paper [8] was one of the first significant works that was related with detecting cyber-
bullying on the social media with the help of NLP. Earlier this problem was studied
mostly by the psychiatrists and social scientists. Authors believed that anonymity and
lack of meaningful supervision has put out this topic into more attention to computa-
tional linguists. They moreover believed that the appropriate action on this content
through NLP could prevent the tragic outcomes of bullying on social media platforms.
Authors modeled this problem into 2 parts. First one is to determine whether the given
comment is sensitive or not making it a binary classification task. And if the comment is
sensitive then classify it as the comment revolving around one of the areas in sexuality,
race culture or intelligence which made this as multi-class classification subtask.

The very strong baseline paper [9] aimed to tackle the serious issue of cyber-bullying
on social media platforms using NLP techniques so that the further investigation on
the same domain could be explored in future by the NLP community with the help of
their baseline results. They moreover formulated the bullying on the social platforms
into 4 major NLP tasks namely text classification, role labeling, sentiment analysis and
topic modeling. They defined the bullying traces as the posts by the individuals who
have been a victim of bullying. Objective of subtask A was to distinguish the bullying
episodes from the non-bullying traces in the dataset examples. The subtask B dealt
with the role labelling of the tweets into Author’s role and Person mention role which
can be thought of a binary text classification task. The other thing in this subtask is
to further sequentially tag these roles into one of the 5 categories which are Accuser
(A), bully (B), reporter (R), victim (V) and other (0). Subtask C was related with
classifying the sentiment of the bullying episode to understand their motivation. This
task was a binary task of classifying the tweet as teasing or not. Subtask D was related
with latent content modelling which extracts the main topics to better understand the
bullying traces.

The authors in [10] realized the need of automating and improving the task of detecting
the bullying content on the social media. Authors clearly outlined that the previous
work has only investigated the bullying detection only from individual comments. They
strongly felt that taking user characteristics and profile information could definitely yield
in better model for characterizing the sensitive content.

Authors in [11] believed that they were among the first one to investigate the topic of
hate speech detection on social media particularly when it comes to racist comments
Background 14

which were proliferating hugely on these online platforms because of freedom of speech
and the features of re-tweeting. They analysed the racism against the black as they
constituted 25% of the whole twitter population at that time and they were the highest
one to face the racism on the platform. They formulated the problem as the binary
classification task to determine whether the tweet is racist or non-racist.

The motivation of authors in the work [12] was triggered by the murder of Drummer
Lee Rigby in Woolwich, London, UK. They observed and believed that immediately
after such events there is a potential opportunity of spreading the hate speech on online
platforms. They formulated this problem as the binary classification task of predicting
whether the tweet is hateful/ antagonistic or not with a focus on race, ethnicity or
religion. They derived their features from the context of each tweet and experimented
with various probabilistic, rule-based, and spatial-based classifiers with a voted ensemble
meta-classifier. They believe that their contribution is closely related to the effective
decision and policy making.

In [13], the authors took the problem of detecting hate speech on social media using
user comments. The motivation of their work was developed because of increasing cases
of hate on the online platforms causing the decline in the online business and the user
experience. They aim to develop a low dimensional representation of comments using
neural language models which they would be sending to classification algorithms. They
addressed the problem of high dimensionality and sparsity due to previous work of BOW
model and aimed to achieve accurate predictive models for the domain.

Authors in [14] were the first one up till 2016 to focus on the problem of detecting the
offensive language on online platforms rather than just focussing on the specific type of
abuse such as cyber-bullying or hate-speech as evident in the past years. They brought
out the attention to the drawbacks of regular expression and black-lists in the extreme
cases of hate speech which are more subtle and less obvious than regular ones. They
tend to outperform the state of the art deep learning approach to build a predictive
model and also created a new corpus of user comments as a unique one of its kind.

Authors in [15] raised the challenge of separating the hate speech with other types of
offensive content as the previously used lexical methods failed to distinguish between the
different offense types. It is important to accurately identify the hate speech particularly
from the other offensive content as both tend to have the different implications on the
society and the individuals. They worked with twitter data that was labelled into 3
categories namely hate speech, offensive language and neither of these two for which
they trained multi-class classifier.
Background 15

In [16] the authors extended the main challenge from [15] to distinguish the hate speech
from general profanity or offense to come up with a lexical baseline for the same.

The work presented in [17] aimed to identify and empirically analyse the relationship;
particularly similarities and differences between the different subtasks that come under
the offensive language detection. They also proposed a topology that could help the
researchers for data annotation and feature construction.

Authors in [18] believed that the offensive content is increasing at an unprecedented rate
on the online platforms which have potential to hamper the mental lives of individual
and groups leading to a negative effective in the society. They believed that the manual
moderation is practically impossible with this rapid generation of the data. Therefore
they introduced this problem at TRAC workshop in 2018 and modelled the problem
as the multi-class classification task of differentiating the posts and the comments into
Overtly Aggressive, Covertly Aggressive, and Non-aggressive texts. They also identified
the important challenges in the shared task mainly revolving around the annotation and
the language issues.

The related work presented in [19] also aims to distinguish hate speech from the regular
type of offensive content / profanity which has not been studied much in the previous
work. They used the n-grams, skip-grams and clustering based techniques for the feature
representations. They analysed that the previous work formulated the task of hate
speech detection as the binary classification problem in which the systems are likely to
misclassify the instances as non-offensive that did not contain any objectionable words
but are actually offensive in a semantic and deeper sense. They also focussed on ensemble
classifiers rather than the single classifiers which were mainly used in the past studies.
The work considered that the opinion variation of the annotators is also an issue for this
problem which needs to be resolved.

The authors in [20] dealt with the topic of offensive language on the German tweets
which was announced as the shared task in GermEval 2018 competition. The problem
was formulated into 2 subtasks. First one is the binary classification task to differentiate
between the offensive and non-offensive German tweets. The second one is the further
classification of offensive tweets into profanity, insult, abuse and other.

Authors in [21] believed that the text with high toxicity on the internet can cause
personal attacks, bullying behaviours, threatening and harassment. They utilized the
Wikipedia dataset released under Kaggle competition of toxic comment classification
which was modelled as the 6-class classification problem. They employed CNN for the
learning the structure of words in the document due to the advances in hardware and
cloud computing.
Background 16

The work in [1] considered the problem of the offensive language detection on the social
media as the whole rather than focussing on the specific type of abusive content on the
web as evident by the previous studies. They came up with the new dataset with three
level of annotation scheme providing an opportunity for the researchers to delve deeper
into the topic. Subtask A is to determine whether the post is offensive or not. Subtask
B further identifies the type of the offense content whether it is targeted or untargeted.
Subtask C then categorizes the target of the offensive post into either one of 3 classes
namely individual, group and other.

2.3.2 Datasets Involved in Previous Work

Related work [8] used the labeled corpus of 4500 YouTube video comments that were
scrapped from web. The sensitive examples from the dataset were labeled into one of
the 3 categories namely sexuality, race culture and intelligence.

In paper [9], the authors used the sampled version containing labeled 1762 tweets from
the TREC corpus developed in 2011.

At that time as no dataset for the bullying detection was publically available so authors
in [10] scrapped the comments from the top 3 videos of different categories found in
YouTube movies. The final dataset contained 4626 user comments consisting of 3856
distinct users and were manually labeled as bullying and non-bullying. Comment history
of each user was also recorded and on an average the dataset contained 54 comments
per user profile.

Authors in [11] prepared a balanced dataset from the twitter accounts that contained
24582 carefully labeled tweets by annotators as racist or non-racist.

The authors in [12] collected the data from twitter API within the 2 week window after
the murder of the drummer as they believed that majority of the hate comments bound
to happen in the 2 weeks’ time and gradually decline. Therefore they wanted to capture
this immediate reaction. A total of 450,000 tweets were captured during the study
window and labeled as ‘Yes’ or ‘No’ meaning that the comment is hateful or not.

