ZERO-SHOT INTENT RECOGNITION IN CONVERSATIONAL AI VIA SEMANTIC PROTOTYPE LEARNING

Intent recognition is a fundamental capability of conversational AI systems, but an existing models require large amount of labeled training samples or have difficulty in transferring to new intents.In this paper, we propose a semantic prototype learning paradigm for zero-shot intent recognition to be capable of recognizing user intentions that do not appear in the training set.Different from the intent representations learned with only annotated examples, our approach generates semantic prototypes from intent descriptions and domain bits, while mapping both user utterances and intent definitions to a joint semantic space.The intention is then classified in light of the semantic alignment rather than textual pattern matching.Experiments on multi-domain dialogue datasets demonstrate that the approach achieves significant improvements in performance over unseen intents, while preserving high accuracy on known classes.Furthermore, it facilitates interpretability of the model and allows for scalable domain transfer without expensive data annotation.