Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization

A new research framework, Multi-Agent Perspectivist Preference Optimization (MAP-PO), has been developed to improve sexism detection in text by accounting for diverse human perceptions. Unlike traditional Natural Language Processing (NLP) systems that collapse annotator disagreement into a majority vote, MAP-PO identifies distinct groups of annotators based on their labeling behavior and fine-tunes separate Large Language Models (LLMs) to reflect each group's perspective. These agents are then coordinated through preference optimization. This approach recognizes the subjective nature of sexism perception, moving beyond a singular 'correct' interpretation.

• Human annotators often disagree on sexism detection, not due to error, but due to differing genuine perceptions.
• Traditional NLP methods for sexism detection typically discard this nuanced disagreement by relying on majority voting.
• The MAP-PO framework clusters annotators based on their labeling behavior, not demographic attributes, to capture divergent perspectives.
• Separate LLM agents are fine-tuned for each identified cluster, enabling the system to represent multiple viewpoints on sexism.
• Agent coordination utilizes preference optimization, incorporating both individual and team-level rewards.

Reported by arXiv — Computers and Society · International. This briefing summarises the publication for executive use; the document itself is not reproduced here.