Source: Gao et al. 2022: HyDE (Precise Zero-Shot Dense Retrieval without Relevance Labels)
Paper introducing Hypothetical Document Embeddings: embed an LLM-written fake answer instead of the query.
Contents
- What: research paper (ACL 2023), arXiv:2212.10496
- Author: Luyu Gao, Xueguang Ma, Jimmy Lin, Jamie Callan
- Date: 2022-12-20
- Original: arXiv:2212.10496
Summary
An instruction-following LLM writes a hypothetical document that answers the query; an unsupervised contrastive encoder embeds it, and that vector is used to find real documents.
Key claims
- The hypothetical document may contain false details; the encoder’s dense bottleneck is expected to filter them out.
- Significantly outperforms the unsupervised Contriever retriever and is comparable to fine-tuned retrievers.
- Works across web search, QA and fact verification, and in several languages (e.g. Swahili, Korean, Japanese).