In [13], the authors prepared the largest dataset available till 2015 for hate speech
detection for over 6 months and named as WWW-2015 dataset. They collected the
comments from the Yahoo Finance website containing a total of 951,736 user comments
out of which only 5% were hate speech comments and rest 95% of the comments were
clean without the use of any offensive language. This was therefore modelled as the
binary classification task as well.
Background 17

The dataset in [14] was the extended version of the one used in [6]. It contained the
user comments from the Yahoo finance as well as the Yahoo news adding the diversity
to it because of presence of all types of abusive language like hate speech, profanity
or derogatory language. They marked each user comment as the “abusive” or “clean”
making it a binary classification problem. In the news dataset only 16% of the data was
abusive and in the finance dataset only 7% of the comments were abusive.

Authors in [15] prepared the dataset from twitter API that contained 24,802 tweets each
of them labeled as hate speech (5%), offensive (76%) or neither of these two (19%).

The authors in [16] used the open source dataset containing 14,509 English tweets which
are labeled as HATE (16%), OFFENSIVE (33%), or OK (51%) reflecting the absence
of any offensive content.

The dataset in [18] contained roughly 12,000 training comments from Facebook. For the
development they provided the participants with 3000 comments in English and Hindi
language respectively. Each comment was annotated with a label indicating the levels
of aggression namely Overtly Aggressive (OAG), Covertly Aggressive (CAG) and Non-
Aggressive (NAG) making it a multi-class classification task. The test set contained 916
and 970 comments in English and Hindi respectively.

The authors in [19] used the same dataset as in [16] making it a 3-class classification
problem of distinguishing between hate speech and general profanity.

The dataset in [20] contained over 8500 annotated tweets which were labeled as OF-
FENSE (33.7%) and OTHER (66.3%). Each offensive tweet was further annotated as
ABUSE (20.4%), INSULT (11.9%), PROFANITY (1.4%) and OTHER (66.3%).

Authors in [21] used a dataset having Wikipedia comments such that each comment is
annotated into either of 6 classes namely toxic, severe toxic, obscene, threat, insult and
identity hate.

In the work [1], the authors came up with a new dataset OLID for the offensive language
detection in social media. At level A, each tweet was assigned a label as Not Offensive
(NOT) or Offensive (OFF). At level B, the offensive tweets were annotated by either of
2 labels namely Targeted Insult (TIN) and Untargeted (UNT). At level C, the targeted
offensive tweets were given an either of 3 labels namely Individual (IND), Group (GRP),
Other (OTH).
Background 18

2.3.3 Existing Methodology

Authors in [8] applied the standard pre-processing techniques like stemming, and removal
of stop-words and unimportant characters from the textual data. They experimented
with three classifiers Naı̈ve Bayes, SVM and Decision Tree with 10-fold cross validation.
They divided their feature space into general features and label specific features. For
the general features they used TF-IDF and for the specific features they used unigrams
and bigrams.

In [9] the authors trained the model to detect and analyze the cyber-bullying on the
social media platforms. They divided their work into 4 subtasks. For the subtask
A, after the regular pre-processing techniques they applied three techniques for the
featurization namely unigram, unigram+bigram and POS colored unigram+bigram. For
the classification task they experimented with Naive Bayes, SVM with linear kernel,
SVM with RBF kernel and Logistic Regression with 5 fold cross validation using the
WEKA implementation. In subtask B, for the author’s role they used the same classifiers
as the subtask A with 10-fold cross validation and also tuned the best model jointly
with 5-fold cross validation with grid search CV. For categorizing the person mention
role into respective categories they used the named entity recognition and trained the
linear CRF and SVM respectively with 10-fold cross validation. In subtask C, they used
the same feature representation, classifiers and parameter tuning as for the previous 2
subtasks with 10-fold cross validation. They used LDA as well as its variational inference
implementation as their exploratory tool to discover the relevant topics from the bullying
trace in the subtask D.

Authors in [10] used three feature sets to train the cyber-bullying classifier which were
content-based, cyber-bullying based and user-based features. For the pre-processing
they removed all the stop words and applied stemming to their dataset. They trained a
Support Vector Machine to classify the bullying comments and non-bullying comments
with 10-fold cross validation.

To deal with the problem of hate against black community on twitter the authors in [11]
trained a Naı̈ve Bayes classifier to able to classify the new tweet as racist or nonracist.
They pre-processed the dataset by eliminating the URL’s, mentions, stop words and
punctuation along with lowercasing and replacing the wrong spellings with the correct
ones. Authors found that 86% of the tweets that were racist only because they contained
the offensive words so they preferred unigram model to featurize the training data.

Work in [12] was a binary classification task of predicting whether the comment is hateful
or not. They followed a pipeline starting from Data collection and annotation, Feature
selection, Data pre-processing, Feature preparation and finally Model selection. They
Background 19

realized that the offensive words from the tweet could be the important features so
they utilized the frequency of occurring of unigram and bigram. As the offensive tweet
contain the certain instances following a particular pattern therefore for the extraction
of typed dependencies within the tweet text they employed a Stanford lexical parser
along with a context free lexical parsing model which represented the syntactic gram-
matical relationship in a sentence that are used a important features for the classifier.
They came out with more common sense type of reasoning approach for this feature
extraction phase. For the pre-processing phase they followed a generalized pipeline of
tokenization, lowercase conversion, removal of stop words and alphanumeric characters,
stemming. To preserve the context of words and the surrounding they employed uni-
grams to trigrams. They experimented with the 2 approaches of n-grams and collection
of derogatory or hateful terms to check the contribution of other terms in determining
the strong predictors. They ran a Bayesian Logistic Regression using all the typed de-
pendencies features and came up with the vector representation of the tweet containing
list of ngrams that included words, typed dependencies or combination of both. They
used the three classifiers Bayesian Logistic Regression, Random Forest Decision Tree
and Support Vector Machine for this binary classification task. They also employed the
meta voting ensemble classifier made from these classifiers.

After applying the standard pre-processing techniques, they [13] divided their work
into two parts for the detection of hate speech from the user comments. First they
employed paragraph2vec to learn the distributed representation of comments and words
using the neural language model of the continuous BOW (CBOW). This produced a
low dimensional embeddings where the semantically similar comments resided in the
same part of the space. Secondly a logistic regression classifier was trained on these
embeddings to classify the type of user comment as hateful or clean.

Authors in [14] used the Vowpal Wabbit’s regression model to measure the different
aspect of the user comments using NLP features. They divided their features into 4
categories which were N-grams, Linguistics, Syntactic and Distributional Semantics.
Due to noise found in the data they performed some mild-preprocessing for the first
three features but did not performed any normalization for the fourth feature.

In [15], the pre-processing part was undertaken as to convert the tweet into lowercase and
performed stemming through the porter stemmer. After that they featurized the tweets
as weighted TF-IDF unigrams, bigrams and trigrams followed by the construction of the
POS tagging using NLTK. They used Flesch-Kincaid Grade Level and Flesch Reading
Ease scores to capture the quality of each tweet and also assigned the sentiment scores
to each of the tweet. For the hashtags, mentions, retweets and URL’s, they included
binary and count indicators and for the number of characters, words and syllables, they
Background 20

included features. They tried various models in Scikit-learn like Logistic regression,
Naı̈ve Bayes, Random Forest, Decision Tree and Linear SVM to train the model using
5-fold cross validation along with L1 regularization to reduce the dimensionality of the
text data. They also performed the grid search parameter tuning to find the optimal
parameters.

Authors in [16] used a LIBLINEAR SVM implementation for this multi-class classifi-
cation task which has proven to be very effective on Native language identification and
temporal text identification. For the features they used character n-grams, word n-grams
and word skip-grams.

In [17], the authors defined their topology based on the prior work in the field of detection
of different types of abusive language. They considered a 2 fold approach where the first
aspect is to analyse the target of the abuse and another aspect is to analyse the degree
to which the abuse is explicit. They also laid the implications of this topology on the
annotation and the modeling of this problem. They suggested that the data annotation
strategies should be dependent on the type of the abuse that is intended to be identified.
On the other hand to select the most relevant features for the modeling, it is important
to identify whether the abuse is directed, generalized, explicit or implicit.

In TRAC workshop proceedings [18], there were a total of 30 teams who submitted their
systems for English and Hindi Language. Participants applied various techniques like
LSTM, CNN, SVM, BiLSTM, Logistic Regression, Random Forest and many more to
classify the English and Hindi Facebook comments.

