Unsupervised Dialog Structure Learning

• DOI: 10.48550/arxiv.1904.03736
• Publication Date: 2019-01-01
• Authors: Weiyan Shi; Tiancheng Zhao; Zhou Yu
• Keywords: Autoencoder; Dialog system

Learning a shared dialog structure from set of task-oriented dialogs is an important challenge in computational linguistics. The learned can shed light on how to analyze human dialogs, and more importantly contribute the design evaluation systems. We propose extract structures using modified VRNN model with discrete latent vectors. Different existing HMM-based models, our based variational-autoencoder (VAE). Such able capture dynamics beyond surface forms language. find that qualitatively, method extracts meaningful structure, quantitatively, outperforms previous models ability predict unseen data. further evaluate model's effectiveness downstream task, system building task. Experiments show that, by integrating into reward function design, converges faster better outcome reinforcement learning setting.