Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Hate Speech, Offensive Language Detection and

Blocking on Social Media Platform using Feature
Engineering Techniques and Machine Learning
Algorithms a Comparative Study
By

Mwayi Malemia
21251377011

Guide

Dr. Glorindall

Project Report

Submitted

In partial fulfillment of the requirements for the

Masters of Science in Computer Science

April, 2023

DMI - ST. EUGINE UNIVERSITY

ZAMBIA CAMPUS

IJISRT23MAY166 www.ijisrt.com 1129

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
ABSTRACT
The increasing use of social media and information sharing has given major benefits to humanity. However, this has
also given rise to a variety of challenges including the spreading and sharing of hate speech messages. Thus, to solve this
emerging issue in social media sites, recent studies employed a variety of feature engineering techniques and machine
learning algorithms to automatically detect the hate speech messages on different datasets. However, to the best of my
knowledge, not much research has been done to compare the variety of machine learning algorithms to evaluate which
machine learning algorithm outshine on a standard publicly available dataset. Hence, the aim of this paper is to compare
the performance of machine learning algorithms to appraise their performance on a publicly available dataset having
three distinct classes. The study has proved that the bigram features when used with the support vector machine algorithm
best performed with 79% off overall accuracy. My study holds practical implication and can be used as a baseline study in
the area of detecting automatic hate speech messages.

IJISRT23MAY166 www.ijisrt.com 1130

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
TABLE OF CONTENTS

Title Page no.

CHAPTER ONE 1132
INTRODUCTION
 Background of the Study 1132
 Statement of the Problem 1132
 Objectives of the Study 1132
 Research Question 1132
 Significance of the Study 1132
 Scope of the Study 1133
CHAPTER TWO 1134
LITERATURE REVIEW
 Introduction 1134
 Main Literature Review 1134
CHAPTER THREE 1136
RESEARCH DESIGN
 Introduction 1136
 Research Design 1136
 Population of the Study 1138
 Sampling Procedure 1138
 Sample Size 1138
 Sampling Area 1138
 Sources of Data Collection 1138
 Methods of Data Collection 1138
 Tools for Data Collection 1138
 Tools for Data Analysis 1138
 Limitations of the Study 1138
CHAPTER FOUR 1139
RESULTS
 Introduction 1139
 Graphs 1139
 Interpretation 1140
CHAPTER FIVE 1141
DISCUSSIONS OF FINDINGS
 Discussions of Findings 1141
 Areas for Further Research 1141
CHAPTER SIX 1142
CONCLUSION
REFERENCES 1143

IJISRT23MAY166 www.ijisrt.com 1131

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER ONE
INTRODUCTION
Social networking sites are the most efficient way to meet new people. The assortment in content on social media has made
people became creative and open minded. Social media has thus developed into really powerful medium to share ideas and
opinions. However, due to the rapid growth and popularity of social networking sites, many users have discovered an illegal and
immoral way to use them. The most commonly encountered and most dangerous misuses of online social media are the expression
of hate and harassment.

Hate Speech in relation to social media, is a kind of writing that disparages and is likely to cause harm or danger to the
victim. It is a bias-motivated, hostile, malicious speech aimed at a person or a group of people because of some of their actual or
perceived innate characteristics [12]. It is a kind of speech that demonstrates a clear intention to be hurtful, to incite harm, or to
promote hatred. The environment of social media provides a particularly fertile ground for creation, sharing and exchange of hate
messages against a perceived enemy group [5].

However, identifying and removing hate speech content has proved to be labor-intensive and time consuming. Owing to
these worries and prevalent hate speech content on the internet, there is a strong motivation to implement an automated hate
speech detection system. The automatic detection of hate speech has proved to be a challenging task because of disagreements on
different hate speech definitions as perceived by many. Detection of hate speech and offensive language has been considered as an
emerging application in numerous research problems associated with the domain of Natural Language Processing [13]

 Background of the study

The abuse on social media has shown to be more and more of significant in the last decade hence, the process of noticing or
eliminating such content manually from the web is a tedious task. So, there is a need of developing an automated model that is
able to notice such toxic content on the web.