The participants of the shared task of GermEval [20] used tokenization, POS-tagging,
lemmatization and stemming and parsing as the methods for tweets pre-processing. They
used SVM, Logistic Regression, Naı̈ve Bayes, CNN, LSTM, GRU and the combination
for the classification of the tweets.

The authors in [21] compared word embeddings and CNN against the BOW approach
with the classifiers such as SVM, Naı̈ve Bayes, k-NN and LDA that were applied on the
Document Term Matrix.

Related work as evident in [1] applied various machine learning and deep learning tech-
niques for each of the subtask within the problem domain. They first applied linear SVM
trained on word unigrams followed by BiLSTM with the softmax activation function in
the final layer with FastText embeddings. Finally they also experimented with CNN on
this dataset.
Background 21

2.3.4 Results Analysis and Evaluation Metrics

In [8] the authors achieved the best accuracy of 80.2% with the help of rule based JRIP
whereas kappa measure was highest for SVM with the average value of 0.75.

In paper [9], for the subtask A the best model came out to be Linear SVM with a
combination of unigrams and bigrams achieving an F-measure of 0.77. For the subtask
B they achieved the cross validation accuracy of 61% with SVM linear along with the
combination of unigram and bigram for the author’s role. For the person mention role
linear CRF outperforms the SVM achieving the accuracy of 87% and F-measure of 0.47.
For the subtask C the best validation accuracy of 89% is achieved by linear SVM. For
the subtask D, LDA discovered 5 topics from the bullying post namely feeling, suicide,
family, school, verbal bullying and physical bullying.

The results from the work in [10]clearly indicated that the detection accuracy was
boosted upon adding the bullying specific features and the user context. The model
achieved the F-measure of 0.64 when they used all the three feature space to train the
model.

Authors in [11] achieved a 10-fold cross validation accuracy of 76% using the Naı̈ve Bayes
classifier and an error rate of 24%.

Results as evident from [12] indicated that the most efficient features proved to be a
combination of ngram typed dependencies and hated terms. All the classifiers BLR,
SVM, RFDT and Ensemble performed equally well on the test set with an F-measure
of 0.77

In [13] the authors achieved the Area Under Curve (AUC) value of 0.80 with the para-
graph2vec representation trained on a logistic regression classifier.

In [14] the authors achieved the best F-measure of 0.79 on the finance data and 0.81
on the news data when trained considering all of the features rather than the selective
features.

Results from the work in [15] indicated that Logistic regression with L2 regularization
performed significantly better than the other models with the F1 measure of 0.90.

In [16] , the authors achieved a good accuracy of 78% with the character 4-gram model
with Linear SVM evaluated using the stratified 10-fold cross validation.

In the shared task of TRAC [18], the best ranked team used LSTM with the data aug-
mentation strategy. They used a combination of CNN and RNN on the surprize twitter
Background 22

dataset for the feature representation. The team performed the spelling correction, emo-
jis conversion and sentiment score computation as the part of text pre-processing and
achieved F-score of 0.64 for both the Hindi and English on the Facebook test set. They
also managed to come up as the team that significantly performed better on the twitter
dataset despite being trained on the Facebook dataset achieving the F-measure of 0.60
and 0.50 for English and Hindi respectively.

The results from GermEval [20] displayed that the top performing systems came out to
be CNN and LSTM for the both the shared tasks. They achieved an average macro F1
score of 76.77% and 52.71 % for the binary-class coarse-grained subtask A and multi-class
fine-grained subtask B respectively.

Authors in [21], for the 6-class classification problem of toxic comments into respective
categories achieved a descent accuracy of 91.2% with CNN.

In [1], as the dataset is fairly imbalanced for each of the levels so the authors used per
class precision, recall and F1-score with the weighted average to compute the perfor-
mance of each of the models. For the detection of the offensive language, CNN came
out to be the best model with macro average F1 score of 0.80. For the categorization
of the offensive language, CNN was the best model with the average macro F1 score of
0.69. In the identification of the target of the offensive posts, both CNN and BiLSTM
performed equally well achieving an F1-macro of 0.47.

2.4 Current State of the Art

In this section we would be critically analyzing the above discussed related work on
the detection of abusive/ offensive language on the various online platforms. This will
eventually lead and form a strong basis to our contribution in the problem area through
this research project.

The work presented in [10], [8], [9] focussed on the detection of Cyber-Bullying on
the social media platform like Youtube and Twitter. Some of them modeled this as the
binary classification problem of a content being sensitive or not while one work took the
user characteristics and profile information into the account. One work also formulated
the problem as sentiment analysis and topic modeling.

Related work [21] focussed mainly on the detection and classification of Toxic Com-
ments into 6 respective categories namely toxic, severe toxic, obscene, threat, insult
and identity hate.
Background 23

Majority of the related work exists in the detection of the Hate Speech because of its
large presence on the social media when compared to the other types of the offensive
content on these platforms. The work presented in [12], [13], [11] are carried out with the
same objective. Some formulated the problem as the binary classification of distinguish-
ing racist content from the non-racist one, while some of them posed it as differentiating
the hateful comments from non-hateful/clean ones based on race, ethnicity and religion.
Work presented in [15], [16], [19] primarily contributed in distinguishing the hate speech
from general profanity and modeled the problem as 3-class classification task with the
annotation of the instances as Hate, Offensive or Clean.

The work in [18] was the identification of the Aggression on the social media par-
ticularly the English and Hindi language comments of Facebook and Twitter. The
contribution formulated the idea as the 3-class classification task of classifying the posts
into 3 categories namely non-aggressive, covertly aggressive and overtly aggressive.

Related work presented in [14], [20], [1] focussed on the detection of Offensive language
rather than the specific type of offence or the abuse on these platforms. Some modeled
the problem as the binary classification task of categorizing the content as abusive
or clean while other focussed on the further multi class classification of the detected
offensive German tweets as the profane, insult or abuse. The most recent work came
out with the very suitable dataset for the detection of offensive language with the 3
levels of annotation scheme helpful in identifying the type and the target of the offensive
posts for an in-depth analysis and moderation.

Upon analyzing each of the above work we can reach to the conclusion that the previous
studies revolving around the detection of abusive/offensive content on the social media
platforms can be classified into 5 main categories. These were mainly Cyber-bullying
detection, Hate speech detection, Aggression identification, Toxic comment classification
and the last but not least is the detection of overall Offensive content that includes all
these previous types.

We strongly believe that OLID dataset [1] aims to capture the differences and similar-
ities between the pre-existing datasets revolving around the above mentioned 5 tasks of
offensive language detection. This dataset also treats the problem as the whole because
it covers all these 5 aspects which was not evident in the previous studies. It also at the
same time enable the identification of the type and the target of the offensive tweets for
the further analysis.

As the OLID dataset was recently released in 2019, therefore it opens up a new oppor-
tunity for the researchers to explore this novel dataset for the further improvement in
Background 24

the field. We also aim to contribute and come with the different techniques in order to
make significant improvements after the work carried in [1]
Chapter 3

Problem - Offensive Language

Detection in Social Media

This chapter introduces us with the problem definition and description. This section
particularly give a brief about the motivation for choosing the topic of offensive language
detection on social media and the reasoning behind choosing the OLID dataset for our
problem followed by choice of sub-task withing the dataset that we are opting for our
research project. Next section will give a detailed description about the OLID dataset
giving us the insights about the features and the target class distribution. Finally we end
this chapter by the research objectives that we need to achieve to successfully contribute
in this problem area.

3.1 Problem Definition

Now a days with the increasing usage of social media platforms like Facebook and
Twitter, we often see people to misuse this freedom of speech. Some of them try to use
this platform to post the offensive or abusive things about a person or a group. This in
turn can negatively impact the mental health of a community/ group or an individual.
Considering the sensitivity of the topic, we aim to tackle this problem of detecting the
offensive language in social media through cutting edge techniques in Machine Learning,
Deep Learning and Natural Language Processing.

To proceed further with our topic we choose the Offensive Language Identification
Dataset (OLID) which was released in 2019. But the question is that, why we choose
this particular dataset when there exists various datasets in the problem domain. The
previous datasets only aims to capture a particular type of offense such as hate speech

25
Problem - Offensive Language Detection in Social Media 26

or cyber bullying, but OLID is a well diverse dataset which covers all the offense types.
Therefore we choose this recent novel dataset which opened up various new research
opportunities to contribute in.

