Ensemble Classifier for Hindi Hostile Content Detection

Detection of hostile content from social media posts ( Facebook TM , Twitter etc.) is a demanding task in the field Natural Language Processing (NLP). Daily growing nature different electronic opened up new challenges language understanding. It becomes more difficult regional languages. AI-based solution required to identify on large scale. Though satisfactory amount researches has been carried out English language, finding languages still under progress due unavailability suitable datasets and tools. In terms number speakers, Hindi ranks third world first Indian Subcontinent. The objective article design detection system using coarse-grained (binary) classification fine-grained (multi-class, multi-label) classification. We noted that baseline learning method with pre-trained models perform differently. Using Constraint 2021 Dataset, this research proposes Bidirectional Encoder Representations Transformers (BERT) based contextual embedding technique concatenation emoji2vec Embedings classify Devanagari script as or non-hostile. Additionally, for tasks where are sub-categorized defamation, fake, hate, offensive, we develop an Ensemble Classifier varying methods models. With F1-Score 0.9721, it found our proposed Indic-BERT+emoji model outperforms other existing task. have also observed giving good results than 0.43, 0.82, 0.58 0.62 offensive classes respectively. code data available https://github.com/skarifahmed/hostile.