IterCQR: Iterative Conversational Query Reformulation with Retrieval Guidance

TL;DR

This work tackles conversational search by reframing context-dependent queries into stand-alone forms without relying on human rewrites. It introduces IterCQR, an iterative, IR-guided training framework that starts from an LLM-generated rewrite dataset and progressively improves via Minimum Bayes Risk training and Top-1 candidate selection, using a cosine-similarity reward between candidate rewrites and gold passages. Across TopiOCQA and QReCC, IterCQR achieves state-of-the-art results for both dense and sparse retrievers, and exhibits strong generalization to unseen datasets as well as robustness in low-resource settings. The approach reduces dependency on manually crafted rewrites while aligning reformulations with retriever behavior, offering practical gains for real-world conversational search systems.

Abstract

Conversational search aims to retrieve passages containing essential information to answer queries in a multi-turn conversation. In conversational search, reformulating context-dependent conversational queries into stand-alone forms is imperative to effectively utilize off-the-shelf retrievers. Previous methodologies for conversational query reformulation frequently depend on human-annotated rewrites. However, these manually crafted queries often result in sub-optimal retrieval performance and require high collection costs. To address these challenges, we propose Iterative Conversational Query Reformulation (IterCQR), a methodology that conducts query reformulation without relying on human rewrites. IterCQR iteratively trains the conversational query reformulation (CQR) model by directly leveraging information retrieval (IR) signals as a reward. Our IterCQR training guides the CQR model such that generated queries contain necessary information from the previous dialogue context. Our proposed method shows state-of-the-art performance on two widely-used datasets, demonstrating its effectiveness on both sparse and dense retrievers. Moreover, IterCQR exhibits superior performance in challenging settings such as generalization on unseen datasets and low-resource scenarios.

Figures (6)

• Figure 1: In the CQA task, the user's queries are dependent on the previous dialogue context. CQR task reformulates conversational queries into stand-alone queries, which are then fed into the off-the-shelf retrievers.
• Figure 2: Overview of IterCQR. IterCQR trains on the candidates generated by the previous iteration model. We define reward as the cosine similarity between the frozen dense passage embeddings and dense candidate embeddings.
• Figure 3: OnlyMBR is the model trained only with the MBR algorithm, and OnlyTop1 is trained only with the Top-1 candidate selection.
• Figure 4: IterCQR dense retrieval performance on TopiOCQA and QReCC datasets for each iteration.
• Figure 5: Effect of iterative setting on queries. Overlapping tokens in (a) and (c) is shown by the Sørensen-Dice coefficient, (b) is reported by the average token length of the reformulated queries, and (d) represents the proportion of distinct 3-gram tokens per query. All results are derived from the TopiOCQA test set.