There are three sub-tasks for this new dataset which involved the detection, predicting
the type and the target of the offensive tweet respectively. We particularly focus on
predicting whether the post is offensive or not which is the first and the most crucial
sub-task for this dataset making it a binary classification problem.

3.2 Dataset Description

To carry out the research project we would be using the publicly available benchmark
dataset [1] which is named as OLID (Offensive Language Identification Dataset). This
dataset generalizes the task of detecting the all types of offensive content that was de-
scribed in the background chapter using its three level annotation scheme. This types
of dataset is very first of its kind and opens up the new opportunity to explore it even
further which we also aim to do in our research project.

This OLID dataset annotate each tweet instance/post with the 3 level annotation scheme
meaning that each instance is labeled with 3 corresponding type of labels. The first level
denotes whether the tweet is offensive (OFF) or not (NOT). The second level identi-
fies the type of the offensive tweet which could be targeted insult (TIN) or un-targeted
(UNT). The third level identifies the target of the offensive post categorized mainly into
individual (IND), group (GRP) or other (OTH).

OLID is a collection of 14,100 annotated tweets obtained using the twitter API. The
training dataset consists of 13240 annotated tweets while the test partition contained
about 860 tweets. The whole problem can be formulated as the 3 sub-tasks of detecting
the offensive content, identifying the type of the offensive content and lastly to catego-
rize the target of the targeted offensive content.

The distribution of OLID dataset is reflected with the figure 3.1 and the distribution for
the training set is depicted below in the hierarchical order of annotation.

1. For sub-task A – The distribution of the each of the labels in the training set
for this subtask A is below

• OFF – 4480 tweets (33.23%)

• NOT- 8840 tweets (66.76%)
Problem - Offensive Language Detection in Social Media 27

Figure 3.1: Distribution of 3 levels of annotation in OLID

2. For sub-task B – The distribution of the each of the labels in the training set
for this subtask B is below

• TIN- 3876 (88.09%)

• UNT- 524 (11.90%)

3. For sub-task C – The distribution of the each of the labels in the training set
for this subtask C is below

• IND – 2407 (62.10%)

• GRP – 1074 (27.70%)
• OTH – 395 (10.19%)

3.3 Research Objectives

1. As our main task, we need to determine whether a tweet is offensive or not. Since
the dataset is new and not much exploration is done on that, we aim to explore
numerous techniques in various phases of the project pipeline and evaluate the
impact of each of them on the performance. Finally we, aim to develop and come
up with an efficient system for this task and try to improve the previous results
on the dataset.

2. In data exploration phase, we found out that this dataset is very messed up with
full noisy data to be fed into a Machine/Deep learning model. Therefore we will
perform a bunch of pre-processing steps to clean this dataset. Much details would
be provided in the next chapter.

3. The another important objective of the project is to balance the training set of
our OLID data by experimenting with various Undersampling and Oversampling
techniques for re-balancing it.
Problem - Offensive Language Detection in Social Media 28

(a) WordCloud- Non-Offensive Tweets (b) WordCloud- Offensive Tweets day

Figure 3.2: WordCloud for Non-Offensive and Offensive Tweets

4. For feature extraction, we will experiment with various available techniques like
TF-IDF, BOW, Uni-gram, Bi-gram, Tri-gram and their combinations. We will
also compare the performance with pre-trained embeddings like Word2Vec and
GloVe.

5. For the purpose of model building, we will apply a whole range of Classical Machine
learning models like SVM, Naive Bayes, Logistic Regression, Decision Tree along
with the ensembles like Random Forest, Gradient Boosting and Ada Boost to
classify our tweets.

6. Further to investigate in much detail, we will build the Deep learning models like
CNN, GRU, LSTM along with state of the art techniques such as BERT by Google,
fine-tuned on our dataset.

7. To evaluate the performance of our Machine learning and Deep Learning models,
we will be using per class Precision, Recall and F1-score. The overall performance
of the model is evaluated using the Macro F1 score as the test set is imbalanced
as evident in the exploration phase.

8. Finally, we will present the top Feature extraction mechanisms along with the
Machine/Deep learning models that worked best for our problem of detecting the
offensive posts. In the next chapter we would be reporting the results of our
experiments in much details through comparative analysis.
Chapter 4

Implementation Approach

In this chapter, we will start of with the experiment design and methodology where we
will discuss the techniques employed in data pre-processing, various feature extraction
mechanisms, techniques for handling the imbalance and various modeling techniques for
both Machine learning and Deep learning. Further we will proceed with the implementa-
tion details and the process flow of our research project in terms of softwares, platforms
and programming language used to experiment with the techniques in each phase of
the pipeline of our process flow. Then we present the results of various feature extrac-
tion techniques with variety of machine learning models followed by the results of deep
learning models on our dataset. This chapter ends with the evaluation and comparative
analysis of these machine learning and deep learning models to come up with the best
model for our problem.

4.1 Experiment Design and Methodology

4.1.1 Data Pre-processing

Text Pre-processing is the technique of cleaning the text data so as to discard the
irrelevant information from it which makes it more meaningful and reduces the size of
our dataset. This can reduce the training time of our model as well.

Our dataset OLID has two features, first one is tweet and the second one is the cor-
responding label Offensive(OFF) or Not-Offensive(NOT). Therefore there is a need to
pre-process the tweet column for both our training and the test set. If the cleaned text
is fed into our algorithms, then it will produce more reliable results when compared
with the uncleaned or garbage data where the model will not be able to interpret much
information which is semantically or syntactically insignificant.
29
Implementation Approach 30

Below are the sequence of steps in text pre-processing that can be employed with respect
to our dataset to get a better version of the twitter text.

• Tokenization: Tokenization is the process of splitting of sentences or paragraph

(sequence of strings) into respective tokens which are individual words. This is
important as it gives us the lot of insights about the meaning of the text through
the extraction of words present in it. It can give us the total word count or
individual word frequency.

• Stemming: Inflected forms of a word is formed by adding the affixes to the

root/base form of the word. Stemming is the process of converting or reducing the
inflected or derivationally related forms of a word to its base form by removing
the suffix from a word. It is not necessary that the root word produced has some
dictionary meaning. For example beautiful and beautifully will be both stemmed
to beauti which has no dictionary meaning in english.

• Lemmatisation: Lemmatisation also aims to convert the inflected form of a word

to its base form by removing the inflectional endings. But this root form generated
is the dictionary form of a word called Lemma. It takes into account the meaning
of word in a sentence to convert the word into its root form. The process takes
the support of a vocabulary and performs the morphological analysis of words.
For example, beautiful and beautifully are lemmatised to beautiful and beautifully
without changing the meaning of words. But good, better and best are lemmatized
to good since all the words have same meaning

• Lower Case Conversion: This is mainly converting all the words in our text to
lowercase form. This is done to normalize our textual data.

• Stop Words Removal: When we have to create the meaningful features from
the text, then these stopwords have no significance. These are generally most
frequently occurring words in a text which can be articles such as a, an, the,
conjunctions such as for, yet, but, so or prepositions such as in, towards, before.
This is done to limit the computation time by reducing the size of our vocabulary.

• Punctuation Removal: Punctuations are the symbols which are used in writing
to clarify the meaning of the text by separating the sentences and their elements.
These are of no use to convey the meaning of text and only used for good readability
so we will remove these from our dataset.

• User and URL Mentions Removal: We have our text data full of USER and
URL mentions which doesn’t convey any meaning to our text, so we have removed
their presence from our training and the test data.
Implementation Approach 31

• Special Characters Removal: Special characters are often the non-alphanumeric

characters which does not make any sense for our task, so we choose to remove
them.

• Numbers Handling: There are two ways to handle the numbers in the text data.
We can either remove them or convert them to their textual form. In our case, as
numbers don’t make any contribution in predicting whether the tweet is offensive
or not, we choose to remove them.

• Emojis Handling: Emojis are visual representations that convey an idea or an

emotion while including them in a text. There are two ways to handle them.
First is to substitute an emoji with a textual word and second one is to simply
remove them. For our task we initially proceeded with removing them as they
don’t contribute much in our text data.

