Source: Wang et al. 2023: Query2doc
EMNLP 2023 paper (Microsoft) that expands a query with an LLM-written pseudo-document, improving BM25 by 3-15% without any fine-tuning.
Contents
- What: research paper (EMNLP 2023), arXiv:2303.07678
- Author: Liang Wang, Nan Yang, Furu Wei (Microsoft Research)
- Date: 2023-03-14 (v2 2023-10-11)
- Original: Query2doc: Query Expansion with Large Language Models
Summary
A query-expansion method: few-shot prompt an LLM to write a pseudo-document for the query, then concatenate the query and pseudo-document and search with that. It is close in spirit to HyDE but keeps the original query in the search input and targets sparse (BM25) as well as dense retrievers.
Key claims
- No retriever or LLM fine-tuning is needed; the LLM is used few-shot.
- Improves BM25 by 3% to 15% on MS MARCO and TREC DL ad-hoc retrieval.
- Also helps state-of-the-art dense retrievers, in-domain and out-of-domain.
- The generated text helps disambiguate the query.