Regardless of the extensive amount of work that researchers have so far done, it remains problematic to make comparisons
on the performance of these approaches to categorize hate speech content that is flooded on the social media. To my knowledge
based on literature that I have read so far, the prevailing studies lack the comparative analysis of dissimilar feature engineering
techniques and ML algorithms.

 Statement of the problem

Identifying and removing hate speech content using manual process has proved to be labor-intensive and also time
consuming. Manual annotation and removal of the hate speech isn't possible because of the huge amount of data processed every
second. For example, as of 2020, there are more than 6000 tweets sent every second [17].

 Objectives of the study

To manage the enormous volumes of data existing in this virtual sphere, there is an urgent requirement for intelligent
systems that are capable of automatically able to flag and classify the content using numerous machine learning models and
feature engineering techniques. In this paper, I propose an approach to devise machine learning models which will be combined
together with feature engineering techniques and later evaluate their performance based on a publicly available dataset having
three distinct classes for classification. Machine learning has lately been used in wide variety of fields like intelligent healthcare,
smart homes, cybersecurity and many more [18].

 Main Objective
This paper will discuss the ways in which machine learning and feature engineering techniques are used to control hate
speech and abusive language on social media with a given dataset.

 Specific Objective
My study is quite significant as it donates to resolving the problem at hand, by relating three feature engineering and eight
Machine Learning classifiers on standard hate speech datasets having three distinct classifiers. The study holds applied reputation
and therefore, serves as a base line for new researchers in the domain of automatic hate speech detection on social media platform.

 Research Question
Which combination of three feature engineering and eight Machine Learning classifiers out performs on a standard hate
speech dataset?

 Significance of the Study

My study holds a practical implication and can be used as a baseline study in the area of detecting automatic hate speech
messages.

IJISRT23MAY166 www.ijisrt.com 1132

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Scope of the Study
The study is focusing on three branches of Artificial Intelligence (AI), they include; (a) Machine Learning, (b) Deep
Learning and (c) Natural Language Processing. Below, I have explained the basics of each topic and how they relate to my
research.

 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)
 Machine learning “(ML) 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).
 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.

 Definition of Unfamiliar Terms.

Table 1 Text Cataloging

S. No. Concept Abbreviation Description
1 Feature Extraction FE It is mapping from text data to real valued vectors
2 Bigram - It's a feature engineering technique which represents two
adjacent words in a single numeric feature while creating
master feature vectors for words.
3 Term Frequency - TFIDF It's a feature representation technique that represents “word
Inverse Document importance” is to a document in the document set. It works in
Frequency a combination of the frequency of word appearance in a
document with no. of documents containing that word.
4 Word2vec It is a technique used to learn vector representation of words,
which can further be used to train machine learning models
5 Doc2vec It is an unsupervised technique to learn document
representations in fixed-length vectors. It is the same as
word2vec, but the only difference is that it is unique among all
documents.
6 Machine Learning ML Classifiers These are applied to numeric features vector to build the
Classifiers predictive model which can be used for prediction class labels.
7 Naïve Bayes NB It's a probabilistic based classification algorithm, which uses
the “Bayes theorem” to predict the class. It works on
conditional independence among features.
8 Random Forest RF It's a type of ensemble classifier consisting of many decision
trees. It classifies an instance based on voting decision of each
decision trees class predictions.
9 Support Vector Machines SVM It's a supervised classification algorithm which constructs an
optimal hyperplane by learning from training data which
separates the categories while classifying new data.
10 K Nearest Neighbor KNN It's a simple text classification algorithm, which categorize the
new data using some similarity measure by comparing it with
all available data.
11 Decision Tree DT It is a supervised algorithm. It generates the classification
rules in the tree-shaped form, where each internal node
denotes attribute conditions, each branch denotes conditions
for outcome and leaf node represents the class label.
12 Adaptive Boosting AdaBoost It is one of the best-boosting algorithms, which strengthens
the weak learning algorithms.
13 Multilayer Perceptron MLP It is a feedforward artificial neural network. It produces a set
of outputs using a set of inputs
14 Logistic Regression LR It is a predictive analysis. It uses a sigmoid function to explain
the relationship between one independent variable and one or
more independent variables

IJISRT23MAY166 www.ijisrt.com 1133

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER TWO
LITERATURE REVIEW
 Introduction
Research has proved that a lot of study has been done from across the world on hate speech and abusive language
recognition written in diverse languages such as English, Dutch, French, Hindi, Greek. Several studies have so far been concluded
in the field of the detection of hate speech and abusive language using machine learning models, deep learning architectures,
language models, etc. To one side, while recognizing the variety in models and architectures, different works have different data
which are annotated for different aspects or labels. Similarly, the dataset might be in different languages. Each language has their
own lexical, morphological, and syntactic structures.

