SKIM : Few-Shot Conversational Semantic Parsers with Formal Dialogue Contexts

The traditional dialogue state tracking (DST) task tracks the dialogue state given the past history of user and agent utterances. This paper proposes to replace the utterances before the current turn with a formal representation, which is used as the context in a semantic parser mapping the current user utterance to its formal meaning. In addition, we propose TOC (Task-Oriented Context), a formal dialogue state representation. This approach eliminates the need to parse a long history of natural language utterances; however, it adds complexity to the dialogue annotations. We propose Skim, a contextual semantic parser, trained with a sample-efficient training strategy: (1) a novel abstract dialogue state machine to synthesize training sets with TOC annotations; (2) data augmentation with automatic paraphrasing, (3) few-shot training, and (4) self-training. This paper also presents MultiWOZ 2.4, which consists of the full test set and a partial validation set of MultiWOZ 2.1, reannotated with the TOC representation. Skim achieves 78% turn-by-turn exact match accuracy and 85% slot accuracy, while our annotation effort amounts to only 2% of the training data used in MultiWOZ 2.1. The MultiWOZ 2.4 dataset will be released upon publication.