Semi-Supervised Few-Shot Intent Classification and Slot Filling

• DOI: 10.48550/arxiv.2109.08754
• Publication Date: 2021-01-01
• Authors: Samyadeep Basu; Karine lp Kiun Chong; Amr Sharaf; Alexander Fischer; Vishal Rohra; Michael Amoake; Hazem El-Hammamy; Ehi Nosakhare; Vijay Ramani; Benjamin Han
• Keywords: Margin (machine learning); Supervised Learning; Natural language understanding; Labeled data

Intent classification (IC) and slot filling (SF) are two fundamental tasks in modern Natural Language Understanding (NLU) systems. Collecting annotating large amounts of data to train deep learning models for such systems is not scalable. This problem can be addressed by from few examples using fast supervised meta-learning techniques as prototypical networks. In this work, we systematically investigate how contrastive unsupervised augmentation methods benefit these existing pipelines jointly modelled IC/SF tasks. Through extensive experiments across standard benchmarks (SNIPS ATIS), show that our proposed semi-supervised approaches outperform methods: losses conjunction with networks consistently the state-of-the-art both IC SF tasks, while strategies primarily improve few-shot a significant margin.