CATCH: A Framework for Controllable Theme Detection in Dialogue Systems

This paper addresses the challenge of theme detection in user-centric dialogue systems, a crucial task for understanding user intent without predefined schemas. It highlights the limitations of existing methods in handling sparse utterances and user-specific preferences. The proposed CATCH framework offers a novel approach by integrating context-aware topic representation, preference-guided topic clustering, and hierarchical theme generation. The use of an 8B LLM and evaluation on a multi-domain benchmark (DSTC-12) suggests a practical and potentially impactful contribution to the field.