Reproduce “ChatRetriever: Adapting Large Language Models for Generalized and Robust Conversational Dense Retrieval” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kelong Mao, Chenlong Deng, Haonan Chen, Fengran Mo, Zheng Liu, Tetsuya Sakai, Zhicheng Dou
ChatRetriever, from Renmin University of China and collaborators, adapts Large Language Models for end-to-end conversational dense retrieval using a Contrastive Session-Masked Instruction Tuning (CSIT) approach. The model achieves competitive performance with state-of-the-art LLM-based query rewriting methods, showing absolute NDCG@3 improvements of 6.8% on CAsT-20 and 12.2% on CAsT-21 compared to prior dense retrievers, while demonstrating enhanced robustness to varied conversational contexts.
Key insights
- Directly adapting LLMs for end-to-end conversational dense retrieval can achieve effectiveness comparable to or surpassing multi-stage LLM-based query rewriting methods.
- Specialized training techniques like session-masked instruction tuning are crucial for enabling LLMs to robustly generalize to complex multi-turn conversational contexts, beyond ad-hoc focused LLM retrievers.
- Combining diverse training data, including general conversational instruction (UltraChat) and ad-hoc search-oriented data (MSMARCO), is essential for developing a universal and effective conversational retriever.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.