Aligning Query Representation with Rewritten Query and Relevance Judgments in Conversational Search

TL;DR

This work addresses context-dependent query understanding in conversational search by proposing QRACDR, a neural retriever that aligns a session query representation with both a rewritten query and a relevant document. It introduces a theoretically motivated alignment framework, using mean-squared-error losses and contrastive learning, augmented with hard negatives to push the session query toward a balanced anchor near $q_n^{\prime}$ and $d_n^+$. The approach is validated across eight datasets, showing consistent gains over state-of-the-art CQR and CDR baselines, with extensive ablations illustrating the importance of supervision signals, negative mining, and the interaction between rewriting and relevance judgments. The results demonstrate improved robustness and transferability to low-resource and ad-hoc scenarios, highlighting the practical potential of joint supervision for conversational dense retrieval.

Abstract

Conversational search supports multi-turn user-system interactions to solve complex information needs. Different from the traditional single-turn ad-hoc search, conversational search encounters a more challenging problem of context-dependent query understanding with the lengthy and long-tail conversational history context. While conversational query rewriting methods leverage explicit rewritten queries to train a rewriting model to transform the context-dependent query into a stand-stone search query, this is usually done without considering the quality of search results. Conversational dense retrieval methods use fine-tuning to improve a pre-trained ad-hoc query encoder, but they are limited by the conversational search data available for training. In this paper, we leverage both rewritten queries and relevance judgments in the conversational search data to train a better query representation model. The key idea is to align the query representation with those of rewritten queries and relevant documents. The proposed model -- Query Representation Alignment Conversational Dense Retriever, QRACDR, is tested on eight datasets, including various settings in conversational search and ad-hoc search. The results demonstrate the strong performance of QRACDR compared with state-of-the-art methods, and confirm the effectiveness of representation alignment.

Figures (6)

• Figure 1: A conceptual illustration for the three types of methods. QRACDR has a Query Representation Alignment goal to help achieve more effective end-to-end optimization toward search within an ongoing conversation.
• Figure 2: A conceptual illustration of the defined hyper-sphere $\mathcal{S}_\epsilon$ and the corresponding conceptions on it. The region outside the hyper-sphere denotes the whole representation space $\mathcal{X}$. The goal of achieving query representation alignment is to enable the learned session query $q_n^s$ fall into the aligned area with the help of the fixed representation of rewritten query $q_n^{\prime}$ and relevant document $d_n^+$.
• Figure 3: Impacts of decomposition terms in MSE based on our best QRACDR (CL + MSE) and regularization in ConvDR.
• Figure 4: Model behavior at the average dot product score of two representations of the same query on two datasets.
• Figure 5: Model behavior at the average dot product score of queries and their nearest relevant documents on two datasets.

Theorems & Definitions (4)

• definition 1: Anchor
• definition 2: Aligned Area
• definition 3: Non-Aligned Area
• theorem 1