The study introduces CAFFEINE, a context-aware refinement framework for contradictory personas in long-term conversations. This framework aims to improve dialogue systems by refining and expanding personas based on their contextual backgrounds. The paper highlights the issues with human-authored datasets that often provide uninformative persona sentences, thus hindering the quality of responses. The authors propose to transform these contradictory personas into sentences that contain richer speaker information. This approach to persona expansion in multi-session settings is pioneering, and the framework may facilitate better and more human-like response generation.
Publication date: 26 Jan 2024
Project Page: https://caffeine-15bbf.web.app/
Paper: https://arxiv.org/pdf/2401.14215
