Can Hallucination Reduction in LLMs Improve Online Sexism Detection?

Online sexism is a pervasive problem with a significant impact on the targeted individuals and social inequalities. Automated tools are now widely used to identify sexist content at scale, but most of these tools do not provide any further explanations beyond generic categories such as ‘toxicity’, ‘abuse’ or ‘sexism’. This paper explores the impact of hallucination reduction in LLMs on enhancing sexism detection across three different levels: binary sexism, four-categories of sexism, and fine-grained vectors, with a focus on explainability in sexism detection. We have successfully applied Neural Path Hunter (NPH) to GPT-2, with the purpose of “teaching” the model to hallucinate less. We have used hallucination-reduced GPT-2, achieving accuracy rates of 83.2% for binary detection, 52.2% for four-categories classification and 38.0% for the 11-vectors fine-grained classification, respectively. The results indicate that: i) While the model performances may slightly lag behind the baseline models, hallucination-reducing methods have the potential to significantly influence LLM performance across various applications, beyond just dialogue-response systems. Additionally, this method could potentially mitigate model bias and improve generalization capabilities, based upon the dataset quality and the selected hallucination reduction technique.