ConvMix: A Mixed-Criteria Data Augmentation Framework for Conversational Dense Retrieval

Conversational dense retrieval depends on relevance judgments that reveal a user’s search intent across multiple context-dependent turns, but suitable training data are scarce and costly to annotate. ConvMix introduces a mixed-criteria augmentation framework that uses large language models to generate scalable, two-sided relevance judgments from multiple perspectives. Semantic diversity controls, noise filtering, and near-distribution supervision help maintain sample quality and reduce distribution mismatch when human-annotated and generated data are combined. Evaluations on five conversational search benchmarks demonstrate consistent improvements over existing augmentation approaches.

📄 In Proceedings of the 40th AAAI Conference on Artificial Intelligence (AAAI 2026), Volume 40, Issue 18, pp. 15555–15563.
🔗 Paper Link (AAAI)