Machine Learning Approach to Classify Twitter Hate Speech

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 7 Issue 5, September-October 2023 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD59873 | Volume – 7 | Issue – 5 | Sep-Oct 2023 Page 165
Machine Learning Approach to Classify Twitter Hate Speech
Subrata Saha1, Md. Motinur Rahman2, Md. Mahbub Alam3
1,2
Scientific Officer, Institute of Electronics,
3
Senior Scientific Officer, Institute of Computer Science,
1, 2, 3
Bangladesh Atomic Energy Commission, Dhaka, Bangladesh
ABSTRACT
In this modern age, social media platforms have become
indispensable tools for communication and information sharing.
However, this unprecedented connectivity has also given rise to a
concerning proliferation of hate speech and offensive content. This
research article presents a comprehensive study on the development
and evaluation of machine learning (ML) models for the automatic
detection of hate speech on Twitter. We leverage a diverse dataset
collected from Twitter, encompassing a wide range of hate speech
categories, including hate speech targeting race, gender, religion, and
more. To address the multifaceted nature of hate speech, we employ a
hybrid approach that combines traditional natural language
processing (NLP) techniques with state-of-the-art machine learning
algorithms. Our methodology involves extensive preprocessing of the
text data, including tokenization, stemming, and feature extraction.
We then experiment with various machine learning algorithms,
including Naïve Bayes (NB), K-nearest Neighbor (KNN), Random
Forest (RF), and Support Vector Machines (SVM). The models are
trained and fine-tuned on a labeled dataset and evaluated using robust
metrics to assess their performance.
KEYWORDS: Hate speech classification, cyberbullying, NLP,
Machine learning
How to cite this paper: Subrata Saha |
Md. Motinur Rahman | Md. Mahbub
Alam "Machine Learning Approach to
Classify Twitter Hate Speech" Published
in International
Journal of Trend in
Scientific Research
and Development
(ijtsrd), ISSN:
2456-6470,
Volume-7 | Issue-5,
October 2023,
pp.165-169, URL:
www.ijtsrd.com/papers/ijtsrd59873.pdf
Copyright © 2023 by author (s) and
International Journal of Trend in
Scientific Research and Development
Journal. This is an
Open Access article
distributed under the
terms of the Creative Commons
Attribution License (CC BY 4.0)
(http://creativecommons.org/licenses/by/4.0)
I. INTRODUCTION
Social media platforms, particularly Twitter, have
emerged as powerful tools for communication,
information dissemination, and social interaction in
the digital era. They offer a platform for individuals
and communities to express their thoughts, engage in
discussions, and connect with a global audience.
However, this unprecedented connectivity has also
brought to light a deeply concerning issue - the
widespread dissemination of hate speech and
offensive content [1].
Hate speech on Twitter, characterized by its
derogatory, discriminatory, or threatening nature,
poses serious challenges to the principles of online
civility, inclusivity, and safety. Instances of hate
speech targeting race, gender, religion, ethnicity, and
other personal characteristics have become
distressingly common, leading to real-world
consequences, including cyberbullying, harassment,
and the exacerbation of social tensions. [2]
The exponential growth of social media platforms and
the sheer volume of content posted daily make manual
moderation an impractical and insufficient solution.
The need for automated tools to detect and mitigate
hate speech has never been more critical. Machine
learning, with its ability to analyze vast datasets and
identify patterns, offers a promising avenue for
addressing this complex problem. [3]
This research article presents a comprehensive study
on the development and evaluation of machine
learning models for the automatic detection of hate
speech on Twitter. We recognize the urgency of this
task in fostering safer and more inclusive online
environments, as well as in protecting vulnerable
individuals and groups from the harms of online hate.
Our research employs a multifaceted approach to
address the challenges posed by hate speech on
Twitter. We explore our dataset encompassing various
hate speech categories and apply a hybrid
methodology that combines traditional natural
language processing (NLP) techniques with cutting-
edge machine learning algorithms. The complete
process in shown in Fig 1.
IJTSRD59873

