Tuesday 25 March 2025
Scientists have made a significant breakthrough in detecting hate speech in Telugu, a language spoken by over 80 million people in India. For years, researchers have struggled to develop effective methods for identifying and combating hate speech in low-resource languages like Telugu, which lack the massive datasets that high-resource languages like English enjoy.
The new study employed transformer-based models, including TeluguHateBERT, HateBERT, DeBERTa, Muril, IndicBERT, Roberta, and Hindi-Abusive-MuRIL, to classify hate speech in Telugu. The researchers fine-tuned these models using Low-Rank Adaptation (LoRA) to optimize efficiency and performance.
One of the key findings was that translating Telugu hate speech into English improved accuracy for most models. However, this came at a cost: the F1-score, which measures both precision and recall, decreased significantly. This suggests that translation can help models leverage richer linguistic patterns from well-trained English embeddings but also introduces errors due to the loss of nuances specific to the Telugu language.
The researchers also found that Hindi-Abusive-MuRIL, a model specifically trained on abusive speech in Hindi, outperformed other models in both the original Telugu dataset and the translated dataset. This suggests that models trained on similar languages can generalize well even for unrelated languages like Telugu.
The study’s results highlight the importance of developing language-specific fine-tuning strategies and hybrid approaches that combine multilingual and monolingual training techniques. By leveraging the strengths of each approach, researchers may be able to create more effective hate speech detection systems for low-resource languages like Telugu.
The implications of this research are significant. As social media platforms and online communities become increasingly important spaces for communication and expression, the need for efficient and accurate hate speech detection algorithms grows. The ability to detect and combat hate speech in low-resource languages can help create a safer and more inclusive online environment for speakers of these languages.
Moreover, this study demonstrates the potential for multilingual models to be adapted for use in other low-resource languages. By developing more effective hate speech detection systems for these languages, researchers can help promote linguistic diversity and address the social and cultural needs of underserved communities.
In the future, the authors plan to expand their Telugu hate speech dataset and explore alternative fine-tuning techniques to further improve model performance. They also aim to develop cross-lingual embeddings that retain language-specific characteristics while enhancing generalization across multiple languages.
Cite this article: “Detecting Hate Speech in Telugu: Breaking Barriers with Multilingual Models”, The Science Archive, 2025.
Hate Speech Detection, Telugu Language, Low-Resource Languages, Multilingual Models, Fine-Tuning Strategies, Hybrid Approaches, Natural Language Processing, Machine Learning, Social Media, Online Communities.







