Leveraging Multilingual BERT Variants for Hate Speech Detection in Hinglish Social Media Content

The framework of Transformer-based hate speech detection in Hinglish social media text introduced in this paper was powerful and suitable to scale to large volumes of data, considering the specific issues of the code-mixed language, irregular grammar, and transliteration variability. Through the use of fine-tuned multilingual and monolingual BERT, it can be shown that the proposed approach is capable of capturing the relationships between context and semantics in Hindi-English mixed text, which are hard to model in traditional machine learning and sequence-based neural architecture models.The extensive experiments implemented on a large-scale dataset of Hindi-English mix text of 76,564 samples show that all the tested BERT-based models perform well. The best overall trade-off in terms of accuracy, recall, and macro F1-score of these is multilingual BERT (mBERT) with test accuracy of 98% and macro F1-score of 0.95. The obtained high ROC-AUC values by all the models further reinforce the high discriminative capacity of the models. The analysis of the confusion matrixes indicates that there is little misclassification between hate and non-hate categories, which means that the suggested framework can be trusted and applied in the context of the real-world content moderation process.Competitive analysis of the existing methods reveals that the offered Transformer-based framework is highly robust and generalized, which is why it is more likely to work successfully in the context of the real-world content moderation. On the whole, this paper has shown that multilingual Transformer models are very useful in automated hate speech detection in code-mixed social media settings as well as serve as a promising basis to build safer and more responsible platforms of online communication.

[1]   T. Davidson, D. Warmsley, M. Macy, and I. Weber, “Automated Hate Speech Detection and the Problem of Offensive Language,” Proc. ICWSM, 2017. [2]   Z. Waseem and D. Hovy, “Hateful Symbols or Hateful People? Predictive Features for Hate Speech Detection on Twitter,” Proc. NAACL, 2016. [3]   Y. Zhang, Y. Wu, and Z. Li, “Deep Learning for Hate Speech Detection,” Pattern Recognition Letters, vol. 117, pp. 83–89, 2018. [4]   K. Rudra, A. Sharma, N. Ganguly, and S. Goyal, “Understanding Language Preference for Expression of Opinion and Sentiment in Hindi–English Code-Mixed Text,” Proc. EMNLP, 2016. [5]   R. Swami, T. Chakraborty, and S. Nandi, “Detecting Humor and Hate Speech in Hinglish,” Proc. ICON, 2018. [6]   A. Bohra, D. Vijay, V. Singh, and A. Kumar, “Benchmark Dataset for Hate Speech Detection in Hinglish,” ACL Workshop, 2022. [7]   A. Vaswani et al., “Attention Is All You Need,” Advances in Neural Information Processing Systems (NeurIPS), 2017. [8]   J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding,” Proc. NAACL, 2019. [9]   M. Mozafari, R. Farahbakhsh, and N. Crespi, “Hate Speech Detection Using Deep Learning and BERT,” IEEE Access, vol. 7, pp. 166782–166790, 2019. [10][A. Conneau et al., “Unsupervised Cross-lingual Representation Learning at Scale,” Proc. ACL, 2020.