 Main Literature Review

Gaydhani et al., 2018 [6] They proposed a solution to the detection of hate speech and offensive language on Twitter
through machine learning using n-gram features weighted with TFIDF values. They performed comparative analysis of Logistic
Regression, Naive Bayes and Support Vector Machines on various sets of feature values and model hyper parameters. The results
showed that Logistic Regression performs better with the optimal ngram range 1 to 3 for the L2 normalization of TFIDF. Upon
evaluating the model on test data, we achieved 95.6% accuracy. It was seen that 4.8% of the offensive tweets were misclassified
as hateful. This problem can be solved by obtaining more examples of offensive language which does not contain hateful words.
The results can be further improved by increasing the recall for the offensive class and precision for the hateful class. Also, it was
seen that the model does not account for negative words present in a sentence. Improvements can be done in this area by
incorporating linguistic features.

Parihar et al., 2021 [18] Hate speech detection is a very difficult task and continues to be a societal problem. There is a very
fine line between what is a hate speech and what is not. For example, a satire might also be considered as a possible threat but it is
not actually a hate speech. The annotation and collection of data for building a model for hate speech detection is thus a very
troublesome task. As discussed, this problem can be solved by narrowing down the criteria for annotations. Similarly, there is a
need to focus research on code-mixed languages and regional languages as well. Language models and deep learning models have
shown promising results in hate speech classifications. For tackling with unbalanced data, the up sampling or down sampling
techniques based on language models should be researched upon. The challenges discussed above must be tackled with more
research in the domain so that the internet becomes more inclusive, welcoming and free from hate.

Mahibha et al, [13] They concluded that, from the output obtained from the different models it could be inferred that deep
learning models outperform the machine learning models considering the offensive language classification problem for the data
set provided by HASOC@FIRE-2021 for Task 1 associated with code mixed Tamil. Among the deep learning model transformer-
based models has done the more accurate predictions compared to recurrent models, hence more scope for transformer-based
models could be identified for research based on Dravidian languages and in specific Hate and Offensive language-based
researches.

Zeerak Waseem et al. [20] classify the hate speech on twitter. In their research, they employed character Ngrams feature
engineering techniques to generate the numeric vectors. The authors fed the generated numeric vector to the LR classifier and
obtained overall 73% F-score. While, Chikashi Nobata et al. [6] used the ML -based approach to detect the abusive language in
online user content. In their research authors employed character Ngrams feature representation technique to represent the
features. The authors fed the features to the SVM classifier. The results showed that the classifier obtained overall 77% F-score.
Shervin Malmasi et al [14] used an ML -based approach to classify hate speech in social media. In their research, the authors
employed 4grams with character grams feature engineering techniques to generate numeric features. The authors fed the generated
numeric features to the SVM classifier. The authors reported maximum of 78% accuracy.

Al-Hassan and Al-Dossari, 2019 [1] Arab regions and worldwide are now more aware of the problem of spreading hate
through the social networks. Many countries are working hard in regulating and countering such speech. This attention raised the
need for automating the detection of hate speech. In this paper we analyzed the concept of hate speech and specifically “cyber
hate” which is conducted in the means of social media and the internet sphere. Moreover, they differentiated between the different
anti-social behaviors which include (Cyberbullying, Abusive and offensive language, Radicalization and hate speech). After that
they presented a comprehensive study on how text mining can be used in social networks. we investigated some challenges which
can be a guide for the implementation of Arabic hate speech detection model. In addition, these recommendations will help in
drawing a road map and a blueprint for the future model. The future work will include incorporating the latest deep learning
architectures to build a model that is capable to detect and classify Arabic hate speech in twitter into distinct classes. A data set
will be collected from twitter, and for intensifying the training of our neural network they will including data from additional
platform “e.g. Facebook” as it is the most used platform in the Arab region.

