Multilingual Hate Speech Detection Using FastText Embeddings and Transformer-Based Architectures

Abstract

Hate speech detection meanly automatically identifying harmful or abusive type of language written against individuals or communities. As social media is increasingly translated into multiple languages, identifying hate speech in various languages is becoming a more difficult task. Traditional machine learning models are not very good at grasping context. This could impact their classification abilities. In this study, it is suggested to address these limitations by using a multilingual hate speech detection framework using the support of the FastText embeddings alongside the machine learning models and deep learning and transformer-based models. FastText is able to store representations at the subword level, which means that the language acquires adaptability. The experimental result confirms that transformer-based model outperform traditional model with BERT model having highest accuracy of 77%. The findings show the importance of contextual embeddings for strong multilingual hate speech detection making it the best-performing model in this study. The novelty of this study lies in its unified comparative framework and multilingual transformer models for hate speech detection across multiple languages.