Learning to Bridge Metric Spaces: Few-shot Joint Learning of Intent Detection and Slot Filling

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 3190–3200 August 1–6, 2021. ©2021 Association for Computational Linguistics 3190 Learning to Bridge Metric Spaces: Few-shot Joint Learning of Intent Detection and Slot Filling Yutai Hou∗, Yongkui Lai∗, Cheng Chen, Wanxiang Che†, Ting Liu Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology {ythou, yklai, cchen, car, tliu}@ir.hit.edu.cn Abstract In this paper, we investigate few-shot joint learning for dialogue language understanding. Most existing few-shot models learn a sin- gle task each time with only a few exam- ples. However, dialogue language understand- ing contains two closely related tasks, i.e., in- tent detection and slot filling, and often bene- fits from jointly learning the two tasks. This calls for new few-shot learning techniques that are able to capture task relations from only a few examples and jointly learn multiple tasks. To achieve this, we propose a similarity-based few-shot learning scheme, named Contrastive Prototype Merging network (ConProm), that learns to bridge metric spaces of intent and slot on data-rich domains, and then adapt the bridged metric space to specific few-shot do- main. Experiments on two public datasets, Snips and FewJoint, show that our model sig- nificantly outperforms the strong baselines in one and five shots settings. 1 Introduction Few-Shot Learning (FSL) that committed to learn- ing new problems with only a few examples (Miller et al., 2000; Vinyals et al., 2016) is promising to break the data-shackles of current deep learning. Commonly, existing FSL methods learn a single few-shot task each time. But, real-world applica- tions, such as dialogue language understanding, usually contain multiple closely related tasks (e.g., intent detection and slot filling) and often benefit from jointly learning these tasks (Worsham and Kalita, 2020; Chen et al., 2019; Qin et al., 2019; Goo et al., 2018). In few-shot scenarios, such re- quirements of joint learning present new challenges for FSL techniques to capture task relations from only a few examples and jointly learn multiple tasks. *Equal contributions. †Corresponding author. Train Domain 1 Support Examples: This skirt product has stains issue. | intent: ReturnExchange Search for red feature dress product. | intent: FindProduct Query Example: This waistcoat is not the right size. Train Domain 2 Support Examples: Play the Harry Potter film on the TV device . | intent: PlayVideo Read the book Harry Potter book on headphones device. | intent: PlayVoice Query Example: Play The Lord of the Rings on my mobile. Multi-Media Domain Test Domain Shopping Domain Support Examples: Where can I buy face masks product nearby? | intent: FindShop Find the nearest barbecue food. | intent: FindRestaurant Query Example: Where can I buy bread. Map Domain Figure 1: Examples of the few-shot joint dialogue lan- guage understanding. On each domain, given a few la- beled support examples, the model predicts the intent and slot labels for unseen query examples. Joint learn- ing benefits from capturing the relation between intent and slot labels, but such relation is hard to learn from a few sparse examples and hard …

Content truncated to 500 words