BSVM: A BERT-Based Support Vector Machine for Hindi Hostile Content Detection

Social media platforms are increasingly providing hostile content. This has caused a requirement for accurate hostile post detection so that appropriate countermove can be made. Increasingly hostile content in various electronic media has created new obstacles for language comprehension. Regional languages make it more challenging. Despite being a good number of studies have been done in the English language, there has not been much progress in Regional languages because the appropriate datasets and tools are not yet available. Hindi is the native language of 615M persons. This research offers a Bidirectional Encoder Representations from Transformers (BERT)-based contextual embedding approach with a combination of Support Vector Machine (SVM) in order to categorize social media posts in Hindi Devanagari script as hostile or non-hostile using the Constraint 2021 Hindi Dataset. Offensive, fake, defamatory, and hateful posts are further evaluated to determine their status. In this research work, several SOTA BERT-based techniques are also subjected to comparative analysis. Our proposed model is found to perform better than the baseline model for all the hostile subclasses (defamation, fake, hate, and offensive).