Towards a scalable dialogue state tracker

Goal-Oriented dialogue systems help users accomplish various tasks, for instance book a flight ticket or make a restaurant reservation, etc. In order to complete the tasks, these assistants provide a conversational interface to multiple third-party services/APIs. Dialogue State Tracking (DST) is a core component of such systems. After every user utterance, it outputs the dialogue state which is essentially a summary of the conversation till the current run. With ever-increasing number of services and constantly changing ontologies, existing approaches fall short as they often rely on a single static ontology to work with. In this work, we focus on a specific condition, in which 1) a guided schema is provided, 2) target slot value can be found as word segments in the dialogue context. Our proposed model, an end to end zero-shot dialogue state tracker, uses BERT to encode the dialogue utterance and the guided schema, so to extract the slot values from dialogue context, meanwhile it’s capable to support unseen domains and services without asking for additional training data. In addition, our model supports multilingual dialogues. Empirical evaluation shows the model achieves competitive performance on Google’s schema-guided dataset.