Download as pdf or txt
Download as pdf or txt
You are on page 1of 2

COUNTERING HATE SPEECH ON SOCIAL

MEDIA
Anirudhan Adukkathayar C
Dept. of Computer Science and Engineering
St. Joseph Engineering College
Mangaluru, India
rudhanster@gmail.com

Abstract—The proliferation of social media enables people The vast majority of studies have investigated the
to express their opinions widely online. However, at the same development of computational models for hate speech and
time, this has resulted in the emergence of conflict and hate, offensive language identification tasks predominantly in a
making online environments uninviting for users. Although single language, English. These efforts have relied on simple
researchers have found that hate is a problem across multiple feature engineering methods, including Bag-of-Words
platforms, there is a lack of models efficient system to solve the (BoW), Term Frequency–Inverse Document Frequency (TF-
issue. This research work is intended to identify current IDF), word-level and character-level n-grams [5]
research gap of the systems and propose better solutions for accompanying different traditional supervised classifiers
the same using emerging technologies.
such as Multinomial Naïve Bayes, SVM, and Random Forest
[1], [2], [6]. Furthermore, the impacts of additional features
Keywords—Machine Learning, Multimodal, Deep learning,
Emotion recognition, Multi classification, hate speech, social
such as syntactic and linguistic features, distributional
media semantics (word-embeddings), and user-based and platform-
base metadata have been investigated [10], [7], [8].
I. INTRODUCTION The multilingual nature of social media has underscored
Online hate, described as abusive language , aggression , the importance of hate speech detection in multilingual
cyberbullying, hate-fulness, insults , personal attacks , settings. Several studies have investigated the multilingual
provocation, racism, sexism , threats , or toxicity , has been classification of hate speech and offensive language using
identified as a major threat on online social media platforms. multilingual, cross-lingual, or joint-learning approaches.
Pew Research Center reports that among 4248 adults in the
United States, 41% have personally experienced harassing Authors in [7] employed two different approaches on the
behaviour online, whereas 66% witnessed harassment Waseem and Hovy dataset [5]. The first approach applied the
directed towards others. Around 22% of adults have deep learning structures CNN, LSTM with Random Embed-
experienced offensive name-calling, purposeful ding and GloVe features. The second approach involved
embarrassment (22%), physical threats (10%), and sexual train- ing FastText, CNNs, and LSTM to represent task-
harassment (6%), among other types of harassment. Social specific word embeddings used as features for deep learning
media platforms are the most prominent grounds for such classifiers, including GBDT. LSTM and GBDT achieved the
toxic behaviour. Even though they often provide ways of best result with an F1-score of 93%, outperforming the state-
fagging offensive and hateful content, only 17% of all adults of- the-art.
have fagged harassing conversation, whereas only 12% of In [8], authors treated hate speech and abusive language
adults have reported someone for such acts Manual detection as a multi-class classification problem and
techniques like fagging are neither effective nor easily employed SVM, NB, and RF, for detecting hate speech on
scalable [4] and have a risk of discrimination under Indonesian Twitter data. They used three sets of previously
subjective judgments by human annotators. Since an used datasets [9]–[11] in addition to a new dataset containing
automated system can be faster than human annotation, new tweets they collected. The authors trained their
machine learning models to automatically detect online hate classifiers using various combinations of n-grams and
have been gaining popularity and bringing researchers from sentiment lexicon features. They achieved the best
different fields together. performance using an RF classifier trained on the word
unigram feature. The highest accuracy achieved on their
II. RELATED WORK dataset was77.36%.
Previous work in this area has focused on different
Authors in [12] introduced two ensemble-based models
aspects of hate speech, including, but not limited to: (1) its
to detect hate speech. The first model combined linear SVC,
definition [1], [2] and typology; the data collection and
LR, and RF. The second model combined RF, KNN, and LR.
annotation process (3) investigation of automatic machine
In addition, they used deep learning with word embed- dings.
learning and deep learning [3], [27], [28] classification
They evaluated their models on the Waseem and Hovy
models and their generalisations [9]; (4) investigation of the
dataset [5]. They achieved the best result by using the first
most effective features of hate speech classification [1], [3];
model.
(5) the unintended bias(es) in datasets or classification
models [9]; and (6) some of the relevant ethical principles
[4].
III. RESEARCH GAP IDENTIFIED
Abusive language emerging in a variety of forms,
including hateful and offensive expressions, toxicity, Some of the issues found in the previous works are
misogyny, racism, sexism, cyberbullying, etc. is widely discussed below
poisoning the online social media environment. A wide range Text based / Uni modal
of studies has therefore been dedicated to developing
automatic methods to detect these types of content in social Almost all of the implemented system uses text based
media. identification of hate speech content. Most of the times, the
data is found in a combination of text , images and Video.

