Source: Ma et al. 2023: Query Rewriting for Retrieval-Augmented LLMs

EMNLP 2023 paper proposing Rewrite-Retrieve-Read, with a small rewriter trained by RL from the reader's feedback.

Contents
  • What: research paper (EMNLP 2023), arXiv:2305.14283
  • Author: Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao, Nan Duan
  • Date: 2023-05-23
  • Original: arXiv:2305.14283

Summary

Argues that the query, not just the retriever or reader, should be adapted: rewrite the input before retrieval.

Key claims

  • Pipeline: rewrite the query with an LLM, retrieve with a web search engine, read with a frozen LLM.
  • A trainable variant uses a small language model as rewriter, trained with reinforcement learning from the reader’s feedback.
  • Consistent gains on open-domain and multiple-choice QA.

Feeds

Retrieval-augmented generation

Suggest changes →