Reproduce “Generative and Pseudo-Relevant Feedback for Sparse, Dense and Learned Sparse Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Iain Mackie, Shubham Chatterjee, Jeffrey Dalton
This paper extends generative-relevance feedback (GRF) to dense and learned sparse retrieval, using language-model-generated text to enrich query representations alongside pseudo-relevance feedback (PRF) from retrieved documents. It also fuses their rankings within each retrieval paradigm. Across 18 settings, fusion has the best Recall@1000 in 14 and significantly beats PRF in 17; this is a recall result, not consistent top-rank improvement.
Key insights
- Generative Relevance Feedback (GRF) is a robust query augmentation technique, effectively improving performance across sparse, dense, and learned sparse retrieval paradigms.
- GRF and traditional Pseudo-Relevance Feedback (PRF) provide complementary signals: GRF injects external world knowledge, while PRF offers corpus-specific grounding.
- GRF is generally more effective for challenging queries with poor initial retrieval, whereas PRF performs better on queries with high initial precision.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.