2022 IEEE

Handling only a single type of data will not give accurate Number footnotes separately in superscripts. Place the
results. actual footnote at the bottom of the column in which it was
cited. Do not put footnotes in the abstract or reference list.
Uni Lingual Use letters for table footnotes.
Most of the time the concerned data will be represented
in local languages. But the studies and data set are depending Unless there are six authors or more give all authors’
more on English language only. names; do not use “et al.”. Papers that have not been
published, even if they have been submitted for publication,
Binary Classification should be cited as “unpublished” [4]. Papers that have been
In machine learning approach normally binary accepted for publication should be cited as “in press” [5].
classification of data is done. That is the content will be Capitalise only the first word in a paper title, except for
classified as a ‘hate speech’ or as a ‘non hate speech’. But in proper nouns and element symbols.
social media platform, in due course of time a non hate For papers published in translation journals, please give
speech may turn out as a hate speech and trigger some the English citation first, followed by the original foreign-
consequence. language citation [6].
Emotion recognition
It is difficult to interpret the meaning and semantics of a 1. Castelle M. The linguistic ideologies of deep abusive language
statement without identifying the emotions of the user. Due classification. In: Proceedings of the 2nd workshop on abusive
to this, Same statement can be considered as a hate speech as language online (ALW2), Brussels; 2018. P. 160–70
well as non hate speech. It's very difficult to address this 2. Kumar S, et al. Community interaction and conflict on the web. In:
Proceedings of the 2018 world wide web conference on world wide
issue. web; 2018. P. 933–43
Lack of periodic Time based checking 3. Hosseinmardi H et al (2015) Analyzing labeled cyberbullying
incidents on the instagram social network. Soc Inf 2015:49–66
A statement which acts like a non hate speech, can
4. Wachs S et al (2019) Understanding the overlap between
become a hate speech in future, due to a number of triggering cyberbullying and cyberhate perpetration: moderating effects of toxic
factors that occurred in the society. So it's vital to recheck the online disinhibition. Crim Behav Mental Health 29(3):179–188.
contents periodically. https://doi.org/10.1002/cbm.2116
5. Salminen J, et al. Anatomy of online hate: developing a taxonomy and
IV. PROPOSED SOLUTIONS machine learning models for identifying and classifying hate in online
My work proposes a novel system which rectifies the news media. In: Proceedings of the international AAAI conference on
identified research gaps. web and social media (ICWSM 2018), San Francisco; 2018
6. Sood SO et al (2012) Automatic identification of personal insults on
The system should be capable of handling the multi social news sites. J Am Soc Inform Sci Technol 63(2):270–285
modal data. 7. Waseem Z. Are you a racist or am i seeing things? Annotator
influence on hate speech detection on twitter. In: Proceedings of the
It should be able to work on multiple languages and on first workshop on NLP and computational social science; 2016. P.
multiple social media platforms. 138–42
Instead of binary classification, the system should be able 8. Chatzakou D, et al. Measuring #GamerGate: A tale of hate, sexism,
and bullying. In: Proceedings of the 26th inter- national conference
to distinguish the data into different levels. For example, the on world wide web companion, Geneva; 2017. P. 1285–90
data should be classified to hate speech , non hate speech,
9. [1] C.-C. Wang, M.-Y. Day, and C.-L. Wu, “Political Hate Speech
and an intermediate level too. So that it can be used to Detection and. Lexicon Building: A Study in
recheck the contents periodically and ensure that it will not Taiwan,” IEEE Access, vol. 10, pp. 44337–44346, 2022, doi: 10.1109/
turn out into a hate speech. ACCESS.2022.3160712.
10. [2] S. Khan et al., “HCovBi-Caps: Hate Speech Detection Using
Emotion recognition system can also be incorporated to Convolutional and Bi-Directional Gated Recurrent Unit With
get a better result. Capsule Network,” IEEE Access, vol. 10, pp. 7881–7894, 2022, doi:
10.1109/access.2022.3143799.
11. [3] H. Watanabe, M. Bouazizi, and T. Ohtsuki, “Hate Speech on
V. IMPACT ON THE SOCIETY Twitter: A Pragmatic Approach to Collect Hateful and Offensive
Expressions and Perform Hate Speech Detection,” IEEE Access, vol.
Hate speeches have unfortunately begun playing a larger 6, pp. 13825–13835, 2018, doi: 10.1109/access.2018.2806394.
role in hate crimes . For instance, suspects in several recent 12. [5] F. M. Plaza-Del-Arco, M. D. Molina-González, L. A. Ureña-
hate related terror attacks had an extensive social media López, and M. T. Martín-Valdivia, “A Multi-Task Learning Approach
history of hate speeches. So it is the right time to to Hate Speech Detection Leveraging Sentiment Analysis,” IEEE
differentiate hate speeches and the non hate speeches by the Access, vol. 9, pp. 112478–112489, 2021, doi: 10.1109/
technology available. The proposed work should definitely ACCESS.2021.3103697.
prevent the uncontrolled spread of hate among various social
media platforms that targets the marginalised people or
groups. To conclude with, my work would definitely be a
better approach to identify the hate speeches and thereby
taking the counter measures for the same.

References
The template will number citations consecutively within
brackets [1]. The sentence punctuation follows the bracket
[2]. Refer simply to the reference number, as in [3]—do not
use “Ref. [3]” or “reference [3]” except at the beginning of a
sentence: “Reference [3] was the first ...”

You might also like