Core Concepts
Bias and unfairness are emerging as critical challenges in information retrieval (IR) systems that integrate large language models (LLMs), threatening the reliability and trustworthiness of these systems. This survey provides a unified perspective on these issues as distribution mismatch problems and systematically reviews the causes and mitigation strategies across different stages of LLM integration into IR.
Abstract
This survey presents a comprehensive analysis of the emerging bias and unfairness challenges in information retrieval (IR) systems that integrate large language models (LLMs). It first provides a unified perspective on these issues, framing them as distribution mismatch problems.
In the data collection stage, the survey discusses two key types of bias: source bias, where IR models favor LLM-generated content over human-authored content, and factuality bias, where LLMs produce content that deviates from factual information. Mitigation strategies include data augmentation, data filtering, and leveraging external knowledge bases.
During model development, the survey covers four types of bias: position bias, where LLM-based IR models prefer content from specific input positions; popularity bias, where models prioritize popular items; instruction-hallucination bias, where models deviate from user instructions; and context-hallucination bias, where models generate content inconsistent with the context. Mitigation approaches involve prompting, data augmentation, rebalancing, and improving LLM memory and processing capabilities.
In the result evaluation stage, the survey examines selection bias, where LLM-based evaluators favor responses at specific positions or with certain ID tokens; style bias, where evaluators prefer responses with particular stylistic features; and egocentric bias, where evaluators exhibit a preference for outputs generated by themselves or similar LLMs. Mitigation strategies include prompting, data augmentation, and rebalancing.
The survey also discusses fairness issues, categorizing them into user fairness, where IR systems should provide equitable and non-discriminatory services to different users, and item fairness, where systems should afford more opportunities to weaker items. Mitigation approaches span data augmentation, data filtering, rebalancing, regularization, and prompting.
Finally, the survey highlights several key challenges and future directions, including the need to address feedback loops, develop unified mitigation frameworks, provide theoretical analysis and guarantees, and establish better benchmarks and evaluation protocols.
Stats
The training data of LLMs often contains a significant amount of low-quality, factually incorrect, and long-distance repetitive content, which can harm the factual correctness of the text generated by LLMs.
LLMs-based IR models tend to favor content generated by LLMs over human-authored content with similar semantics.
LLMs-based IR models often exhibit a preference for content positioned at the beginning or end of a list, neglecting the contributions of items in the middle.
LLMs-based IR models are more prone to generating unfair outcomes for items compared to traditional models.
Quotes
"LLMs often struggle to adhere fully to users' instructions across various natural language processing tasks, such as dialogue generation, question answering and summarization."
"When acting as evaluators, LLMs demonstrate a clear bias towards outputs generated by themselves over those from other models or human contributors."
"Utilizing explicit user-sensitive attributes like gender or race in LLMs may lead to the generation of discriminated recommendation results or unfair answers to specific questions."