Anti-social Behavior Detection using Multi-lingual Model

Abstract

In the current era, social media has emerged as a very useful and reliable means of communication between different people and communities. However, with the leverage of communication platforms and billions of social media users, it became more challenging to stop hateful, abusive, or offensive content spread by extremists that are various aspects of Anti-social Behavior (ASB). Multiple users from several regions use different languages (a mix of native, local and other languages) to express their emotions. Roman Urdu-English and Roman Hindi-English are the two most commonly used languages on social media in the South Asia region. Therefore, the ASB detection with multilingual (multiple languages) model settings represents a wide area of interest for all kinds of social media platforms. Failing to properly address this issue over time on a global scale has already led to morally questionable real-life events, human deaths, and the perpetuation of hate itself. In this paper, we perform a sentimental analysis of the Roman Urdu-English and Roman Hindi-English languages using transformer based mBERT and XLM-R models. Moreover, we process the negatively classified sequences for detection of the ASB.