DETECTION OF ONLINE HUMAN BEHAVIOUR IN HINDI LANGUAGE POSTS BY LEVERAGING MULTILINGUAL BERT

Abstract

Hindi is the third highest spoken language in the word with almost 662 million people speaking worldwide. Recent trend shows the increase in usage of Hindi as internet language. When individuals use online platforms to express their perspectives, share knowledge, recount personal experiences, and convey emotions, a significant issue arises when these interactions transform into a space for offensive remarks, comments, and conversations. This project is being built with the aim of detecting antisocial and prosocial online behaviour with a reward or feedback system. The whole project is planned in two phases. In the first phase detection of antisocial online post has been done. Prosocial behaviour and feedback system is developed in the second phase. Antisocial behaviourr is mainly divided into four category abusive/offensive, cyberbullying, Targeted group and hate speech. Except these few more categories are like fake, defamation also there. More than 8000 label data have been used in the both the phases of the project. mBERT model with cross entropy loss function and AdamW activation function, was applied. The model achieve maximum F1 score of 0.79. Integration of the two classifiers has been done using weighted average to get a unified score for detection of human behavior.