• Contraction Expansion: Contractions are the short form of words which are
created using removing one of the vowels from them. For example contraction of
do not is don’t. We need to convert these contractions to the expansion forms
which helps in text standardization.

• White Space Removal: After we perform all the pre-processing steps there is
the generation of the white spaces that are of no use as will take unnecessary of
memory space so better to remove them.

4.1.2 Feature Extraction Techniques

Feature Extraction with respect to NLP tasks is to extract the features and produce the
numerical representations of the text data so that it can be supported by the Machine
Learning and Deep Learning algorithms for predictive calculations.

4.1.2.1 Feature Extraction with Statistical Models

1. Bag Of Words: Raw text data has the variable length, but most of the machine
learning algorithms expect the numerical feature vectors as the input. Vectoriza-
tion is the process of converting the text documents into the numerical feature
vectors. Bag of Words [22] does this with the help of Tokenisation, Counting and
Normalization.

This model takes in the whole corpus of text documents and produces a sparse
matrix where each row in the matrix represents the word frequency of one docu-
ment and each column or feature is the corresponding token or unique word in the
Implementation Approach 32

whole corpus. This model ignore the relative position information of the words
in the document and based only on the word occurrences in each document. The
basic idea is shown by the below figure 4.1

Figure 4.1: Feature Extraction through BOW

2. Bag of Words Unigram, Bigram, Trigram and their Combinations: A

n-gram is the sequence of n words which aims to capture more meaning from
the text document by preserving the sequence information (order in which the
word appears in the document) when compared to the simple unigram BOW. This
representation changes both the scope and the vocabulary of our model. This way,
the each feature of our obtained matrix through n-grams model is the combination
of n consecutive words that occur in the whole corpus.

Each row represents the combination frequency of each feature of that particular
document in the text corpus. Bi-Gram model forms the features as all the possible
combinations of two consecutive words occurring in the corpus text whereas the
Tri-Gram captures the possible combinations of three consecutive words occurring
in the corpus for feature extraction. Therefore, we can either individually consider
the Unigram, Bigram, Trigram or Ngram in BOW model or can make different
combinations of them for extracting the important features from the text.

3. Tf–Idf term weighting: When we create the vector from a document based
on word frequency, then the most frequent words/terms which carry very little
information about the content, starts to dominate over the rarer words which are
more significant and interesting. So instead of scoring the terms in the documents
by the simple frequency in one document, we rescale it to the frequency of the
words appearing in all the documents. This will incur the penalty for the words
that are frequent in a document but also frequent across all the documents.
Implementation Approach 33

Tf in Tf-Idf denotes the score based on frequency of the word in the current
document and Idf gives the score based on how rare that word is across all the
documents. The multiplication of both these scores gives us the final numerical
weight of the word/term when converting a document to a vector representation.
The basic idea of Tf-Idf is that, the most frequent words are less contextually
important than the domain specific or rare words which are more significant in the
corpus. Therefore this approach of feature extraction tends to associate each word
of the documents present in the corpus with a less weight if it is frequent across
all the documents and a high weight if it is rare across all the documents.

4. Tf–idf term weighting Unigram, Bigram, Trigram and their Combina-

tion: We have seen the concepts of Bi-grams, Tri-grams and N-grams in the above
paragraphs. There we saw that, this idea can be integrated with BOW method
for extracting the useful features. Here also we can use these combinations with
Tf-idf weighting scheme for feature extraction.

4.1.2.2 Feature Extraction with Word Embeddings

1. Word2Vec: This technique was introduced by the researchers of Google in 2013.

The paper [23] aims to address the shortcomings of the previous approaches for
vectorizing the text documents like BOW or Tf-Idf. These traditional approaches
results in the large sparse matrices giving only the information about the docu-
ments but not display the meaning of the words. This Word2Vec technique takes
the semantic meaning into consideration. This word embedding technique receives
the input as the large corpus of text and produces the vector space field of sev-
eral hundreds of dimensions with the help of shallow neural network; where each
unique word in the corpus is assigned the vector in that space. In short Word2Vec
is a methodology to convert the words to their numerical vector form preserving
the meaning and the context of the words in the vector space. Closer words in the
vector space signify that they share the same context and are similar.

2. GloVe : The abbreviation GloVe denotes the Global Vectors was first introduced
in the work [24] at Stanford University. This technique also aims to find the d-
dimensional vector representations for each word in vector space. The distance
between these words is dependent on the semantic similarity. GloVe is based on
matrix factorization techniques on the word-context matrix and is one of the most
popular embeddings used in NLP.

3. BERT: BERT stands for Bidirectional Encoder Representations from Transform-

ers [25] is the significant state of the art technique in Natural Language Un-
derstanding given by the researchers of Google AI Language. The main aim of
Implementation Approach 34

this BERT model is to produce the pre-trained vector representations of the words
and can be used for range of NLP tasks by just little fine tuning and addition of
just one output layer. Bi-directional in the model signifies that it has learnt the
representation from both left and right context for all the processing layers which
makes it even more efficient at understanding the context of words in a sentence.
Encoder representation means that it refers to the algorithm which has trained this
model to represent the words from huge corpus of text which is from Wikipedia
and Google’s book corpus estimating around 3300 million words. Architecture of
BERT is based on Transformer layers containing multiple attention heads.

4.1.3 Handling Imbalance

Upon the Exploratory data analysis, we found that our training set was imbalanced as
the Offensive tweets were 33% and Non-offensive ones were 66%. We have experimented
with various oversampling and undersampling techniques for handling this issue. It is
very essential to handle this large imbalance because otherwise the results would be
biased for the majority class only. Below figure 4.2 shows the class imbalance in the
training set of OLID. The main goal of these resampling techniques is to balance the
distribution of majority and minority classes in the dataset.

Figure 4.2: Training set distribution of class labels

4.1.3.1 Oversampling Techniques

The main goal of Oversampling techniques is to rebalance the dataset by increasing the
instances of minority class to match the distribution of majority class in the dataset.
This overall increases the size of the dataset.

The main drawbacks of Oversampling is that our model might overfit on the training
data.
Implementation Approach 35

1. Random Oversampling: This technique will randomly increase the size of mi-
nority class by randomly duplicating instances from the minority class until the
distribution of the b classes are equal and balanced. This is similar to observation
weighting.

2. SMOTE: This technique [26] allows us to create the new artificial data points in
our dataset. Below are the steps that are need to be taken in order to create these
artificial points.

2.a) Select a data point from the minority class.

2.b) Select the k nearest neighbours of the selected data point in the feature

space.

2.c) Select any one random neighbours from k of them.

2.d) Calculate the difference vector between the original data point and

this randomly selected neighbour.

2.e) Multiply this vector by the random number between 0 and 1 to get

the intermediate random value.

2.f ) This random value obtained is the artificial data point in the feature

space that will lie between the original data point and it’s neighbour.

4.1.3.2 Undersampling Techniques

The main goal of Undersampling techniques is to rebalance the dataset by decreasing

the instances of majority class to match the distribution of minority class in the dataset.
This overall decreases the size of the dataset.

The main drawback of Undersampling is that we may loose the important information
from our dataset.

1. Random Undersampling: This technique will randomly remove data from the
majority class until the distribution of the two classes are equal and balanced.

2. Tomek Links: Tomek links is the method of under-sampling and removes the
instances of majority class in areas of significant overlap between the classes.

• Let x be an instance of class A and y an instance of class B.

• Let d(x, y) be the distance between x and y.
• (x, y) is a Tomek link, if for any instance z, d(x, y) < d(x, z) < d(y, z).
Implementation Approach 36

• This means that y is the nearest neighbour of x or that x is the nearest

neighbour of y.
• Tomek links can be used to reduce the level of overlap between the majority
and minority classes.

4.1.4 Algorithmic Modeling

Algorithmic Modeling is the process of building a range of both Machine learning and
Deep Learning models to carry out the predictions on the unseen or the test data.

4.1.4.1 Machine Learning Based Models

Below are the Machine Learning algorithms that we have used to determine whether a
given tweet is offensive or not.

