End-to-End Multi-Domain Task-Oriented Dialogue Systems with Multi-level Neural Belief Tracker

Hung Le, Doyen Sahoo, Chenghao Liu, Nancy F. Chen, Steven C.H. Hoi

TL;DR: We proposed an end-to-end dialogue system with a novel multi-level dialogue state tracker and achieved consistent performance on MultiWOZ2.1 in state tracking, task completion, and response generation performance.

Abstract: It has been an open research challenge for developing an end-to-end multi-domain task-oriented dialogue system, in which a human can converse with the dialogue agent to complete tasks in more than one domain. First, tracking belief states of multi-domain dialogues is difficult as the dialogue agent must obtain the complete belief states from all relevant domains, each of which can have shared slots common among domains as well as unique slots specifically for the domain only. Second, the dialogue agent must also process various types of information, including contextual information from dialogue context, decoded dialogue states of current dialogue turn, and queried results from a knowledge base, to semantically shape context-aware and task-specific responses to human. To address these challenges, we propose an end-to-end neural architecture for task-oriented dialogues in multiple domains. We propose a novel Multi-level Neural Belief Tracker which tracks the dialogue belief states by learning signals at both slot and domain level independently. The representations are combined in a Late Fusion approach to form joint feature vectors of (domain, slot) pairs. Following recent work in end-to-end dialogue systems, we incorporate the belief tracker with generation components to address end-to-end dialogue tasks. We achieve state-of-the-art performance on the MultiWOZ2.1 benchmark with 50.91% joint goal accuracy and competitive measures in task-completion and response generation.

Keywords: task-oriented, dialogues, dialogue state tracking, end-to-end