Reproduce “Continual Generalized Intent Discovery: Marching Towards Dynamic and Open-world Intent Recognition” on OpenResearch

About this paper

Xiaoshuai Song, Yutao Mou, Keqing He, Yueyan Qiu, Pei Wang, Weiran Xu

The Beijing University of Posts and Telecommunications team introduced Continual Generalized Intent Discovery (CGID) to enable dynamic, open-world intent recognition for dialogue systems. Their proposed method, PLRD, continually discovers new intents from unlabeled out-of-domain data streams and incrementally expands the classifier, achieving an average accuracy improvement of 4.06% on Banking and 2.13% on CLINC datasets over baselines while mitigating catastrophic forgetting.

Key insights

• Continuously learning and integrating new, unlabeled user intents is essential for building adaptive and robust dialogue systems that operate in open-ended environments.
• Combining a minimal memory replay with prototype-guided learning for new class discovery and feature distillation effectively mitigates catastrophic forgetting during continual intent recognition.
• The overall gains from continually discovering new intents significantly outweigh the performance loss on old classes, demonstrating the net benefit of the CGID paradigm.

Summary by alphaXiv.