Domain Specific Embeddings in RNN Frameworks for Hate Span Detection and Classification

The unrestricted use of the internet has led to a significant rise in the dissemination of hate speech online, posing serious threats of harm and violence to individuals and society at large. Social media data often comprises informal and fragmented sentences, interspersed with multiple languages. While numerous researchers are dedicated to identifying and mitigating hate speech in social media content, existing approaches primarily focus on detecting hate speech as a whole, rather than targeting specific hateful words or phrases. To encourage research in this direction, a shared task was formulated during SemEval 2021, a task called Toxic Span Identification, to detect hateful words present in a sentence. Leveraging the dataset provided by the organizers of the task, our approach entails constructing a model comprising three fundamental layers: feature extraction, a Bidirectional Long Short-Term Memory (BiLSTM) layer, and at last, a Conditional Random Field (CRF) layer. Among various embedding techniques, we found that GloVe embeddings yielded superior results with our base model, achieving an F1 score of ${6 1. 2 3 \%}$ when combined with dropout layer. Further Toxic BERT and HateBERT models were used to classify the comments as hateful or non-hateful, HateBERT outperformed the Toxic BERT model.