Proceedings of the Seventh International Conference on Communication and Electronics Systems (ICCES 2022)

IEEE Xplore Part Number: CFP22AWO-ART; ISBN: 978-1-6654-9634-6

Hate Speech Detection using Machine Learning

P. Preethy Jemima Bishop Raj Majumder Bibek Kumar Ghosh Farazul Hoda
2022 7th International Conference on Communication and Electronics Systems (ICCES) | 978-1-6654-9634-6/22/$31.00 ©2022 IEEE | DOI: 10.1109/ICCES54183.2022.9835776

Assistant Professor, Department of Student, Department of Computer Student, Department of Computer Student, Department of Computer
Computer Science and Engineering Science and Engineering Science and Engineering Science and Engineering
SRM Institute of Science and SRM Institute of Science and SRM Institute of Science and SRM Institute of Science and
Technology Technology Technology Technology
Ramapuram Campus, Chennai, Ramapuram Campus, Chennai, Ramapuram Campus, Chennai, Ramapuram Campus, Chennai,
Tamil Nadu, India Tamil Nadu, India Tamil Nadu, India Tamil Nadu, India
[email protected] [email protected] [email protected] [email protected]

Abstract – A lot of methods have already been created for the

automation of hate speech detection online. There are two II. ORGANISATION
elements to this process: identifying the qualities that these terms The following is a summary of the content of this document.
utilize to target a certain group and classifying textual material as Section II provides an overview of the topic at issue. System
hate or non-hate speech. Due to time restraints, research efforts
are initiated on the latter issue in this project. For this reason,
architecture explanations and diagrams are given in Section III.
detecting hate speech is a more challenging endeavor, as our Section IV describes how the system has been implemented,
research of the language used in typical datasets reveals that hate the modules and a few diagrams to explain better the working
speech lacks distinctive, discriminatory characteristics. Deep of the program. And finally Section V concludes the study as
neural network topologies are very useful for capturing the well as mentions future enhancements that could be made.
meaning of hate speech and are thus proposed as feature
extractors. Data from social media sites such as Twitter are used III. PROBLEM STATEMENT
to test the effectiveness of these procedures, and they reveal a 6 "Tweets" are messages that members post and respond to on
percentage point improvement in macro-average F1 or a 9 percent
Twitter, one of the most popular social networking sites.
improvement for content that has been labeled as hateful,
respectively. Individuals can use this to share their ideas and opinions on a
variety of topics [5]. The sentiment analysis of such tweets has
Keywords— Random Forest, NLP, Machine Learning, Text mining, been used by a wide range of entities, including consumers and
Sentiment Analysis, Back Propagation Neural Network, ANNs, marketers, to get insights about products or conduct market
RNN, LSTN. study. The accuracy of the sentiment analysis prediction
models increases with the recent advances in machine learning
I. INTRODUCTION techniques [6]. We want to use a variety of machine learning
People may now communicate and share material more methods to analyze "tweets" in our research. We begin by
easily than ever before, but social media platforms like Twitter determining the tweet's polarity, or whether it is good or
and other community forums are also being used to disseminate negative. A tweet with both positive and negative aspects will
hate speech and organize events based on hatred [1]. Such be classified as either positive or negative depending on which
media's anonymity and mobility make it easier for hate speech mood is more prominent [7]. According to Kaggle's data set,
to propagate, which eventually leads to hate crime in a virtual we are utilizing a crawling dataset that has been classified
place that traditional law enforcement can't reach. Using the positive or negative. Emotional expressions and hashtags are
word "hate speech" refers to any kind of communication that also included in the data. As a result, they will need to be
disparages someone or a group of people because of their processed and then transformed into a format that is standard.
qualities, such as religion, ethnicity, sexual orientation or We'll also need to sift through the content to find important
gender [2]. Recent events, such as Britain's decision to leave information [8]. As a type of "tweet," these properties include
the European Union and the terrorist attacks in Manchester and unigrams and bigrams, which may be used to convey
London, have resulted in an uptick in hate speech directed information. Although individual models did not provide a high
towards immigrant and Muslim populations in the UK. Hate level of accuracy, we selected a small number of the best
speech is on the rise among young people in the European models to create an ensemble [9]. To boost the accuracy of
Economic Area (EEA) area, according to EU polls and studies prediction, we employ ensembling, a form of meta learning
[3]. According to a survey, 80 percent of respondents had algorithm technique that combines several classifiers. And at
experienced hate speech online, and 40 percent have been the end, we report our findings and experimental results [10].
intimidated or attacked as a result of their opinions. Since
Trump's election, there has been an increase in hate speech and
IV. EXISTING SYSTEM
crime in the United States. An increasing number of worldwide
efforts have been developed in an effort to better understand the A document classification task is all that exists in terms of
problem and devise effective solutions [4]. existing methodologies. SVM [12], Naive Bayes [13], and
Logistic Regression [14] (traditional approaches) rely on