1. Support Vector Machine: The Support Vector Machine was first introduced in
the work [27]. They are the supervised machine learning algorithms that are used
for both classification and regression tasks which can be linear or non-linear. The
main of this algorithm is to come up with the hyper-plane such that the marginal
width is maximum. Marginal width is the distance between the 2 marginal lines
where each marginal line is at closest data point of each of the classes present in
our dataset. These closest points belonging to each class from the hyper plane are
called support vectors [28]. SVM works very well in the cases of high dimensional
data and also in the cases where number of dimensions are greater than the number
of instances or data points.

2. Naive Bayes: This type of classifiers are the class of probabilistic classifiers
??. The algorithm is based on the mathematical formulation of Bayes Theorem.
It is called naive because the assumption of this algorithm is that, the features
are independent of other features in the dataset; which is never fully true. The
formulation of Naive Bayes Classifiers is depicted below by the equation 4.1 where
’p(x)’ is the probability of feature given values for ’y’ features.

p
Y
p(x|y) = p (xi |y) (4.1)
i=1

3. Logistic Regression: This supervised machine learning algorithm as presented

in the book [29] is generally used for classification problems rather than regression.
In linear regression the output can be anything positive or negative values, but
Implementation Approach 37

here in logistic regression, we squeeze our output between 0 and 1. The aim here
also is to find the decision surface which best separates the 2 classes in our dataset.
The data points below this decision line will result in positive values and the ones
above the line results in negative values. These values are passed into the sigmoid
function such that the positive values are greater than 0.5 and the negative values
are less than 0.5. Therefore we can easily classify our 2 classes. The idea can be
extended to multi-class classification problems also, but generally works best for
binary classification problems.

4. Decision Tree: A decision tree [30] is a graphical representation of a decision

process. This type of machine learning model takes input as the vector of feature
values and produces the prediction class/value. It is a flow like structure where the
nodes of the tree represents the features in our dataset and the leaves denotes the
predicted class. Nodes are the test on the attribute, each branch is the outcome
of the test and leaf or terminal node holds a class label where the topmost node
is the root node. These algorithms are simple to interpret and, understand and
visualize. These also explicitly performs the feature selection and can handle both
numerical and categorical data.

5. Ensemble Techniques: Ensemble methods in Machine learning utilize or com-

bine the results of various learning algorithms to obtain a better predictive perfor-
mance that could be obtained from the same base learner or multiple base learners.
All of the ensembles have two things in common. First is that, they construct mul-
tiple, diverse predictive models from adapted versions of the training data. Second
is that they combine the predictions of these models in some way, often by simple
averaging or voting. Ensembles can be broadly classified into two types which
are Bagging (Random Forest [31]) and Boosting (Gradient Boosting [32]/ Ada
Boosting)

4.1.4.2 Deep Learning Based Models

Some of the Deep Learning Architectures that we have experimented are stated below:-

1. Artificial Neural Networks

2. Convolutional Neural Netoworks

3. Long Short Term Memory Networks

4. Bidirectional Long Short Term Memory Networks

5. Hybrid Architectures of CNN and LSTM

Implementation Approach 38

4.1.5 Evaluation Metrics

Evaluation Metrics are the ways to measure the performance and the quality of our Ma-
chine learning or Deep learning models. These metrics should be choosen very carefully
depending upon the test dataset we have at our disposal. For example, if we choose
accuracy on the imbalanced test set then it might give a good overall accuracy, but
actually the majority class accuracy is dominating over minority class accuracy. So here
a better metrics could be F1-score or confusion matrix to get the actual insights of the
performance of all the classes present in our test dataset.

To evaluate the performance of our Machine learning and Deep Learning models for
detecting the offensive language, we will be using per class Precision, Recall and F1-
score. The overall performance of the model is evaluated using the Macro F1 score as
the test set is imbalanced as evident in the below figure 4.3

Figure 4.3: Test set distribution of class labels

Some concepts that will be important to understand the further concepts are as follows
and can be depicted by the figure 4.4

• True Positives (TP)-Number of instances of majority class that were correctly

classified as majority class.

• True Negatives (TN)-Number of instances of minority class that were correctly

classified as minority class.

• False Positives (FP)-Number of instances of minority class that were incorrectly

classified as majority class.

• False Negatives (FN)-Number of instances of majority class that were incor-

rectly classified as minority class.
Implementation Approach 39

Figure 4.4: Representation of Confusion Matrix

Below are some evaluation metrics that we will be using to determine the prediction
performance of our models on this OLID test set.

1. Precision: It is the measure of the positive predictive value. It tells how confident
we can be that any instance predicted as belonging to a certain class actually
belongs to that class.The formula for the precision for any class is as follows

T rue P ositives
P recision = (4.2)
T rue P ositives + F alse P ositives

2. Recall:It is the measure of sensitivity or the true positive rate. It tells how
confident we can be that all instances belonging to a specific class have been
correctly classified by the model. The formula for the recall for any of class present
in the target feature is as

T rue P ositives
Recall = (4.3)
T rue P ositives + F alse N egatives

3. F1-Score: It is interpreted as the weighted average of precision and recall for a

particular class. For F1 score to be high of any class, both the precision and recall
values should be high as well. The best value for f1 score is 1 and worst value is
0. The formula for F1-score is

Recall ∗ P recision
F 1 Square = 2 ∗ (4.4)
Recall + P recision

4. Confusion Matrix: It is the representation of the information about the actual

and predicted classifications by the classification algorithm. It gives the insight
into the accuracy of each of the classes that are present in the target feature
meaning that we get the number of test instances for each class that are correctly
as well incorrectly predicted by our model.

5. Classification report: This report provides an insight into the precision, recall
and f1 score broken down by the classes present in our dataset. For our problem
as the data is imbalanced, so we will not see the normal accuracy, rather our
evaluation for this biased classification would be on the basis of confusion matrix
(indication of the accuracy of each of the class). Also we would be considering
Implementation Approach 40

the f1 score across each of the 2 classes which tell about how good the balance
is between the precision and recall values. Moreover this classification report also
generates the mean and weighted average of precision, recall and f1 score for each
of the individual classes in our target feature. So we would be considering the
macro average and weighted average of f1 scores for both the classes as the one of
the metric for evaluation.

4.2 Implementation Details and Results

4.2.1 Process Flow and Implementation Details

Our whole process of carrying out this research project is divided into 5 major steps and
implemented in Python Programming Language :-

1. Text Pre-processing: For carrying out this step of processing our text into some
cleaned version, we took the help of Natural Language Toolkit [33] (NLTK). This
platform provides the various libraries and built in modules to deal with the textual
data.

2. Feature Extraction: We experimented with various feature extraction techniques

like BOW and TF-IDF present in Scikit-learn API [22]. Further we explored the
pre-trained word embeddings like Word2Vec and GloVe with the help of Gensim
library [34].

3. Handling Imbalance: For handling the major issue of imbalance in our training
data, we leveraged the imbalanced-learn API [35] whose packages and modules we
have utilized for this project. This toolkit implemented the overall solution of im-
balance with four techniques. These were under-sampling, over-sampling, combi-
nation of both over sampling and under sampling and ensemble learning methods.
We only experimented with over-sampling and under-sampling approaches.

4. Model Building: For building and using various machine learning models for
detecting the offensive tweets, we used the Scikit-learn API [22]. And for the deep
learning algorithms, we used the TensorFlow API [36]

5. Model Evaluation: For evaluating our model performance, we came up with

various evaluation metrics available in Scikit-learn API [22] such as precision, recall
and F1-score. Finally we reported the model performance with confusion matrix
insights and macro F1 score.
Implementation Approach 41

4.2.2 Various Machine Learning Models with Different Feature Ex-

traction Techniques

We found that Oversampling techniques such as SMOTE and Random Oversampling

did not performed well when compared with Random Undersampling. The results of
these oversampling methods on the test set were biased towards the majority class that
is Non-Offensive tweets. Therefore, we choose Random Undersampling technique to
compare various Feature Extraction Mechanisms.

In this sub-section we will present the results of trying the different Feature Extraction
Mechanisms like BOW, Tf-idf along with their Bi-gram and Trigram Combinations with
Random Undersampling. We will also see the results with pre-trained embeddings like
Word2Vec and GloVe.

