User‐Based Evaluation of Explainability Techniques for Misogyny Detection in Code‐Mixed Hindi–English

ABSTRACT Although artificial intelligence (AI) models have demonstrated great success and efficiency in automatically countering online hate speech and misogyny, they suffer from a lack of explainability and transparency. Explainable artificial intelligence (XAI) is an apparent solution to make opaque black‐box models more explainable, but remains under‐explored for low resource and code‐mixed languages. In this study, user‐based evaluation of 3 explainability techniques for a misogyny classifier has been performed through simulatability, what‐if explainability, and a qualitative feedback that covers several key dimensions. The participants were also required to mark the rationales to justify their predictions, and required to answer an open‐ended question regarding their opinion on the explanations. Analysis of the results from 10 participants highlights that Integrated Gradients (IG) was more consistent and plausible than other techniques, and their preference for token attribution heatmaps and bar plots over natural language explanations (NLE).