International Journal of Trend in Scientific Research and Development @ www.ijtsrd.com eISSN: 2456-6470
The subsequent sections of this paper provide some
related studies, and a detailed account of our research
methodology, experimentation, results, and
conclusions. By leveraging machine learning and
data-driven insights, our aim is to contribute to the
ongoing efforts to combat the proliferation of hate
speech on social media platforms, ultimately fostering
a more inclusive and respectful online community.
II. Related works
Research in the domain of hate speech detection on
social media platforms, particularly Twitter, has
gained considerable attention in recent years. Several
studies have explored various approaches and
methodologies to tackle this pressing issue. In this
section, we review key related works and their
contributions to the field of hate speech detection,
providing valuable insights and context for our
research.
Davidson et al. conducted a seminal study on hate
speech detection, focusing on the development of
machine learning models to identify offensive
language and hate speech on Twitter. Their work laid
the foundation for subsequent research in this area and
highlighted the challenges of distinguishing hate
speech from offensive language [4].
Waseem et al. investigated the predictive features for
hate speech detection, emphasizing the importance of
not only textual content but also user behavior and
interaction patterns. Their work underscores the
complexity of the hate speech detection task and the
need for a holistic approach. [5]
SemEval-2019 Task 5 challenge aimed to advance
hate speech detection by focusing on multilingual
content targeting specific demographics. Participants
in this challenge developed and evaluated models for
detecting hate speech against immigrants and women,
addressing the need for language diversity and
inclusivity [6].
Fig 1 Block diagram of the complete process
In their study, Davidson et al. presented a
comprehensive exploration of automated hate speech
detection on Twitter. They investigated the challenges
of distinguishing between offensive language and
genuine hate speech and offered valuable insights into
the detection of online hate speech. [7]
Zhang et al. focused on the use of deep learning
techniques, specifically Convolutional Neural
Networks (CNNs) and Gated Recurrent Units (GRUs),
for hate speech detection on Twitter. The study
explored the effectiveness of deep learning in
addressing this challenging task. [8]
Founta et al. presented a large-scale crowdsourcing
effort to characterize abusive behavior on Twitter.
Their work provided a valuable dataset and insights
into different forms of online abuse, including hate
speech. [9]
III. Methodology
The machine learning approaches can construct
classifiers to complete sentiment classification by
extracting feature vectors, which incorporates data
collecting and preprocessing, features extraction,
training the data with the classifier, and analyzing
results.
It is required to separate the dataset into a training
dataset and a test dataset. The test dataset assesses the
classifier's performance while the training sets are
intended to help the classifier learn the text features.
A. Dataset description and pre processing
The dataset utilized in this study includes tweets from
users of the social media platform Twitter about
various subjects directed at people in a variety of
professions, including actors, models, athletes,
musicians, and politicians. Using the Twitter API, web
scraping was used to gather 31596 tweets for this
dataset. Tweets and sentiment are the two variables
(columns) in the dataset. Each Tweet fell into one of
five categories: religion, ethnicity, age, gender, and
not cyberbullying or neutral. There are 6342 tweets
related to religious views, 6334 related to ethnicity,
6328 related to age, 6296 related to gender and finally
6296 tweets belong to not cyberbullying. Later,
religion, age, gender, ethnicity and not cyberbullying
categorical values in the sentiment column were
replaced by numerical values namely 1, 2, 3, 4 and 5
respectively.
Text preprocessing is a fundamental step in sentiment
analysis that involves transforming raw textual data
into a format suitable for analysis. This section
outlines the various text preprocessing techniques
employed in our research to ensure the quality and
consistency of our dataset.
The first step involves removing any non-essential
elements from the text, such as HTML tags, special
characters, and punctuation. This is done to ensure
that the text is in a clean and uniform format. Then, to
maintain consistency and reduce the dimensionality of
the data, all text is converted to lowercase.
Tokenization is the process of breaking down text into
individual words or tokens. We employ word
tokenization to split the text into its constituent words.

