Episode-Based Prompt Learning for Any-Shot Intent Detection.

Emerging intents may have zero or a few labeled samples in realistic dialog systems. Therefore, models need to be capable of performing both zero-shot and few-shot intent detection. However, existing zero-shot intent detection models do not generalize well to few-shot settings and vice versa. To this end, we explore a novel and realistic setting, namely, any-shot intent detection. Based on this new paradigm, we propose E pisode-based P rompt L earning (EPL) framework. The framework first reformulates the intent detection task as a sentence-pair classification task using prompt templates and unifies the different settings. Then, it introduces two training mechanisms, which alleviate the impact of different prompt templates on performance and simulate any-shot settings in the training phase, effectively improving the model’s performance. Experimental results on four datasets show that EPL outperforms strong baselines by a large margin on zero-shot and any-shot intent detection and achieves competitive results on few-shot intent detection.