4.2.2.1 Feature Extraction with BOW

Results of BOW (Uni-gram) with Random Undersampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.70 0.77 0.67
Naive Bayes 0.67 0.68 0.73
Logistic Regression 0.73 0.83 0.65
Decision Tree 0.66 0.72 0.66
Random Forest 0.73 0.81 0.66
Gradient Boosting 0.74 0.94 0.49
Ada Boost 0.73 0.92 0.49
Table 4.1: Evaluation Metric for Feature Extraction with BOW (Uni-gram)

Results of BOW (Uni-gram + Bi-gram) with Random Undersampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.72 0.90 0.50
Naive Bayes 0.64 0.62 0.77
Logistic Regression 0.73 0.84 0.62
Decision Tree 0.66 0.73 0.64
Random Forest 0.71 0.87 0.54
Gradient Boosting 0.73 0.94 0.48
Ada Boost 0.73 0.92 0.49
Table 4.2: Evaluation Metric for Feature Extraction with BOW (Uni-gram + Bi-
gram)
Implementation Approach 42

Results of BOW (Uni-gram + Bi-gram + Tri-gram) with Random Under-

sampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.69 0.94 0.40
Naive Bayes 0.64 0.62 0.78
Logistic Regression 0.73 0.84 0.62
Decision Tree 0.68 0.73 0.67
Random Forest 0.73 0.91 0.52
Gradient Boosting 0.73 0.93 0.48
Ada Boost 0.73 0.92 0.49
Table 4.3: Evaluation Metric for Feature Extraction with BOW (Uni-gram + Bi-gram
+ Tri-gram)

4.2.2.2 Feature Extraction with Tf-idf

Results of Tf-idf (Uni-gram) with Random Undersampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.74 0.85 0.62
Naive Bayes 0.68 0.70 0.72
Logistic Regression 0.72 0.83 0.62
Decision Tree 0.63 0.67 0.66
Random Forest 0.71 0.83 0.60
Gradient Boosting 0.74 0.92 0.51
Ada Boost 0.71 0.93 0.46
Table 4.4: Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram)

Results of Tf-idf (Uni-gram + Bi-gram) with Random Undersampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.71 0.80 0.64
Naive Bayes 0.67 0.67 0.74
Logistic Regression 0.70 0.78 0.65
Decision Tree 0.65 0.69 0.65
Random Forest 0.70 0.85 0.53
Gradient Boosting 0.72 0.93 0.48
Ada Boost 0.71 0.92 0.46
Table 4.5: Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram + Bi-
gram)
Implementation Approach 43

Results of Tf-idf (Uni-gram + Bi-gram + Tri-gram) with Random Under-

sampling

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.71 0.80 0.63
Naive Bayes 0.67 0.69 0.74
Logistic Regression 0.70 0.77 0.65
Decision Tree 0.66 0.70 0.67
Random Forest 0.69 0.84 0.54
Gradient Boosting 0.73 0.93 0.48
Ada Boost 0.71 0.92 0.47
Table 4.6: Evaluation Metric for Feature Extraction with Tf-idf (Uni-gram + Bi-gram
+ Tri-gram)

4.2.2.3 Feature Extraction with Word2Vec

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.73 0.78 0.68
Naive Bayes 0.69 0.80 0.68
Logistic Regression 0.74 0.72 0.74
Decision Tree 0.64 0.70 0.72
Random Forest 0.70 0.85 0.70
Gradient Boosting 0.69 0.90 0.55
Ada Boost 0.73 0.91 0.50
Table 4.7: Evaluation Metric for Feature Extraction with Word2Vec

4.2.2.4 Feature Extraction with GloVe

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
SVM 0.75 0.80 0.71
Naive Bayes 0.70 0.82 0.70
Logistic Regression 0.73 0.73 0.75
Decision Tree 0.67 0.71 0.69
Random Forest 0.72 0.87 0.73
Gradient Boosting 0.71 0.92 0.60
Ada Boost 0.74 0.93 0.55
Table 4.8: Evaluation Metric for Feature Extraction with GloVe
Implementation Approach 44

4.2.3 Results of Various Deep Learning Models

Model Macro F1- Majority Class Minority Class

Score Accuracy Accuracy
Artificial Neural Net- 0.73 0.94 0.52
work (BOW)
CNN+LSTM (BOW) 0.75 0.93 0.60
CNN+BiLSTM 0.61 0.73 0.53
(GloVe)
CNN+BiLSTM (BOW) 0.77 0.89 0.64
BERT 0.82 0.91 0.70
Table 4.9: Evaluation Metric for Deep Learning Models

4.3 Evaluation and Analysis

In this section we will be doing the comparative analysis of various feature extraction
methods to come up with the top Machine learning models for each of these methods.
We will be selecting the best models based on the Macro F1 score and the insights from
the confusion matrix (majority and minority class accuracy). We will also present our
top performing Deep learning models in this section. All this analysis will help us to
come up with the best model for predicting the offensive tweets with our dataset OLID.

4.3.1 Comparative Analysis of Various Feature Extraction techniques

along with Machine Learning and Deep Learning Models

Feature Extraction/Machine Macro F1- Majority Class Minority Class

Learning Model Score Accuracy Accuracy
BOW(Unigram) + SVM 0.70 0.77 0.67
BOW(Unigram+Bigram+Trigram) 0.68 0.73 0.67
+ Decision Tree
TF-IDF(Unigram+Bigram+Trigram) 0.70 0.77 0.65
+ Logistic Regression
TF-IDF(Unigram) + Naive Bayes 0.68 0.70 0.72
Word2Vec + Logistic Regression 0.74 0.72 0.74
GloVe + SVM 0.75 0.80 0.71
BOW + CNN-BiLSTM 0.77 0.89 0.64
BERT 0.82 0.91 0.70

Table 4.10: Comparative Analysis of Various Feature Extraction, Machine Learning

and Deep Learning Techniques
Implementation Approach 45

4.3.2 Best Model for Offensive Language Detection

From the above comparative analysis of various Machine Learning and Deep Learning
models, we found that the BERT achieved the highest F1 score of 0.82 beating the
performance of the previous work carried in [1].
Chapter 5

Conclusions and Future Work

This chapter starts with the main conclusions that we have come up with, in terms of
background, problem description and the solution approach for this research project.
Then we end this chapter with the possible areas of future work that can be undertaken
in order to further contribute meaningfully to the topic of problem area.

5.1 Conclusion

The following are the main conclusion points that are being achieved in effort to con-
tribute to the problem of detecting the offensive language on the social media platforms
through this project:-

1. The extensive literature review from the previous work in the area of Offensive
Language Detection helped us to explore the novel benchmark dataset - OLID in
the great depth as it was recently created.

2. We explored the various techniques of handling the imbalance in the training

set, but for our dataset; the random under-sampling worked best minimizing the
difference between the majority and the minority class accuracy and giving the
higher macro F1 scores.

3. We performed an in depth comparative analysis of various feature extraction mech-

anisms, machine learning and deep learning models to explore the dataset even
further and set the significant base point from which the future research could
proceed with the dataset and the results achieved with this thesis work.

46
Conclusions 47

4. Finally the evaluation and the analysis of our comparisons lead to the fact that the
state of the art model by Google - BERT outperformed the previous performance
achieving the F1 macro score of 0.82 on this dataset.

5.2 Future Work

Below are some future research directions that we have planned if we would have some
more time to contribute on the problem area.

1. We have only focused on the sub-task A that is the detection of offensive language
for this research project. But on the same dataset there are two more sub-tasks A
and B which focused on determining the type and the target of the offensive post.
This can even lead us to further insights and analysis of offensive posts detected
on the social media to take the relevant actions. In future, we would like to solve
these both sub-tasks using variety of approaches to contribute more in the domain.

2. Human language is very much diverse and certain posts may not look offensive
from the surface, but actually they are when analysed by the human annotator.
In future we will like to conduct an analysis to capture the syntactic and semantic
features along with their combination and other pre-trained features. This have
actually the potential to improve our performance than what we have achieved
now.

3. We have clearly seen that to tackle the issue of imbalance, the random under-
sampling has performed the best with our dataset. But at the same time the
big problem with this approach is that we have lost a good amount of significant
information which might be useful in the training of our model. This approach
have reduced our training data to a great amount, therefore we lost important
features from the dataset. In future we may come up with the suitable approach
and exploration that may tackle this problem of huge imbalance in the dataset
efficiently without the reduction in training data.

