Offensive Text Detection for Tamil Language

In recent years, the Internet and social media have sparked a revolution in the way information is exchanged. The growth of social media and micro-blogging sites not only provides platforms for empowering freedom of expression and individual voices but has also led to a rise in anti-social behavior like online harassment, cyberbullying and hate speech. In multilingual social networks, users often engage in code-mixed communication, blending multiple languages in their posts and comments. There has been significant amounts of work in offensive text detection for the English language. However, the lack of adequate data makes it challenging to detect offensive text, particularly in Indian languages like Tamil, Hindi, and Marathi. This work presents a deep learning model for offensive text detection in Tamil and Tanglish. A CNN-BiGRU-based approach using fastText embeddings is proposed which aims to improve the detection of offensive content in Tamil code-mixed social media. The model was found to have an accuracy of 73.5%. Additional real-time testing was also done with the Twitter data.