RAC: Retrieval-Augmented Clarification for Faithful Conversational Search

TL;DR

RAC addresses the challenge of generating clarifying questions in open-domain conversational search that are faithfully grounded in the available corpus. It introduces Retrieval-Augmented Clarification (RAC), a framework that conditions clarifications on top-$k$ retrieved passages, and trains a large language model through supervised fine-tuning and direct preference optimization (DPO) to prioritize evidence-based questions. A novel noise-based unfaithful-question generation strategy enables contrastive learning for faithfulness alignment, and the joint objective $\mathcal{L}_{\text{RAC}}(\theta) = \gamma \mathcal{L}_{\text{DPO}}(\theta) + (1-\gamma) \mathcal{L}_{\text{SFT}}(\theta)$ with $\gamma=0.5$ balances faithfulness and fluency. Evaluations on four datasets plus LLM-based judgments show RAC outperforms baselines in both relevance and corpus-faithfulness, with DPO providing additional gains over SFT. The work advances corpus-grounded clarification in conversational search and offers practical methods for producing faithful clarifications conditioned on retrieved evidence.

Abstract

Clarification questions help conversational search systems resolve ambiguous or underspecified user queries. While prior work has focused on fluency and alignment with user intent, especially through facet extraction, much less attention has been paid to grounding clarifications in the underlying corpus. Without such grounding, systems risk asking questions that cannot be answered from the available documents. We introduce RAC (Retrieval-Augmented Clarification), a framework for generating corpus-faithful clarification questions. After comparing several indexing strategies for retrieval, we fine-tune a large language model to make optimal use of research context and to encourage the generation of evidence-based question. We then apply contrastive preference optimization to favor questions supported by retrieved passages over ungrounded alternatives. Evaluated on four benchmarks, RAC demonstrate significant improvements over baselines. In addition to LLM-as-Judge assessments, we introduce novel metrics derived from NLI and data-to-text to assess how well questions are anchored in the context, and we demonstrate that our approach consistently enhances faithfulness.

Figures (3)

• Figure 1: Overview of RAC. Given an ambiguous user query, the system first retrieves the top-$k$ passages ((a) passage retrieval). A mixture of the fine-tuned model and the base model is then used to generate unfaithful clarifying questions. Both faithful and unfaithful clarifying questions are subsequently leveraged for preference optimization via the DPO algorithm ((b) training pipeline). During inference, the trained model directly generates faithful clarifying questions.
• Figure 2: Overview of our proposed training pipeline.
• Figure 3: NLG metrics on ClariQ: impact of varying the number of passages (left) and comparison of retrieval strategies (BM25, TCT, random) using the top 5 retrieved passages (right).