4. In future we would also like to combine both machine learning and deep learning
models to create a significant ensemble model which can help to improvise upon
the current results.

5. In our previous chapter, it is quite visible that deep learning models are not per-
forming that good because of less depth in their architectures. In future we aim
to increase the complexity of these models by increasing the layers and the config-
uration settings given the computational resources and the time.
Conclusions 48

6. We also aim to perform the hyper-parameter tuning of both machine learning and
deep learning models to come up with the optimal values of these hyper-parameters
for our dataset. This may also contribute in the boosting of the performance by
our models.
Bibliography

[1] M. Zampieri, S. Malmasi, P. Nakov, S. Rosenthal, N. Farra, and R. Kumar, “Pre-

dicting the type and target of offensive posts in social media,” in Proceedings of
the 2019 Conference of the North American Chapter of the Association for Com-
putational Linguistics: Human Language Technologies, Volume 1 (Long and Short
Papers), 2019, pp. 1415–1420.

[2] K. Keppner, “Artificial intelligence in supply chain planning – why a hybrid ai

concept is the better choice,” 2018.

[3] S. Kassel, “Predicting building code compliance with machine learning models,”
2017.

[4] J. K. Gill, “Automatic log analysis using deep learning and ai,” 2018.

[5] J. Chugh, “Types of machine learning and top 10 algorithms everyone should know,”
2018.

[6] Shriyanka, “What is natural language processing ?” 2017.

[7] A. LLC, “Ai, machine learning (ml) and natural language processing (nlp),” 2019.

[8] K. Dinakar, R. Reichart, and H. Lieberman, “Modeling the detection of textual

cyberbullying,” in fifth international AAAI conference on weblogs and social media,
2011.

[9] J.-M. Xu, K.-S. Jun, X. Zhu, and A. Bellmore, “Learning from bullying traces in
social media,” in Proceedings of the 2012 conference of the North American chap-
ter of the association for computational linguistics: Human language technologies.
Association for Computational Linguistics, 2012, pp. 656–666.

[10] M. Dadvar, D. Trieschnigg, R. Ordelman, and F. de Jong, “Improving cyberbullying

detection with user context,” in European Conference on Information Retrieval.
Springer, 2013, pp. 693–696.

49
Conclusions 50

[11] I. Kwok and Y. Wang, “Locate the hate: Detecting tweets against blacks,” in
Twenty-seventh AAAI conference on artificial intelligence, 2013.

[12] P. Burnap and M. L. Williams, “Cyber hate speech on twitter: An application

of machine classification and statistical modeling for policy and decision making,”
Policy & Internet, vol. 7, no. 2, pp. 223–242, 2015.

[13] N. Djuric, J. Zhou, R. Morris, M. Grbovic, V. Radosavljevic, and N. Bhamidipati,

“Hate speech detection with comment embeddings,” in Proceedings of the 24th
international conference on world wide web, 2015, pp. 29–30.

[14] C. Nobata, J. Tetreault, A. Thomas, Y. Mehdad, and Y. Chang, “Abusive language

detection in online user content,” in Proceedings of the 25th international conference
on world wide web, 2016, pp. 145–153.

[15] T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automated hate speech de-
tection and the problem of offensive language,” in Eleventh international aaai con-
ference on web and social media, 2017.

[16] S. Malmasi and M. Zampieri, “Detecting hate speech in social media,” arXiv
preprint arXiv:1712.06427, 2017.

[17] Z. Waseem, T. Davidson, D. Warmsley, and I. Weber, “Understanding abuse: A

typology of abusive language detection subtasks,” in Proceedings of the First Work-
shop on Abusive Language Online, 2017, pp. 78–84.

[18] R. Kumar, A. K. Ojha, S. Malmasi, and M. Zampieri, “Benchmarking aggression

identification in social media,” in Proceedings of the First Workshop on Trolling,
Aggression and Cyberbullying (TRAC-2018), 2018, pp. 1–11.

[19] S. Malmasi and M. Zampieri, “Challenges in discriminating profanity from hate

speech,” Journal of Experimental & Theoretical Artificial Intelligence, vol. 30, no. 2,
pp. 187–202, 2018.

[20] M. Wiegand, M. Siegel, and J. Ruppenhofer, “Overview of the germeval 2018 shared
task on the identification of offensive language,” 2018.

[21] S. V. Georgakopoulos, S. K. Tasoulis, A. G. Vrahatis, and V. P. Plagianakos, “Con-

volutional neural networks for toxic comment classification,” in Proceedings of the
10th Hellenic Conference on Artificial Intelligence, 2018, pp. 1–6.

[22] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,

M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay, “Scikit-learn: Machine
Conclusions 51

learning in Python,” Journal of Machine Learning Research, vol. 12, pp. 2825–2830,
2011.

[23] T. Mikolov, K. Chen, G. Corrado, and J. Dean, “Efficient estimation of word rep-
resentations in vector space,” arXiv preprint arXiv:1301.3781, 2013.

[24] J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word rep-
resentation,” in Proceedings of the 2014 conference on empirical methods in natural
language processing (EMNLP), 2014, pp. 1532–1543.

[25] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep
bidirectional transformers for language understanding,” in Proceedings of the 2019
Conference of the North American Chapter of the Association for Computational
Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers).
Minneapolis, Minnesota: Association for Computational Linguistics, Jun. 2019,
pp. 4171–4186. [Online]. Available: https://www.aclweb.org/anthology/N19-1423

[26] N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, “Smote: syn-

thetic minority over-sampling technique,” Journal of artificial intelligence research,
vol. 16, pp. 321–357, 2002.

[27] J. A. Suykens and J. Vandewalle, “Least squares support vector machine classifiers,”
Neural processing letters, vol. 9, no. 3, pp. 293–300, 1999.

[28] N. Cristianini, J. Shawe-Taylor et al., An introduction to support vector machines

and other kernel-based learning methods. Cambridge university press, 2000.

[29] D. W. Hosmer Jr, S. Lemeshow, and R. X. Sturdivant, Applied logistic regression.

John Wiley & Sons, 2013, vol. 398.

[30] S. R. Safavian and D. Landgrebe, “A survey of decision tree classifier methodology,”

IEEE transactions on systems, man, and cybernetics, vol. 21, no. 3, pp. 660–674,
1991.

[31] A. Liaw, M. Wiener et al., “Classification and regression by randomforest.”

[32] J. H. Friedman, “Greedy function approximation: a gradient boosting machine,”

Annals of statistics, pp. 1189–1232, 2001.

[33] E. Loper and S. Bird, “Nltk: The natural language toolkit,” in Proceedings of
the ACL-02 Workshop on Effective Tools and Methodologies for Teaching Natural
Language Processing and Computational Linguistics, 2002, pp. 63–70.

[34] R. Řehůřek and P. Sojka, “Software Framework for Topic Modelling with Large
Corpora,” in Proceedings of the LREC 2010 Workshop on New Challenges for NLP
Frameworks. Valletta, Malta: ELRA, May 2010, pp. 45–50.
 Bibliography 52

[35] G. Lemaı̂tre, F. Nogueira, and C. K. Aridas, “Imbalanced-learn: A python

toolbox to tackle the curse of imbalanced datasets in machine learning,” Journal
of Machine Learning Research, vol. 18, no. 17, pp. 1–5, 2017. [Online]. Available:
http://jmlr.org/papers/v18/16-365.html

[36] M. Abadi, A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.

Corrado, A. Davis, J. Dean, M. Devin, S. Ghemawat, I. Goodfellow, A. Harp,
G. Irving, M. Isard, Y. Jia, R. Jozefowicz, L. Kaiser, M. Kudlur, J. Levenberg,
D. Mané, R. Monga, S. Moore, D. Murray, C. Olah, M. Schuster, J. Shlens,
B. Steiner, I. Sutskever, K. Talwar, P. Tucker, V. Vanhoucke, V. Vasudevan,
F. Viégas, O. Vinyals, P. Warden, M. Wattenberg, M. Wicke, Y. Yu,
and X. Zheng, “TensorFlow: Large-scale machine learning on heterogeneous
systems,” 2015, software available from tensorflow.org. [Online]. Available:
https://www.tensorflow.org/

View publication stats