History-Aware Conversational Dense Retrieval (HAConvDR): arXiv Paper Overview

Forum topic · 小凯 · 2026-07-05

Summary

HAConvDR (History-Aware Conversational Dense Retrieval) is an arXiv paper (2401.16659, January 2024) by Fengran Mo, Chen Qu, Kelong Mao, Tianyu Zhu, Zhan Su, Kaiyu Huang and colleagues addressing weaknesses in conversational dense retrieval. Current approaches typically fine-tune a pre-trained ad-hoc retriever on entire conversational sessions, which can be lengthy and noisy, and they are constrained by the limited manual supervision signals available in existing datasets. The paper proposes two key ideas: context-denoised query reformulation, which filters irrelevant content from conversation history, and automatic mining of supervision signals based on the actual impact of historical turns, reducing reliance on manual annotation. Experiments on two public conversational search datasets show that HAConvDR improves history modeling capability, with especially strong gains on long conversations involving topic shifts. The work is relevant to conversational search, query reformulation, dense retrieval, and LLM-era retrieval-augmented systems, and is cross-referenced with surveys and related work on agentic conversational search.

History-Aware Conversational Dense Retrieval (HAConvDR)

This post summarizes an academic paper on conversational search originally shared on zhichai.net.

Metadata

| Field | Content | |---|---| | Title | History-Aware Conversational Dense Retrieval | | Authors | Fengran Mo, Chen Qu, Kelong Mao, Tianyu Zhu, Zhan Su, Kaiyu Huang, et al. (7 authors total) | | Published | 2024-01-30 | | Source | https://arxiv.org/abs/2401.16659 | | Type | Academic paper | | Category | Conversational Search |

One-Line Summary

HAConvDR improves conversational dense retrieval through context-denoised query reformulation and automatic mining of supervision signals based on the actual impact of historical turns.

Background and Motivation

Conversational search enables complex information retrieval through multi-turn interactions between users and the system. Supporting such interactions requires a comprehensive understanding of the conversational inputs so that a good search query can be formulated based on historical information. In particular, the reformulated query should include relevant information from previous conversation turns.

However, current approaches for conversational dense retrieval primarily rely on fine-tuning a pre-trained ad-hoc retriever using the whole conversational search session, which can be lengthy and noisy. Moreover, existing approaches are limited by the amount of manual supervision signals in existing datasets.

Core Contributions

Context-denoised query reformulation: reduces the noise from lengthy conversation histories when building the query representation.
Automatic supervision signal mining: derives training signals from the actual impact of historical turns, reducing dependence on limited manual annotations.
Demonstrates improved history modeling on two public conversational search benchmarks, particularly for long conversations with topic shifts.

Evaluation

The paper reports experiments on two public conversational search datasets, comparing against existing conversational dense retrieval baselines. Results show HAConvDR's stronger ability to model conversation history; exact metric values should be verified against the original PDF tables.

Takeaways for Search/Recommendation Practitioners

1. Architecture: cascade retrieval + reranking + generation remains mainstream, but agentic paradigms are making retrieval strategy itself learnable. 2. Data: high-quality instruction data and click/session logs are critical; synthetic data needs leakage and distribution-shift safeguards. 3. Evaluation: gaps between offline metrics and online satisfaction are widening; LLM-as-judge should be cross-validated with human evaluation. 4. Production: latency, cost, explainability, and safety are hard constraints for industrial deployment.

Related Entries

A Survey of Conversational Search (Sep 2025, ACM)
Engineering Conversational Search Systems: A Review (arXiv 2407.00997)
Agentic Conversational Search with Contextualized Reasoning via Reinforcement Learning
CTR-Guided Generative Query Suggestion in Conversational Search (EMNLP)
ChatRetriever: Adapting LLMs for Generalized and Robust Conversational Dense Retrieval (arXiv 2404.13556)
CoSearchAgent: A Lightweight Collaborative Search Agent (arXiv 2402.06360)

Original Abstract

> Conversational search facilitates complex information retrieval by enabling multi-turn interactions between users and the system. Supporting such interactions requires a comprehensive understanding of the conversational inputs to formulate a good search query based on historical information. In particular, the search query should include the relevant information from the previous conversation turns. However, current approaches for conversational dense retrieval primarily rely on fine-tuning a pre-trained ad-hoc retriever using the whole conversational search session, which can be lengthy and noisy. Moreover, existing approaches are limited by the amount of manual supervision signals in the existing datasets. To address the aforementioned issues, we propose a History-Aware Conversational Dense Retrieval (HAConvDR) system, which incorporates two ideas: context-denoised query reformulation and automatic mining of supervision signals based on the actual impact of historical turns. Experiments on two public conversational search datasets demonstrate the improved history modeling capability of HAConvDR, in particular for long conversations with topic shifts.