Target-Based State and Tracking Algorithm for Spoken Dialogue System

Conventional spoken dialogue systems use frame structure to represent dialogue state. In this paper, we argue that using target distribution to represent dialogue state is much better than using frame structure. Based on the proposed target-based state, two target-based state tracking algorithms are introduced. Experiments in an end-to-end spoken dialogue system with real users are conducted to compare the performance between the target-based state trackers and frame-based state trackers. The experimental results show that the proposed target-based state tracker achieve 97% of dialogue success rate, comparing to 81% of frame-based state tracker, which suggests the advantage of target-based state.