Machine Learning Approach for Detection of Harassment and Hate Speech in Sinhala Text Messages on Telegram

Nowadays, social media platforms have become an important communication method to connect and express messages without geographical limitations. Nevertheless, harassment and hate speech are some of the main negative outcomes of these platforms, poisoning digital spaces online and leading to depression and even suicide. Hate speech and harassment target individuals or groups based on attributes such as gender, religion, race, or sexuality. According to Sri Lanka CERT (National Centre for Cyber Security), hate crime incidents rose 38% in 2021, with 182 reported cases against 70 in 2020. Although, to overcome that concern, specifically in Sinhala text messages on Telegram, there is a lack of relevant studies. Consequently, this study aims to develop a machine learning model to efficiently identify and mitigate a spectrum of harassment and hate speech, within Sinhala Telegram text messages. The initial step involved developing a Telegram API and a Python application using the Telethon library to collect Telegram text messages automatically. Telegram messages were manually annotated into two categories based on different factors. Sinhala language specific data preprocessing NLP techniques such as stemming, and tokenization were utilized for normalization and feature extraction. A vectorizer was utilized to convert textual data into a format that can be understood by a classifier. The study found the Support Vector Machine classifier to be effective since, when compared to other developed machine learning classifiers such as Random Forest, Logistic Regression, Decision Tree, and Linear SVC, it has an F1-Score of 0.87. All these models exhibited success to varying degrees and were tested on a balanced dataset.