Hierarchical Hybrid Code Networks for Task-Oriented Dialogue

Abstract

Task-oriented dialog system is a research hotspot in natural language processing field. In recent years, the application of neural network (NN) has greatly improved the performance of dialog agent. However, there is still a big gap of performance between human beings and dialog agent, in which the domain knowledge and semantic analysis are not well exploited. In this paper we propose a model of Hierarchical Hybrid Code Networks (HHCNs), in which a word-character RNN for semantic representation and a NN-based selection for domain knowledge are integrated. Thus the proposed HHCNs can effectively conduct semantic analysis (e.g., identify proper nouns and misspelling word) and select meaningful responses for the dialog. The experimental results on the dataset of Dialog State Tracking Challenge 2 (DSTC2) have shown a superior performance of HHCNs.