Self-supervised Schema Induction for Task-oriented Dialog

Anonymous

Abstract: Hand-crafted schemas describing how to collect and annotate dialog corpora are a prerequisite towards building task-oriented dialog systems. In practical applications, manually designing schemas can be error-prone, laborious, iterative, and slow, especially when the schema is complicated. To automate this process, we propose a self-supervised approach for schema induction from unlabeled dialog corpora. Our approach utilizes representations provided by in-domain language models constrained on unsupervised structures, followed by multi-step coarse-to-fine clustering. We compare our method against several strong supervised baselines, and show significant performance improvement in schema induction on MultiWoz and SGD datasets. We also demonstrate the effectiveness of induced schemas on downstream tasks including dialog state tracking and response generation.