IJISRT23MAY166 www.ijisrt.com 1134

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
Mesa-Jimenez et al., 2022 [15]. In their findings they present a two-stage text classification study in the field of BMS. The
results of the first stage show that XGBoost performs better than the other four, but the others make good candidates for this stage
too, except maybe for multinomial Naive Bayes, which shows slightly worse results. The outperformer in the second stage
classification, the multi label problem, is logistic regression. The top performers that follows are followed by XGBoost algorithm
and, again, the Naive Bayes method performs the worst of the five. The accuracy per tag type shows that certain algorithms may
be better in predicting certain tags than others. In the current paper, we have considered XGBoost and logistic regression to design
the system, but the aim for further work will be a combination of methods for the second stage, using each method for doing only
the classifications they are the best at, to improve the general accuracy of the whole implementation. Sub-dividing the problem
into several problems improves its accuracy for the whole system as expected. In terms of the model's deployment, the assessment
of errors is very important. The main problem for this system's implementation is to locate false positive elements. The false
positives are the incorrectly tagged elements that passed to the building analytics software. These elements may be difficult to
detect, especially for buildings with a big number of points. Increasing the confidence boundary to a higher level may help to
solve this problem and reduce false positives to a minimum. This also may reduce the number of true positives, increasing the
amount of manual work. The findings of this work open a new field of application for text classification methodologies, aiming to
a scientific audience, which may explore the methodologies of this paper further to generalize this field of application for text
processing and categorization, or to industrial professionals who seek to implement this system to reduce operational tagging
times from several days to a couple of minutes. Our research provides a novel multi-stage machine learning solution for the real-
world BMS problem, which can be applied in several systems, or even re-trained with new standards that could appear in the
future.

IJISRT23MAY166 www.ijisrt.com 1135

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER THREE
RESEARCH DESIGN
 Introduction
This section represents a general methodology for building hate speech detection model. The process starts with dataset collection
and goes through the process of annotation or labelling of the data, extraction of features, use of learning algorithms and
evaluation of performance.

 Research Design
The section explains the Methodology that has been adopted in order to categorize tweets into three different distinct classes
namely, “hate speech, offensive but not hate speech, and neither hate speech nor offensive speech”. Fig. 1 below illustrates a
comprehensive research methodology that has been implemented in this research. As presented by the figure below, the
methodology employed has six key steps that are going to be used before the results can be concluded. The steps include; (1) data
collection, (2) data preprocessing, (3) feature engineering, (4) data splitting, (5) classification model construction, (6) and
classification model evaluation. Each step is explained in detail below.

Fig 1 Research Methodology, step by step process

 Data Collection
This study, for the purpose collecting research data, I have used a publicly available open source CrowdFlower dataset.
CrowdFlower provided this dataset as an open source, they compiled and labelled datasets making it very user friendly. The
tweets are labeled into three distinct classes, namely, hate speech, not offensive, and offensive but not hate speech. This dataset
has 14509 number of tweets. Of these, 16% of tweets belong to class hate speech. In addition, 50% of tweets belong to not
offensive class and the remaining 33% tweets are offensive but not hate speech class. The details of this distribution are also
shown in

Fig 2 Crowd Flower Dataset Classification

IJISRT23MAY166 www.ijisrt.com 1136

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Text Preprocessing
Different preprocessing-techniques have been used in order to sieve noisy and non-informative features from the tweets. The
preprocessing technique involves transforming tweets into lower case. The technique correspondingly, is applied to remove all the
URLs, usernames, white spaces, hashtags, punctuations and stop-words using pattern matching techniques from the collected
tweets. Subsequent to preprocessing the tweets also undergo Tokenization and stemming. The tokenization, is the process that is
applied to convert each single tweet into tokens or words, then the porter stemmer converts words to their root forms, such as
offended to offend using porter stemmer.