978-1-6654-9634-6/22/$31.00 ©2022 IEEE 1274

d licensed use limited to: Alvas Institute of Engineering and Technology Department of Library and Information Centre. Downloaded on February 17,2025 at 10:19:44 UTC from IEEE Xplore. Restrictio
 Proceedings of the Seventh International Conference on Communication and Electronics Systems (ICCES 2022)
IEEE Xplore Part Number: CFP22AWO-ART; ISBN: 978-1-6654-9634-6

manual feature engineering; the more contemporary deep In the form of comma-separated values files, the data includes
learning paradigm, on the other hand, uses deep learning tweets and the attitudes they express. Tweet id is a unique
methods or neural networks on raw data to automatically identifier, and sentiment is either 1 (positive) or 0 (negative).
The tweet is enclosed in "." in a CSV training dataset. CSV
generate multi-layers of abstract features from it [15-17]
files of the form tweet id, tweet are used for the test dataset.

A. Disadvantages There is a considerable bias towards non-hate in all datasets

1. Most research on hate speech detection have relied on because hate Tweets account for just 5.8% (DT) to 31.6 % of
Micro-average Precision, Recall, and F1. the total tweets (WZ). Those with "racism" and, as previously
indicated, the extreme scenario in which it contains "both" are
2. The issue here is that micro-averaging might conceal the true even fewer when we look at specific sorts of hate. Aside from
that, there are two possible outcomes. The system's capacity to
performance of minority classes in an unbalanced dataset where
detect 'non-hate' may be overstated when evaluated using a
instances of one class (to be termed the "dominant class") far metric like the micro F1, which considers a system's
outnumber those of other classes (to be called "minority performance throughout the whole dataset, independent of
classes")[18-19]. class. Racist-tagged tweets can still be obscured by tweets
labeled "non-hate," and vice versa, even if a hypothetical
system can reach virtually flawless F1 in the recognition of
V. PROPOSED SYSTEM 'racism' tagged tweets. To begin with, there aren't nearly as
Only 5.8% (DT) to 31.6 percent (WZ) of all tweets are hateful, many hate-related Tweets as there are non-hate-related ones.
which means that all datasets are heavily skewed against Due to the fact that the datasets are being obtained from
Twitter, this scenario may not be as simple to fix as it appears.
hatred. 'Racism,' and the extreme situation that carries 'both,' As a result, substantially more time and effort will be required
become increasingly uncommon when specific sorts of hate are to annotate the non-hate component of training data in order to
examined in detail. Aside from that, there are two possible detect as many occurrences of hateful content as possible.
outcomes. Initially, an evaluation measure, such as the micro
F1 that examines a system's performance on the full dataset and
that is done without consideration to class differences that may
be skewed against the system's capacity to identify non-hate,
will be used to assess the system's performance. In a nutshell, a
theoretical system that achieves virtually flawless F1 in
recognizing 'racism' Tweets in identifying 'non-hate' can
nevertheless hide a low F1 in identifying 'racism' Tweets.
Second, there is a dearth of hate-related data compared to non-
hate-related data. Because the numbers were gleaned from
Twitter and represent the true nature of data imbalance
throughout this sector, it's possible that this is a more difficult
topic to discuss than it first appears. Because of this, we will
very certainly have to devote a significant amount of time and
effort to providing more training data for non-hateful content.

A. Advantages
1. Oversampling or undersampling may not be able to solve
this problem, as we'll illustrate in the next section.
2. There are no distinguishing linguistic traits in hate Tweets
that are not present in non-hate tweets.
One way of measuring how distinct hate and non-hate
Twitter messages are in their vocabulary is to look at how
many words are unique to each class.

