Leveraging Historical Turns for Retrieval-Augmented Response Generation in Conversational Search

ACL ARR 2024 December Submission610 Authors

Abstract:

Conversational search enables the users to interact with the systems by multi-turns to address their complex information needs, which consist of two key components: retrieval and generation. Although retrieval has achieved significant improvement recently by understanding context-dependent queries, response generation has not been well studied. The existing methods only adapt the single-turn retrieval-augmented generation (RAG) pipeline, which overlooks the historical information (e.g., historical search results) as the conversation dives in. In this paper, we first define conversational RAG scenarios and verify the feasibility of leveraging historical turns for current turn RAG, e.g., the historical search results and the turn dependency.
Then, we investigate various strategies toward a better practice for conversational RAG on three public benchmarks and demonstrate the effectiveness of integrating abundant information in historical turns.
We also analyze the potential principle behind our observations, aiming to understand when and why historical information can contribute to the conversational RAG, which could facilitate the build-up of modern conversational search systems.

Paper Type: Long

Research Area: Information Retrieval and Text Mining

Research Area Keywords: Conversational retrieval-augmented generation (RAG), Historical search results, Turn dependency, Conversational search

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 610