Bridging the Gap in Health Video Retrieval: A Multi-Turn Dialogue Benchmark and Two-Stage Framework

The increasing availability of health-related videos offers valuable opportunities for clinical training, patient rehabilitation, and health education. However, existing retrieval methods in this domain remain limited, particularly when users need to refine their information needs through iterative interactions. To address this gap, we introduce the task of interactive multi-turn semantic retrieval for health videos, which enables users to progressively narrow and specify their queries in dialogue with an intelligent retrieval system. To support this task, we construct a new dataset, named MHVRC (Multi-Turn Health Video Retrieval Corpus), by combining VideoChat-Flash–generated video descriptions with DeepSeek-based refinement of coarse queries into realistic follow-up questions. A user study demonstrates that the resulting multi-turn queries better capture fine-grained procedural semantics than existing benchmarks such as MedVidQA. Based on MHVRC, we propose DATR (Dialogue-Aware Two-Stage Retrieval), a two-stage retrieval framework. The first stage adopts a dual-encoder with sparse frame sampling and CLIP-style contrastive learning to achieve efficient coarse retrieval. The second stage re-ranks the top candidates by fusing multi-turn queries and applying a lightweight cross-encoder scoring module, which captures contextual intent shifts and procedural details that single-turn retrieval often misses. Extensive experiments show that DATR achieves state-of-the-art performance on both quantitative metrics and qualitative user evaluations. This work not only establishes the first benchmark for interactive multi-turn health video retrieval but also provides a scalable framework for advancing intelligent medical information access.