Graph Neural Net-Based User Simulator

Xinrui Nie, Zehao Lin, Xinjing Huang, Yin Zhang

Abstract: User Simulators are major tools that enable offline training of task-oriented dialogue systems. To efficiently utilize semantic dialog data and generate natural language utterances, user simulators based on neural network architectures are proposed. However, existing neural user simulators still rely on hand-crafted rules, which is difficult to ensure the effectiveness of feature extraction. This paper proposes the Graph Neural Net-based User Simulator (GUS), which constructs semantic graphs from the corpus and uses them to build Graph Convolutional Network (GCN) to extract feature vectors. We tested our model on examined public dataset and also made conversation with real human directly to verify the effectiveness. Experimental results show GUS significantly outperforms several state-of-the-art user simulators.