大语言模型对检索公平性与无偏性的挑战

大语言模型对检索公平性与无偏性的挑战徐君junxu@ruc.edu.cn 中国人民大学高瓴人工智能学院YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP20242OutlineØBias and Unfairness in IR + LLMsØSource Bias in Document RetrievalYSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024Information Retrieval are Everywhere3Product SearchMusicAppsNew BingVideoYSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024IR Meets LLM4[1] Unifying Bias and Unfairness in Information Retrieval: A Survey of Challenges and Opportunities with Large Language Models, KDD 2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024YSSNLP2024A Unified View of Distribution Alignment5ØBias: Target ground truth distribution represents objective and factual realitiesØUnfairness: Target ground truth distribution reflects human values and social contracts and evolves with the progress of timePredicted Distribution ≠ Target Ground Tru