Dense retrievers show bias in political and dialect query results
Retrieval Sensitivity to Identity Signals in Queries
Computation and LanguageComputers and SocietyInformation Retrieval
Summary
Search tools that find documents based on how a question is asked can be influenced by the identity of the person asking, like their political views or dialect. The authors studied how these tools respond differently to questions that express political beliefs or use African American Language compared to mainstream English. They found that these search tools tend to return results reflecting the political lean of the question and perform worse for questions in African American Language. This could make information access unfair and increase social divides if not addressed.
What this means in practice
- •For search engine developers: Evaluate and reduce identity-related biases in search results to improve fairness for users of different political views and dialects.
- •For health information platforms: Improve retrieval accuracy for health questions written in African American Language to avoid reinforcing health disparities.
Authors
Andrew Tang, Nicholas Deas, Kathleen McKeown, Vishal Misra
Abstract
Dense retrievers decide which documents reach users and the language models that use them, yet they are typically evaluated with neutral queries. We ask whether the identity signals that real users express in their queries---political ideology and dialect---bias what a retriever returns. We design evaluations in two domains, political news and consumer-health questions, each pairing a controlled synthetic set that varies only the identity signal with naturalistic queries. Across five dense retrievers and a sparse baseline, every retriever (i) retrieves articles that align with the query's own political lean and (ii) performs worse for questions written in African American Language (AAL) than in White Mainstream English (WME). Two analyses tie these gaps to queries' identity signals beyond surface vocabulary: partialling out an aggregate lexical-asymmetry score leaves the synthetic gaps largely intact, and linear probes recover lean and dialect from the retrievers' query embeddings beyond token-level features. Left unaddressed, such retrieval biases risk contributing to polarization and reinforcing the health disparities already faced by AAL speakers. Code is available at https://github.com/Andrewtcr/bias-ret.