Fig 3 Data Preprocessing Cycle

 Feature Engineering
Basically, ML algorithms do not recognize the classification rules from the raw text. Hence the need for numerical features
to understand classification rules. This is why feature engineering is highly considered as one among the top key steps in text
classification. At this level it where key features are extracted from raw text and later representing the extracted features in
numerical form. This study has therefore, used three unalike feature engineering techniques. They include; n-gram with TFIDF,
Word2vec and Doc2vec.

 Data Splitting
The table below illustrates the class-wise distribution and also results after splitting of the overall dataset. The table is
showing the number of instances that has been used in Training set and also the number of instances that has been used in Test set.
In the study to split the preprocessed data we have used 80-20 ratio, basically what it means is that 80% of the instances has been
used for Training Data while 20% has been dedicated for Test Data. The whole idea is to ensure that classification models are
trained to learn classification rules.

Table 2 Details of Data Split

S no. Category Total Insurances Training Numbers Testing Numbers
0 Hateful Speech 2397 1908 489
1 Not offensive Speech 7275 5814 1459
2 Offensive but not Hateful Speech 4837 3884 954
Total 14509 1607 2902

 Machine Learning Models

It is significantly commended to apply several unlike classifiers on a master feature vector to detect which one reaches to the
better outcomes. This is because research has proved that there is no any single classifier which best performs on all kinds of
dataset. This is the reason why eight different classifiers have been used for the purpose of this study. These include; NB [21],
SVM [9], KNN [20], DT [2], RF , AdaBoost, MLP and LR [10].

 Classifier Evaluation
At this stage, it is where classes of unlabeled text are predicted by the constructed classifier. The Test Set process follows
that the text is labeled into three distinct classes, namely; (0) hate speech, (1) offensive but not hate speech, (3) neither hate speech
nor offensive speech. The matrixes of True Negative (TN), False Positives (FP), False Negatives (FN) and True Positives (TP) are
calculated in order to evaluate classifier performance..

IJISRT23MAY166 www.ijisrt.com 1137

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
 Population of the Study
The study is targeting social media users from “Crowd Flower” an open source dataset with a population of 14509 records.

 Sampling Procedure
For the purpose of this study a Probability Sampling technique has been adopted in which samples from a larger population
based on the theory of probability has equal chances. This sampling method considers every member of the population and forms
samples based on a fixed process.

 Sample Size
The CrowdFlower dataset has a population of 14,509 number of tweets. This is an open source data which has been provided
freely by CrowdFlower for the purposes of research and other studies.

 Sampling Area
Social media users via world wide web drawn from “Crowd Flower” dataset

 Sources of Data Collection

The dataset used in this study has been provided by “Crowd Flower” as an open source.

 Methods of Data Collection

“Crowd Flower” is an open source dataset which is publicly available to any interested person who has interest to use the
data for research purposes.

 Tools for Data Collection

Since the dataset has been obtained from open source “Crowd Flower” it is apparently difficult to establish the tools that
were used when collecting the data.

 Tools for Data Analysis

In this study Feature Engineering (FE) techniques together with 8 Machine Learning (ML) algorithms has been used to
analyze data for research findings.

 Limitations of the Study

My study has found that most the limitations relating to the research are very much centered of feature extraction while
classifying labels. Labels has a fixed maximum length which is 11words, and in many cases the labels may contain text full of
errors, therefore it has proved to be extremely difficult and very challenging while extracting information in order to perform text
classification into the 3 distinct classes. The other limitations is basically on inconsistencies that are found in the training datasets
after performing data splitting process, hence manually tagged data are used in the system for training the models. The
interpretation is that some engineers while running the models may use slightly different tags sometimes or simply that the
information contained within the label itself is just incomplete and only compensated by personal experience, which limits the
system results.

IJISRT23MAY166 www.ijisrt.com 1138

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER FOUR
RESULTS
 Introduction
In this chapter I will talk about experimental settings that have been used in this study, as mentioned earlier I have used three
types of feature engineering namely n-gram (bigram) with TFIDF, Word2vec and Doc2vec. Therefore, a total of three feature
engineering representations have been computed, and also, eight ML algorithms were used to create three master feature vectors.
As a result, overall 24 analyses (3 master feature vectors x 8 ML algorithms) were evaluated in order to check the effectiveness of
classification models.

