Towards Effective Long Conversation Generation: Dynamic Topic Tracking and Recommendation for Open-Domain Dialogue Systems

Abstract

The dynamic nature of human conversation necessitates effective topic management and evo- lution in open-domain dialogue systems. This thesis presents EvolvConv, a novel approach for real-time conversation topic tracking and evolution in AI dialogue systems. EvolvConv addresses critical limitations in existing open-domain dialogue systems, which often exhibit performance degradation in extended conversations due to inadequate topic management. The system implements real-time tracking of both conversation topics and user preferences, utilizing this information to facilitate natural topic evolution and shifting based on con- versation state. Through comprehensive experimentation, we evaluate EvolvConv's topic evolution and shifting capabilities across increasing conversation lengths. Using the un- referenced evaluation metric UniEval, we demonstrate that EvolvConv maintains conversa- tion coherence while achieving a controlled topic shift rate of 5-8% at any point throughout the conversation. Comparative analysis shows that EvolvConv generates 4.77% more novel topics than baseline systems while maintaining balanced topic groupings. User evaluation studies validate the practical effectiveness of EvolvConv, with participants preferring its generated responses 47.8% of the time compared to baseline systems, positioning it as the leading artificial system among comparative baselines, second only to human responses. This research contributes to the advancement of more natural and engaging open-domain dialogue systems capable of sustained, evolving conversations.