Unsupervised Dialog Structure Learning

Proceedings of NAACL-HLT 2019, pages 1797–1807 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 1797 Unsupervised Dialog Structure Learning Weiyan Shi University of California, Davis wyshi@ucdavis.edu Tiancheng Zhao Carnegie Mellon University tianchez@cs.cmu.edu Zhou Yu University of California, Davis joyu@ucdavis.edu Abstract Learning a shared dialog structure from a set of task-oriented dialogs is an important challenge in computational linguistics. The learned dialog structure can shed light on how to analyze human dialogs, and more impor- tantly contribute to the design and evalua- tion of dialog systems. We propose to ex- tract dialog structures using a modified VRNN model with discrete latent vectors. Different from existing HMM-based models, our model is based on variational-autoencoder (VAE). Such model is able to capture more dynam- ics in dialogs beyond the surface forms of the language. We find that qualitatively, our method extracts meaningful dialog structure, and quantitatively, outperforms previous mod- els on the ability to predict unseen data. We further evaluate the model’s effectiveness in a downstream task, the dialog system build- ing task. Experiments show that, by integrat- ing the learned dialog structure into the reward function design, the model converges faster and to a better outcome in a reinforcement learning setting. 1 Introduction Human dialogs are like well-structured buildings, with words as the bricks, sentences as the floors, and topic transitions as the stairs connecting the whole building. Therefore, discovering dialog structure is crucial for various areas in compu- tational linguistics, such as dialog system build- ing (Young, 2006), discourse analysis (Grosz and Sidner, 1986), and dialog summarization (Murray et al., 2005; Liu et al., 2010). In domain specific tasks such as restaurant booking, it’s common for people to follow a typical conversation flow. Cur- rent dialog systems require human experts to de- sign the dialog structure, which is time consuming and sometimes insufficient to satisfy various cus- tomer needs. Therefore, it’s of great importance to automatically discover dialog structure from exist- ing human-human conversations and incorporate it into the dialog system design. However, modeling human conversation is not easy for machines. Some previous work rely on human annotations to learn dialog structures in su- pervised learning settings (Jurafsky, 1997). But since human labeling is expensive and hard to ob- tain, such method is constrained by the small size of training examples, and by the limited number of application domains (Zhai and Williams, 2014). Moreover, structure annotations on human conver- sation can be subjective, which makes it hard to reach inter-rater agreements. Therefore, we pro- pose an unsupervised method to infer the latent di- alog structure since unsupervised methods do not require annotated dialog corpus. Limited previous work has studied unsuper- vised methods to model the latent dialog struc- ture. Most of the previous methods use the hidden Markov model to capture the temporal dependency within human dialogs (Chotimongkol, 2008; Rit- ter et al., 2010; Zhai and Williams, 2014). We propose to adopt a new type of models, the varia- tional recurrent neural network (VRNN, a recur- …

Content truncated to 500 words