 Graphs

Graph A. Precision Analysis

Graph B. Recall Analysis

Graph C. F-Measure Analysis

IJISRT23MAY166 www.ijisrt.com 1139

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165

Graph D. Accuracy Analysis

 Interpretation
This segment will explain the total results of 24 studies that have been conducted in the in this research. Graph A to graph D
demonstrates (A) the precision, (B) recall, (C) F-measure and (D) accuracy of all 24 studies, respectively. The performance and
classification techniques for each of the different feature representation are displayed graphically. Machine Learning (ML)
algorithms with MLP and also KNN registered the lowest results. i.e. Precision analysis recorded 0.56, Recall analysis recorded
0.58, F-measure analysis recorded 0.48, while Accuracy analysis recorded 58%. Machine Learning (ML) algorithms with SVM
while using TFIDF feature representation with bigram registered the highest results. Precision analysis recorded 0.78, Recall
analysis recorded 0.80, F-measure analysis recorded 0.78, while Accuracy analysis recorded 80%. Feature engineering
representation; Best performance has been registered with Bigram feature as compared to Word2vec and Doc2vec. However
further analysis has revealed a marginal difference recorded from results obtained in bigram and Doc2vec. SVM classifier
registered best performance in the text classification models. SVM results out classed all the eight classifiers. Never the less
AdaBoost and RF classifiers results were lesser than SVM results and were better than LR, DT, NB, KNN, and MLP results.

IJISRT23MAY166 www.ijisrt.com 1140

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER FIVE
DISCUSSIONS OF FINDINGS
As a background to my study I have weighed three different feature engineering techniques over eight Machine Learning
(ML) classifiers. Hence at the end obtaining results from an output of twenty-four analysis namely; (a) the precision analysis, (b)
recall analysis, (c) F-measure analysis (d) accuracy analysis. The findings have revealed SVM algorithm as the best model while,
working with combination of bigram feature engineering with TFIDF FE techniques.

 Feature Engineering
This study has three distinct feature extraction techniques which have been deployed and evaluated in their performance
over a standard dataset. The three feature techniques include; Bigram with TFIDF, word2vec and doc2vec. The findings have
revealed Bigram with TFIDF as the best performing model while Word2vec and Doc2vec has inversely performed lower.
Bigram with TFIDF feature technique maintains a sequence of words unlike Word2vec and Doc2vec, this is probably the reason
why Bigram with TFIDF performed better than the rest. Numerous research studies have showed that the TFIDF representation
technique is better than the binary and term frequency representation Mujtaba et al., 2018 [16]. The likely cause for the lower
performance of Word2vec is because it is unable to handle OOV (out of vocabulary) words specially in the domain of Twitter
data.

 Machine Learning
Previous research has proved that “no single Machine Learning (ML) algorithm has performed better results on all kinds of
dataset. It is against this background that several different ML algorithms have to be deployed in order to determine the best
performer on a given dataset. SVM uses threshold functions to perform data separation and not the number of feature based on
margin, this study has thus revealed SVM and AdaBoost as best classifiers based on this reason. This shows that SVM is
independent upon the presence of the number of features in the data Hornik et al., 2013 [8]. The Kernel functions in SVM gives it
ability to perform much better on non-linear data apart from working with linear data. The adaptive algorithms in AdaBoost
enables the model to learn the classification rules, with much attention focusing on decreasing training error. This is the reason
why AdaBoost has better performance as compared with the rest of the ML algorithms. The study has also revealed that SVM and
AdaBoost classifiers has better results and on the second tier there is RF and LR who also have performed higher than results of
NB, DT, KNN, and MLP which are placed on third tier. The absence of informative features in RF which result to incorrect
predictions could be the reason for its low performance. It is possible that the performance of LR might be lower because its
decision surface is linear in nature and cannot handle nonlinear data adequately Eftekhar et al., 2005 [5]. The reason behind the
poor performance of the MLP classifier is due to not having enough training data that's why it is considered as complex “black
box” Singh and Shahid Husain, 2014 [19]. The KNN had the worst performance due to laziness of the learning algorithm and it
does not work adequately for noisy data Bhatia and Author, 2010 [3]. Therefore, according to this study the KNN has proved to be
not suitable for detecting hate speech tweets.