Later, in stop words removal section, such as "the,"
"and," "is," are common words that often carry little
sentiment information.
Fig 2 Dataset sample after pre processing
Then tweets with less than 3 words and more than 100
words were eliminated as they can be outliers.
Stemming and lemmatization are techniques used to
reduce words to their base or root form. We
experimented with stemming to evaluate their impact
on sentiment analysis accuracy. Emoticons and emoji
are often used to convey sentiments. We preprocess
text to identify and standardize emoticons and emoji,
converting them into textual representations for
analysis. Dataset after filtering is shown in Fig 2.
B. Feature extraction
Feature extraction is a pivotal step in sentiment
analysis, wherein the preprocessed textual data is
converted into a numerical representation that
machine learning models can analyze. This section
outlines the feature extraction techniques employed in
our research to capture relevant information from the
text for sentiment classification.
Researchers employ a number of methods, including
word2vec, TF-IDF, Count Vectorizer, Bag of Words
(BOW), and One Hot Encoding. In our investigation,
the text vectorizer was Count Vectorizer.
C. Classification
In this study, we created a model to recognize twitter
hate speech using four machine learning classifiers,
including Naïve Bayes (NB), k Nearest Neighbors
(KNN), Random Forest (RF), and Support Vector
Machine (SVM). The effectiveness of each classifier
has been evaluated in relation to numerous
performance parameters. In the subsection that
follows, we will discuss various performance
parameters used for the prediction model.
D. Performance metrics
Evaluating the performance of sentiment analysis
models is paramount to gauge their accuracy,
reliability, and suitability for practical applications.
This section presents the performance metrics
employed in our research to assess the effectiveness of
sentiment classification models.
Accuracy:
Accuracy measures the proportion of correctly
classified instances out of the total. It provides an
overall assessment of model performance, shown in
eq 1.
(1)
Precision and Recall:
Precision measures the proportion of true positive
predictions among all positive predictions,
emphasizing the correctness of positive predictions.
Recall, on the other hand, measures the proportion of
true positive predictions among all actual positive
instances, emphasizing the ability to capture all
positive instances as in eq 2 and eq 3.
(2)
(3)
F1-Score:
The F1-score is the harmonic mean of precision and
recall, providing a balanced assessment of model
performance, depicted in eq 4.
(4)
Fig 3 Confusion Matrix
Confusion Matrix:
The confusion matrix provides a detailed breakdown
of model performance, showing the number of true
positives, true negatives, false positives, and false
negatives. Shown in fig 3.
IV. Result analysis
After applying selected classifiers following precision,
recall and f1 score were achieved. Comparison of all
accuracy is shown in Fig 4. From the figure it can be
seen that, 84%, 81%, 94% and 93% accuracy was
achieved by NB, KNN, RF and SVM algorithms
respectively. Best accuracy was demonstrated by RF
algorithm. SVM showed the second best performance
in terms of accuracy. Precision, recall, f1 score for RF
and SVM are shown in Table 1 and Table 2.
For RF algorithm best precision, recall and f1 score
was demonstrated by ethnicity, and second best was
by age. For SVM algorithm precision, recall and f1
score was achieved by ethnicity. The second best
performance was precision, recall and f1 score was
achieved by religion and gender, age and religion and
age.

Confusion matrix of all the algorithms is shown in Fig
5.
Table: 1 Precision, Recall and F1 score RF
algorithm
precision recall f1-score
religion 0.97 0.96 0.96
age 0.98 0.98 0.98
gender 0.95 0.87 0.91
ethnicity 0.99 0.99 0.99
Not bullying 0.81 0.90 0.85
Table: 2 Precision, Recall and F1 score SVM
algorithm
precision recall f1-score
religion 0.97 0.95 0.96
age 0.95 0.97 0.96
gender 0.97 0.87 0.92
ethnicity 0.98 0.98 0.98
Not bullying 0.78 0.88 0.83
Fig 4 Comparison of Accuracies For All
Algorithms
V. Conclusions
Through a rigorous analysis of diverse sentiment
analysis techniques, we delved into the nuances of
feature extraction, model selection, and performance
evaluation. Our findings underscore the pivotal roleof
preprocessing steps in enhancing the quality oftextual
data and the significance of appropriate feature
representations in sentiment analysis. We
demonstrated the versatility of machine learning
algorithms and the effectiveness of deep learning
models, each offering a unique lens through which to
view and interpret sentiment. Furthermore, we
investigate the impact of different feature engineering
techniques, model architectures, and hyperparameter
tuning on the hate speech detection task. Our results
demonstrate that our hybrid approach outperforms
baseline models, achieving high accuracy, precision,
recall, and F1-score in identifying hate speech tweets.
Additionally, we explore the challenges and ethical
considerations in hate speech detection, emphasizing
the importance of addressing bias and fairness
concerns in automated moderation systems. This
research contributes to the ongoing efforts to mitigate
the spread of hate speech on social media platforms
and promotes safer and more inclusive online
environments. The developed models and insights
from this study offer valuable tools for social media
companies and policymakers to combat online hate
speech effectively.
Some of the future prospects of our study includes,
our technique of hate speech detection can include not
only text but also images and videos shared on
Twitter. Multimodal analysis can provide a more
comprehensive understanding of hate speech and its
propagation. Also, develop models that consider the
context in which tweets are posted.
Confusion matrix for all algorithms
Confusion matrix for NB Confusion matrix for KNN

