Go to MSLD 2026 Meeting homepageCode-Mixed Telugu-English Hate Speech Detection
Keywords: Natural Language Processing
Abstract: Social influence dialogue systems are emerging as key tools for facilitating nuanced, persuasive, and ethically sound interactions in diverse conversational settings. Hate speech detection in low-resource languages like Telugu presents significant challenges for NLP, with critical implications for ensuring ethical and constructive communication. This study investigates transformer-based models—including TeluguHateBERT, HateBERT, DeBERTa, Muril, IndicBERT, Roberta, and Hindi-Abusive-MuRIL—for classifying hate speech in Telugu. We fine-tune these models using Low-Rank Adaptation (LoRA) to improve efficiency while maintaining performance. Additionally, we explore a multilingual pipeline by translating Telugu text into English using Google Translate to leverage resource-rich English-language models. To further improve robustness, we propose an ensemble approach that aggregates predictions from multiple models via majority voting. Experimental results show that translation generally improves accuracy, with Hindi-Abusive-MuRIL and DeBERTa performing best on the translated dataset. The ensemble method further enhances overall performance, achieving the highest F1 scores across both original and translated data. Notably, Hindi-Abusive-MuRIL consistently outperforms individual models in both settings, demonstrating strong cross-lingual generalization. By integrating these hate speech detection techniques into dialogue systems, our work provides a foundation for developing socially responsible and linguistically inclusive conversational agents. This research contributes to the broader goal of advancing ethical and effective multilingual NLP in low-resource environments.
Email Sharing: We authorize the sharing of all author emails with Program Chairs.
Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.
Submission Number: 70
Loading