Cross-language dialog state tracking using hierarchical attention mechanism

Abstract

The dialog state tracking, which refers to identifying the user&#39

s intent from the utterances, is one of the most important tasks in dialog management. In this paper, we present our dialog state tracker developed for the fifth dialog state tracking challenge (DSTC5), which focused on cross-language adaptation using machine-translated training data, which was very scarce compared to the size of the ontology. Our dialog state tracker is based on the bi-directional long short-term memory (LSTM) with a hierarchical attention mechanism in order to spot important words in user utterances. The intent prediction is done by finding the most relevant keyword in the ontology to the attention-weighted word vector. Our tracker thus has various advantages over existing approaches, such as predicting out-of-vocabulary (OOV) intent due to scarce training data or machine translation. We show that our tracker outperforms other trackers submitted to the challenge in terms of almost all performance measures.