A Domain-Transfer Meta Task Design Paradigm for Few-Shot Slot Tagging

Few-shot slot tagging is an important task in dialogue systems and attracts much attention of researchers. Most previous few-shot methods utilize meta-learning procedure for training strive to construct a large number different meta tasks simulate the testing situation insufficient data. However, there widespread phenomenon overlap between two domains tagging. Traditional ignore this special cannot such realistic scenarios. It violates basic principle which consistent with real task, leading historical information forgetting problem. In paper, we introduce novel domain-transfer design paradigm tackle We distribute domain each target based on coincidence degree labels these domains. Unlike classic only rely small samples domain, our aim correctly infer class query both abundant data scarce domain. To accomplish propose Task Adaptation Network effectively transfer from carry out sufficient experiments benchmark dataset SNIPS name entity recognition NER. Results demonstrate that proposed model outperforms achieves state-of-the-art performance.