VI. SYSTEM ARCHITECTURE

A. Data Description: Fig 1: System Design

978-1-6654-9634-6/22/$31.00 ©2022 IEEE 1275

d licensed use limited to: Alvas Institute of Engineering and Technology Department of Library and Information Centre. Downloaded on February 17,2025 at 10:19:44 UTC from IEEE Xplore. Restrictio
 Proceedings of the Seventh International Conference on Communication and Electronics Systems (ICCES 2022)
IEEE Xplore Part Number: CFP22AWO-ART; ISBN: 978-1-6654-9634-6

VII. SYSTEM IMPLEMENTATION

A. Modules:

A.1 Exploratory Data Analysis:

The target variable was established in this stage after some
descriptive analysis. Other potentially problematic (high
cardinality) factors were also examined, including the target's
number of classes. Visualizing the target variable in a
histogram, an excellent tool for understanding the distribution
of data, was used to aid with parameter tuning. Fig 2: Use Case Diagram

A.2 Data Cleaning:

As a prelude to pre-processing, the high cardinality variables
were removed in this stage.

A.3 Pre-processing & Transformation:

One-hot encoding was used to change the categorical variable
from the complete data set and subsequently to make it easier
to analyze. When dealing with data in a sparse matrix format,
this is a criterion that must be satisfied by specific algorithms.
Automated software for statistics such as R can handle this
phase when creating models. The data is then populated with
zeros for any missing values. Min-max normalization is then
used to the continuous variables to translate their values onto a
scale from 0 to 1 so that the coefficients are not affected by
variables that are on different scales.
Fig 3: Class Diagram
A.4 Data Partition:
Partitioning the pre-processed data into training and testing
sets is the next step.

A.5 Modelling:
In order to build a k-NN classifier model, 10 neighbor classes
and the Euclidean distance between them were used.

A.6 Evaluation:
The classifier is subsequently evaluated using new test data,
and the R squared values for the training and test datasets are
computed.

Fig 4: Sequence Diagram

978-1-6654-9634-6/22/$31.00 ©2022 IEEE 1276

d licensed use limited to: Alvas Institute of Engineering and Technology Department of Library and Information Centre. Downloaded on February 17,2025 at 10:19:44 UTC from IEEE Xplore. Restrictio
 Proceedings of the Seventh International Conference on Communication and Electronics Systems (ICCES 2022)
IEEE Xplore Part Number: CFP22AWO-ART; ISBN: 978-1-6654-9634-6

