Relieving Bias in Hate Speech Detection with a Small Number of Expert Annotations: A Prompt-based Learning Approach (SSRN)

The research at hand tackles bias in automatic hate speech detection on social media, where linguistic heterogeneity within speaker groups is frequently overlooked by machine learning models trained on datasets labeled by general annotators. A weakly supervised framework that uses contrastive and prompt-based learning strategies based on huge language models in conjunction with a limited number of expert annotations is suggested as a solution to this problem. A group estimator, pair generator, and knowledge injection are all included into the suggested architecture to improve the model’s sensitivity to sociolinguistic subtleties.

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5487166