Multilingual Hate Speech Detection Using Semi-supervised Generative Adversarial Network

Abstract

Online communication has overcome linguistic and cultural barriers, enabling global connection through social media platforms. However, linguistic variety introduced more challenges in tasks such as the detection of hate speech content. Although multiple NLP solutions were proposed using advanced machine learning techniques, data annotation scarcity is still a serious problem urging the need for employing semi-supervised approaches. This paper proposes an innovative solution-a multilingual Semi-Supervised model based on Generative Adversarial Networks (GAN) and mBERT models, namely SS-GAN-mBERT. We managed to detect hate speech in Indo-European languages (in English, German, and Hindi) using only 20% labeled data from the HASOC2019 dataset. Our approach excelled in multilingual, zero-shot cross-lingual, and monolingual paradigms, achieving, on average, a 9.23% F1 score boost and 5.75% accuracy increase over baseline mBERT model.