# Learning When to Retrieve, What to Rewrite, and How to Respond in Conversational QA

**Authors:** Nirmal Roy, Leonardo F. R. Ribeiro, Rexhina Blloshmi, Kevin Small  
**arXiv:** [2409.15515](https://arxiv.org/abs/2409.15515)  
**Date:** 2024-09-23  
**Summary page:** https://www.summarizepaper.com/en/paper/2409.15515v1/

## Key points

- Authors Nirmal Roy, Leonardo F. R. Ribeiro, Rexhina Blloshmi, and Kevin Small focus on augmenting Large Language Models (LLMs) with information retrieval capabilities through Retrieval-Augmented Generation (RAG).
- The study emphasizes the importance of understanding users' contextual search intent in conversational question answering (QA), an area that has been largely understudied.
- Conversational QA presents unique challenges compared to single-turn QA, requiring systems to comprehend conversational context and manage retrieved passages over multiple turns.
- The authors propose a novel method within the SELF-multi-RAG framework that enables LLMs to determine when retrieval is necessary based on the conversational context at hand.
- Their approach showcases enhanced capabilities in retrieving relevant passages and evaluating response quality in conversational settings.
- Experimental results on three conversational QA datasets show a notable improvement of approximately 13% as measured by human annotation, validating the effectiveness of SELF-multi-RAG in enhancing response generation capabilities.

## Summary

In their paper titled "Learning When to Retrieve, What to Rewrite, and How to Respond in Conversational QA," authors Nirmal Roy, Leonardo F. R. Ribeiro, Rexhina Blloshmi, and Kevin Small delve into the realm of augmenting Large Language Models (LLMs) with information retrieval capabilities through Retrieval-Augmented Generation (RAG). This approach has shown significant benefits for knowledge-intensive tasks. The focus of their study is on understanding users' contextual search intent in conversational question answering (QA), a topic that has been largely understudied. Conversational QA poses unique challenges compared to single-turn QA as systems must grapple with comprehending conversational context and managing retrieved passages over multiple turns. To address this issue, the authors propose a novel method that enables LLMs to determine when retrieval is necessary in RAG settings based on the conversational context at hand. If deemed essential, the LLM rewrites the conversation for passage retrieval and evaluates the relevance of returned passages before generating responses. Building upon the single-turn SELF-RAG framework introduced by Asai et al. in 2023, the authors present SELF-multi-RAG specifically tailored for conversational settings. Their approach showcases enhanced capabilities over single-turn variants in terms of retrieving relevant passages by leveraging summarized conversational context and evaluating the quality of generated responses. The effectiveness of SELF-multi-RAG is validated through experiments conducted on three conversational QA datasets, demonstrating a notable improvement of approximately 13% as measured by human annotation. This research contributes valuable insights into enhancing response generation capabilities in conversational QA scenarios and sheds light on the importance of integrating information retrieval techniques within large language models for more effective knowledge dissemination and interaction with users.