The paper introduces a novel framework, CAFFEINE, for managing and refining personas in long-term conversations. It addresses the issue of uninformative persona sentences in human-authored datasets by leveraging commonsense-based persona expansion. Unlike previous solutions that focused on avoiding contradictory personas, CAFFEINE transforms these contradictions into rich speaker information by considering their contextual backgrounds. The framework is the first to explore commonsense-based persona expansion in multi-session settings and enables better response generation in long-term conversations.
Publication date: 26 Jan 2024
Project Page: https://caffeine-15bbf.web.app/
Paper: https://arxiv.org/pdf/2401.14215
