Multi-level Feature Combination in Dialogue State Tracking

Abstract

Dialogue State Tracking is one of the most important component in task-oriented dialog system, which can update the dialogue state and accurately estimate the compact state representation. This paper introduces a multi-level feature combination to capture correlation on the basis of the entire dialog and slots. Based on the Memory-Network [1], which only consider the relevance of current user utterance with historical dialogue context under turn level, we both consider the correlation at the sentence-level and word-level, and give the slot-value pair as input replacing of candidate state pool to estimate the rare slot-value pair. We use Bi-LSTM and self-attention to capture the correlation between internal words from the utterance, the particular slot-value pair and previous unsettled actions, and finally give a confidence degree score to judge whether the particular slot-value pair is the right answer for the current utterance. Our model achieves a good result on the DSTC-2 dataset about 0.74, which is 3% higher than Memory-Network’s 0.71.