Inculcating Context for Emoji Powered Bengali Hate Speech Detection using Extended Fuzzy SVM and Text Embedding Models

The massive growth of social webs offer opportunities to communicate with diverse languages, unstructured text, informal posts, misspelled contents and emojis. Social media users feel comfortable express their emotions specially high intensity (hate speech) in mother tongue. Hate speech any form targets groups individuals that may trigger antisocial activities, hate crimes, terrorist acts. Bengali use for posting implicit or indirect text. Existing detection research considers explicit but actual is expressed more way. In order detect both from low resource content, need highly efficient automated tools. Researchers applied discriminative learning approaches (i.e. SVM, MLP, CNN) distinguish text only clear-cut outcomes detecting direct speech. proposed novel model two parallel approaches: (i) It applies extended fuzzy SVM classifier class imbalanced dataset (FSVMCIL) multilingual BERT (mBERT) embedding first label; (ii) Morphological analysis method content the similarity (HS) scheme second label. Linking labeling methods, this extracts contextual This HS Word2Vec word lexicon. also emoji conversion analysis. study conducts extensive experiments various categories dataset. evaluates performance considering weighted F1 score, precision, recall accuracy parameters. Results reveal significant improvement 2.35% increase F1- score 9.11 % accuracy.