 Areas for Further Research

My work has two distinct boundaries. First, the proposed ML model is inefficient in terms of real-time predictions accuracy
for the data. Finally, it only classifies the hate speech message in three different classes and is not capable enough to identify the
severity of the message. Henceforth, in the future, the objective is to improve the proposed ML model which can be used to
predict the severity of the hate speech message as well.

IJISRT23MAY166 www.ijisrt.com 1141

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
CHAPTER SIX
CONCLUSION
The study has concluded that Bigram combined with TFIDF performed much better than Word2vec and Doc2vec feature
engineering techniques. Furthermore, SVM and RF algorithms in Machine Learning (ML) algorithms have proved better results
when compared with LR, NB, KNN, DT, AdaBoost, and MLP. The lowest performance was observed in KNN. The outcomes
from this research study hold practical importance because this will be used as a baseline study to compare upcoming researches
within different automatic text classification methods for automatic hate speech detection. Furthermore, this study also holds a
scientific value because this study presents experimental results in form of more than one scientific measure used for automatic
text classification.

IJISRT23MAY166 www.ijisrt.com 1142

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
REFERENCES
[1]. Areej Al-Hassan and Hmood Al-Dossari. 2019. DETECTION OF HATE SPEECH IN SOCIAL NETWORKS: A
SURVEY ON MULTILINGUAL CORPUS. In Computer Science & Information Technology (CS & IT), AIRCC
Publishing Corporation, 83–100. DOI:https://doi.org/10.5121/csit.2019.90208
[2]. Asma Ben Abacha, Md. Faisal Mahbub Chowdhury, Aikaterini Karanasiou, Yassine Mrabet, Alberto Lavelli, and Pierre
Zweigenbaum. 2015. Text mining for pharmacovigilance: Using machine learning for drug name recognition and drug–
drug interaction extraction and classification. Journal of Biomedical Informatics 58, (December 2015), 122–132.
DOI:https://doi.org/10.1016/j.jbi.2015.09.015
[3]. Nitin Bhatia and Corres Author. 2010. Survey of Nearest Neighbor Techniques. 8, 2 (2010).
[4]. Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. 2012. Detecting Offensive Language in Social Media to Protect
Adolescent Online Safety. In 2012 International Conference on Privacy, Security, Risk and Trust and 2012 International
Confernece on Social Computing, IEEE, Amsterdam, Netherlands, 71–80. DOI:https://doi.org/10.1109/SocialCom-
PASSAT.2012.55
[5]. Behzad Eftekhar, Kazem Mohammad, Hassan Eftekhar Ardebili, Mohammad Ghodsi, and Ebrahim Ketabchi. 2005.
Comparison of artificial neural network and logistic regression models for prediction of mortality in head trauma based on
initial clinical data. BMC Med Inform Decis Mak 5, 1 (December 2005), 3. DOI:https://doi.org/10.1186/1472-6947-5-3
[6]. Aditya Gaydhani, Vikrant Doma, Shrikant Kendre, and Laxmi Bhagwat. 2018. Detecting Hate Speech and Offensive
Language on Twitter using Machine Learning: An N-gram and TFIDF based Approach. Retrieved April 10, 2023 from
http://arxiv.org/abs/1809.08651
[7]. Njagi Dennis Gitari, Zuping Zhang, Hanyurwimfura Damien, and Jun Long. 2015. A Lexicon-based Approach for Hate
Speech Detection. IJMUE 10, 4 (April 2015), 215–230. DOI:https://doi.org/10.14257/ijmue.2015.10.4.21
[8]. Kurt Hornik, Patrick Mair, Johannes Rauch, Wilhelm Geiger, Christian Buchta, and Ingo Feinerer. 2013. The textcat
Package for n -Gram Based Text Categorization in R. J. Stat. Soft. 52, 6 (2013). DOI:https://doi.org/10.18637/jss.v052.i06