Confusion matrix for RF Confusion matrix for SVM
Fig 5 Confusion Matrix For All Algorithms
References:
[1] K. Dashtipour et al., “Multilingual sentiment
analysis: state of the art and independent
comparison of techniques,” Cogn Comput, vol.
8, no. 4, pp. 757–771, Aug. 2016, doi:
10.1007/s12559-016-9415-7.
[2] Y. Qi and Z. Shabrina, “Sentiment analysis
using Twitter data: a comparative application of
lexicon- and machine-learning-based
approach,” Soc. Netw. Anal. Min., vol. 13, no.
1, pp. 1–14, Dec. 2023, doi: 10.1007/S13278-
023-01030-X/TABLES/3.
[3] K. Arun and A. Srinagesh, “Multi-lingual
Twitter sentiment analysis using machine
learning,” Int J Electr Comput Eng, vol. 10, no.
6, pp. 5992–6000, Dec. 2020, doi:
10.11591/ijece.v10i6.pp5992-6000.
[4] T. Davidson, D. Warmsley, M. Macy, and I.
Weber, “Automated Hate Speech Detection and
the Problem of Offensive Language,” Proc. Int.
AAAI Conf. Web Soc. Media, vol. 11, no. 1, pp.
512–515, May 2017, doi:
10.1609/ICWSM.V11I1.14955.
[5] Z. Waseem and D. Hovy, “Hateful Symbols or
Hateful People? Predictive Features for Hate
Speech Detection on Twitter,” HLT-NAACL
2016 - 2016 Conf. North Am. Chapter Assoc.
Comput. Linguist. Hum. Lang. Technol. Proc.
Student Res. Work., pp. 88–93, 2016, doi:
10.18653/V1/N16-2013.
[6] V. Basile et al., “SemEval-2019 Task 5:
Multilingual Detection of Hate Speech Against
Immigrants and Women in Twitter,” NAACL
HLT 2019 - Int. Work. Semant. Eval. SemEval
2019, Proc. 13th Work., pp. 54–63, 2019, doi:
10.18653/V1/S19-2007.
[7] T. Davidson, D. Warmsley, M. Macy, and I.
Weber, “Automated hate speech detection and
the problem of offensive language,” Proc. 11th
Int. Conf. Web Soc. Media, ICWSM 2017, pp.
512–515, 2017, doi:
10.1609/ICWSM.V11I1.14955.
[8] Z. Zhang, D. Robinson, and J. Tepper,
“Detecting Hate Speech on Twitter Using a
Convolution-GRU Based Deep Neural
Network,” Lect. Notes Comput. Sci. (including
Subser. Lect. Notes Artif. Intell. Lect. Notes
Bioinformatics), vol. 10843 LNCS, pp. 745–
760, 2018, doi: 10.1007/978-3-319-93417-
4_48/TABLES/4.
[9] A. M. Founta et al., “Large Scale
Crowdsourcing and Characterization of Twitter
Abusive Behavior,” 12th Int. AAAI Conf. Web
Soc. Media, ICWSM 2018, pp. 491–500, Feb.
2018, doi: 10.1609/icwsm.v12i1.14991.

Machine Learning Approach to Classify Twitter Hate Speech

More Related Content

Similar to Machine Learning Approach to Classify Twitter Hate Speech (20)

More from ijtsrd (20)

Recently uploaded (20)

Machine Learning Approach to Classify Twitter Hate Speech