Reproduce “Query Expansion Is More Than Generation: Improving Dense Retrieval through Better Integration” on OpenResearch

About this paper

Siyuan Sun, Mihai Surdeanu

Researchers from the University of Arizona introduced AnchorQE, a framework for integrating LLM-generated query expansions into dense retrieval, and SC-AnchorQE, an unsupervised online calibration strategy. This approach consistently improved retrieval performance over baselines and conventional integration methods across 18 benchmarks while maintaining search efficiency.

Key insights

• The method of integrating LLM-generated expansions is a crucial factor in retrieval performance, often overshadowing the quality of the generated text itself, with conventional integration methods frequently degrading performance.
• AnchorQE is theoretically proven to be equivalent to linear score fusion, offering comparable effectiveness to multi-request fusion methods but requiring only a single index request, thereby preserving efficiency.
• For $\alpha < 1/2$, AnchorQE intrinsically bounds the angular deviation between the original query vector and the expanded vector, limiting how much a potentially misleading expansion can alter the original retrieval direction.

Summary by alphaXiv.