AI-based Content Moderator for Devanagari Scripts

Social media platforms have transformed international communication, but they have also turned into havens for hate speech and other offensive material, endangering public safety, social cohesiveness, and mental health. In addition to failing to effectively manage the complexities of many linguistic and cultural settings, traditional content moderation techniques frequently find it difficult to keep up with tremendous online interactions. A fundamental user interface framework specifically designed for texts written in Devanagari-scripted languages, such as Hindi, Marathi, Nepali, and Sanskrit, is the aim of this research. It centralizes on three crucial subtasks: Language Detection, Hate Speech Detection, Target Identification. The models used to inculcate language detection is fastText. The model for hate speech detection is Hate-Speech-CNERG (fine-tuned). The model used for target identification is mBERT. The Precision-Recall Area Under Curve (PR-AUC) is utilized to assess the performance of the models in predicting the number of correct positive predictions made by the models. The PR-AUC score for fastText is 0.92, while Hate-Speech-CNERG has a score of 0.39 and mBERT stands at 0.78. The research proposes a method to combat online toxicity by moderating multilingual content, particularly in Devanagari-scripted languages, using advanced moderation techniques for inclusive communication and diverse voices.