Reproduce “Long Time No See! Open-Domain Conversation with Long-Term Persona Memory” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xinchao Xu, Zhibin Gou, Wenquan Wu, Zheng-Yu Niu, Hua Wu, Haifeng Wang, Shihang Wang
Baidu Inc. and collaborators developed PLATO-LTM, a framework that enhances open-domain dialogue systems with a long-term memory mechanism for dynamic mutual persona management. This system consistently extracts, stores, and retrieves both user and bot persona information, which was shown to improve dialogue consistency by over 100% and increase engagingness in human evaluations.
Key insights
- A two-stage, self-supervised training approach for the Persona Extractor significantly improves its accuracy in identifying persona-related sentences from dialogue history.
- Dynamically managing and explicitly grounding mutual persona information (user and bot) within a generative dialogue model, even without multi-session training data, substantially enhances conversational consistency.
- Differentiating between user and bot personas via role embeddings and special tokens effectively guides the generative model to utilize the retrieved persona information more accurately.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.