[9]. Thorsten Joachims. 1998. Text categorization with Support Vector Machines: Learning with many relevant features. In
Machine Learning: ECML-98, Claire Nédellec and Céline Rouveirol (eds.). Springer Berlin Heidelberg, Berlin, Heidelberg,
137–142. DOI:https://doi.org/10.1007/BFb0026683
[10]. Hui Li and Zeming Li. 2022. Text Classification Based on Machine Learning and Natural Language Processing
Algorithms. Wireless Communications and Mobile Computing 2022, (July 2022), 1–12. DOI:https://doi.org/10.1155/ 2022/
3915491
[11]. Shuhua Liu and Thomas Forss. 2014. Combining N-gram based Similarity Analysis with Sentiment Analysis in Web
Content Classification: In Proceedings of the International Conference on Knowledge Discovery and Information
Retrieval, SCITEPRESS - Science and and Technology Publications, Rome, Italy, 530–537. DOI:https://doi.org/10.5220/
0005170305300537
[12]. Sean MacAvaney, Hao-Ren Yao, Eugene Yang, Katina Russell, Nazli Goharian, and Ophir Frieder. 2019. Hate speech
detection: Challenges and solutions. PLoS ONE 14, 8 (August 2019), e0221152. DOI:https://doi.org/10.1371/ journal.pone.
0221152
[13]. C Jerin Mahibha, Sampath Kayalvizhi, Durairaj Thenmozhi, and Sundar Arunima. Offensive Language Identification using
Machine Learning and Deep Learning Techniques.
[14]. Shervin Malmasi and Marcos Zampieri. 2017. Detecting Hate Speech in Social Media. Retrieved April 11, 2023 from
http://arxiv.org/abs/1712.06427
[15]. Jose Joaquin Mesa-Jiménez, Lee Stokes, QingPing Yang, and Valerie N. Livina. 2022. MACHINE LEARNING FOR
TEXT CLASSIFICATION IN BUILDING MANAGEMENT SYSTEMS. JOURNAL OF CIVIL ENGINEERING AND
MANAGEMENT 28, 5 (May 2022), 408–421. DOI:https://doi.org/10.3846/jcem.2022.16012
[16]. Ghulam Mujtaba, Liyana Shuib, Ram Gopal Raj, Retnagowri Rajandram, and Khairunisa Shaikh. 2018. Prediction of cause
of death from forensic autopsy reports using text classification techniques: A comparative study. Journal of Forensic and
Legal Medicine 57, (July 2018), 41–50. DOI:https://doi.org/10.1016/j.jflm.2017.07.001
[17]. Keisuke Nakagawa, Nuen Tsang Yang, Machelle Wilson, and Peter Yellowlees. 2022. Twitter Usage Among Physicians
From 2016 to 2020: Algorithm Development and Longitudinal Analysis Study. J Med Internet Res 24, 9 (September 2022),
e37752. DOI:https://doi.org/10.2196/37752
[18]. Anil Singh Parihar, Surendrabikram Thapa, and Sushruti Mishra. 2021. Hate Speech Detection Using Natural Language
Processing: Applications and Challenges. In 2021 5th International Conference on Trends in Electronics and Informatics
(ICOEI), IEEE, Tirunelveli, India, 1302–1308. DOI:https://doi.org/10.1109/ICOEI51242.2021.9452882
[19]. Pravesh Kumar Singh and Mohd Shahid Husain. 2014. Methodological Study Of Opinion Mining And Sentiment Analysis
Techniques. IJSC 5, 1 (February 2014), 11–21. DOI:https://doi.org/10.5121/ijsc.2014.5102
[20]. Zeerak Waseem and Dirk Hovy. 2016. Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection
on Twitter. In Proceedings of the NAACL Student Research Workshop, Association for Computational Linguistics, San
Diego, California, 88–93. DOI:https://doi.org/10.18653/v1/N16-2013
[21]. Baoxun Xu, Xiufeng Guo, Yunming Ye, and Jiefeng Cheng. 2012. An Improved Random Forest Classifier for Text
Categorization. JCP 7, 12 (December 2012), 2913–2920. DOI:https://doi.org/10.4304/jcp.7.12.2913-2920

IJISRT23MAY166 www.ijisrt.com 1143

Volume 8, Issue 5, May 2023 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[22]. View of Modeling the Detection of Textual Cyberbullying. Retrieved April 11, 2023 from https://ojs.aaai.org/index.php/
ICWSM/article/view/14209/14058

IJISRT23MAY166 www.ijisrt.com 1144