Dynamic Semantic Matching and Aggregation Network for Few-shot Intent Detection

Few-shot Intent Detection is challenging due to the scarcity of available\nannotated utterances. Although recent works demonstrate that multi-level\nmatching plays an important role in transferring learned knowledge from seen\ntraining classes to novel testing classes, they rely on a static similarity\nmeasure and overly fine-grained matching components. These limitations inhibit\ngeneralizing capability towards Generalized Few-shot Learning settings where\nboth seen and novel classes are co-existent. In this paper, we propose a novel\nSemantic Matching and Aggregation Network where semantic components are\ndistilled from utterances via multi-head self-attention with additional dynamic\nregularization constraints. These semantic components capture high-level\ninformation, resulting in more effective matching between instances. Our\nmulti-perspective matching method provides a comprehensive matching measure to\nenhance representations of both labeled and unlabeled instances. We also\npropose a more challenging evaluation setting that considers classification on\nthe joint all-class label space. Extensive experimental results demonstrate the\neffectiveness of our method. Our code and data are publicly available.\n