REFERENCES
[1] David M. Blei, Andrew Ng, and Michael I. Jordan. 2003. Latent Dirichlet
Allocation. Journal of Machine Learning Research, 3:993–1022.
[2] Peter F. Brown, Peter V. deSouza, Robert L. Mercer, Vincent J. Della
Pietra, and Jenifer C. Lai. 1992.Class-based n-gram models of natural
language. Computational Linguistics, 18(4):467–479.
[3] Karnan, B., Kuppusamy, A., Latchoumi, T. P., Banerjee, A., Sinha, A.,
Biswas, A., & Subramanian, A. K. (2022). Multi-response Optimization of
Turning Parameters for Cryogenically Treated and Tempered WC–Co
Inserts. Journal of The Institution of Engineers (India): Series D, 1-12.
[4] Pete Burnap and Matthew L. Williams. 2015. Cyber hate speech on twitter:
An application of machine classification and statistical modeling for policy
and decision making. Policy & Internet, 7(2):223–242.
[5] Dasari, M. S., & Mani, V. (2020). Simulation and analysis of three-phase
parallel inverter using multicarrier PWM control schemes. SN Applied
Sciences, 2(5), 1-10.
[6] Latchoumi, T. P., & Parthiban, L. (2022). Quasi oppositional dragonfly
algorithm for load balancing in a cloud computing environment. Wireless
Personal Communications, 122(3), 2639-2656.
[7] Pete Burnap, Matthew L. Williams, Luke Sloan, Omer Rana, William
Housley, Adam Edwards, Vincent Knight, Rob Procter, and Alex Voss.
2014. Tweeting the terror: modelling the social media reaction to the
woolwich terrorist attack. Social Network Analysis and Mining, 4(1):1–14.
[8] Pavan, V. M., Balamurugan, K., & Latchoumi, T. P. (2021). PLA-Cu
reinforced composite filament: Preparation and flexural property printed at
different machining conditions. Advanced composite materials.
[9] Ying Chen, Yilu Zhou, Sencun Zhu, and Heng Xu. 2012. Detecting
offensive language in social media to protect adolescent online safety. In
Privacy, Security, Risk and Trust (PASSAT), 2012 International
Conference on and 2012 International Conference on Social Computing
(SocialCom), pages 71–80, Amsterdam, Netherlands, September. IEEE.
[10] Ravipati, S., Mani, V., & Yarlagadda, S. R. (2021). Efficient Control of
Sensorless Hybrid Electric Vehicle Using RBFN Controller. Studies in
Informatics and Control, 30(4), 87-97
[11] Banu, J. F., Muneeshwari, P., Raja, K., Suresh, S., Latchoumi, T. P., &
Deepan, S. (2022, January). Ontology-Based Image Retrieval by Utilizing
Model Annotations and Content. In 2022 12th International Conference
on Cloud Computing, Data Science & Engineering (Confluence) (pp.
Fig 5: Activity Diagram 300-305). IEEE.
[12]Nemanja Djuric, Jing Zhou, Robin Morris, Mihajlo Grbovic, Vladan
Radosavljevic, and Narayan Bhamidipati. 2015. Hate speech detection with
VIII. CONCLUSION comment embeddings. In Proceedings of the 24th International Conference
We conducted a poll to see if hate speech identification could on World Wide Web, pages 29–30, New York, NY, USA. ACM.
[13]Njagi Dennis Gitari, Zhang Zuping, Hanyurwimfura Damien, and Jun
be automated. The most common way to describe this Long. 2015. A lexicon-based approach for hate speech detection.
challenge is as a problem of supervised learning. In a regular International Journal of Multimedia and Ubiquitous Engineering,
order, features that are sufficiently generic, such as word bags 10(4):215–230.
or word embeddings, provide good classification performance. [14]Homa Hosseinmardi, Sabrina Arredondo Mattson, Rahat Ibn Rafiq,
Richard Han, Qin Lv, and Shivakant Mishra. 2015. Detection of
Character-level techniques outperform token-level ones. There cyberbullying incidents on the instagram social network. CoRR,
are several lists of slurs that can aid in categorization, but only abs/1503.03909.
when they are used in conjunction with other traits. Many more [15]Irene Kwok and Yuzhou Wang. 2013. Locate the hate: Detecting tweets
advanced characteristics, such as rely upon information or against blacks. In Marie desJardins and Michael L. Littman, editors, AAAI,
pages 1621– 1622, Bellevue, Washington, USA. AAAI Press.
features that mimic certain linguistic constructions, such as [16] Ravipati, S., Mani, V., & Yarlagadda, S. R. (2021). Efficient Control of
imperatives or politeness, have been demonstrated to be useful.. Sensorless Hybrid Electric Vehicle Using RBFN Controller. Studies in
Textual analysis may not be the only way to determine whether Informatics and Control, 30(4), 87-97
or not someone is spewing hate speech. There is a chance that [17] Dasari, M. S., & Mani, V. (2020). Simulation and analysis of three-phase
parallel inverter using multicarrier PWM control schemes. SN Applied
information gained from other modalities (such as pictures sent Sciences, 2(5), 1-10
along with text messages) might be useful as well. In many [18] C Bhuvaneshwari, A Manjunathan, “Reimbursement of sensor nodes and
situations, the only data sets that may be used to make path optimization”, Materials Today: Proceedings, 2021, 45, pp.1547-1551
judgments regarding the overall efficacy of these complicated [19] Roselin Suganthi Jesudoss, Rajeswari Kaleeswaran, Manjunathan
Alagarsamy, Dineshkumar Thangaraju, Dinesh Paramathi Mani,
characteristics are those that are not publicly available and that Kannadhasan Suriyan, “Comparative study of BER with NOMA system in
exclusively cover a specific subtype of hate speech, such as different fading channels”, Bulletin of Electrical Engineering and
bullying of certain ethnic minority. When it comes to Informatics, 2022, 11(2), pp. 854–861
identifying hate speech, there is a need for a uniform data set
that can be used to compare characteristics and approaches.

978-1-6654-9634-6/22/$31.00 ©2022 IEEE 1277

d licensed use limited to: Alvas Institute of Engineering and Technology Department of Library and Information Centre. Downloaded on February 17,2025 at 10:19:44 UTC from IEEE Xplore. Restrictio