Cyberbullying Detection using BERT for Telugu Language

The rapid proliferation of online communication has introduced cyberbullying as a significant concern affecting individuals' well-being. Existing research employs various techniques like Tf-Idf, XLM-RoBERTa, and machine learning algorithms such as Logistic Regression, Random Forest, and Naive Bayes to detect cyberbullying across mixed and bilingual languages. However, these approaches often struggle with accuracy and fail to effectively discern cyberbullying instances due to language nuances and context misinterpretation. Key challenges faced by previous systems include limited linguistic coverage, contextual understanding, and nuanced interpretation of cyberbullying. The new advancement to address these challenges is the implementation of BERT (Bidirectional Encoder Representations from Transformers) architecture by leveraging bidirectional context understanding, allowing it to capture subtle linguistic nuances and contextual cues, thereby improving accuracy and contextual understanding. The proposed model is advancing further by integrating specialized models like IndicBERT, specifically tailored for languages like Telugu. By focusing on contextual nuances, our model aims to improve precision and accuracy of cyberbullying detection for a local language, Telugu content. This study has developed a local language, Telugu dataset comprising 27,000 sentences and achieve an accuracy rate of 90%, highlighting the efficacy of our approach in overcoming these challenges and contributing to online safety.