Understanding Interpretability: Explainable AI Approaches for Hate Speech Classifiers

Cyberbullying and hate speech are two of the most significant problems in today's cyberspace. Automated artificial intelligence models might be used to find and remove online hate speech, which would address a critical problem. A variety of explainable AI strategies are being developed to make model judgments and justifications intelligible to people as artificial intelligence continues to permeate numerous industries and make critical change. Our study focuses on mixed code languages (a mix of Hindi and English) and the Indian sub-continent. This language combination is extensively used in SARRAC nations. Three transformer-based models and one machine learning model was trained and fine-tuned on the modified HASOC-Identification of Conversational Hate-Speech in Code-Mixed Languages (ICHCL) data for hate speech classification. Several types of explainability techniques have been explored on the respective models, such as Local interpretable model-agnostic explanations (LIME), Shapley additive explanations (SHAP), and model attention, to analyze model behavior. The analysis suggests that better trained models and comparison of Explainable Artificial Intelligence (XAI) techniques would provide better insight.