End-to-End Task-Completion Neural Dialogue Systems

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 733–743, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP End-to-End Task-Completion Neural Dialogue Systems Xiujun Li† Yun-Nung Chen⋆ Lihong Li† Jianfeng Gao† Asli Celikyilmaz† †Microsoft Research, Redmond, WA, USA ⋆National Taiwan University, Taipei, Taiwan ⋆y.v.chen@ieee.org †{xiul,lihongli,jfgao,aslicel}@microsoft.com Abstract One of the major drawbacks of modu- larized task-completion dialogue systems is that each module is trained individu- ally, which presents several challenges. For example, downstream modules are af- fected by earlier modules, and the per- formance of the entire system is not ro- bust to the accumulated errors. This pa- per presents a novel end-to-end learning framework for task-completion dialogue systems to tackle such issues. Our neu- ral dialogue system can directly interact with a structured database to assist users in accessing information and accomplish- ing certain tasks. The reinforcement learn- ing based dialogue manager offers robust capabilities to handle noises caused by other components of the dialogue system. Our experiments in a movie-ticket book- ing domain show that our end-to-end sys- tem not only outperforms modularized di- alogue system baselines for both objective and subjective evaluation, but also is ro- bust to noises as demonstrated by several systematic experiments with different er- ror granularity and rates specific to the lan- guage understanding module1. 1 Introduction In the past decade, goal-oriented dialogue systems have been the most prominent component in to- day’s virtual personal assistants, which allow users to speak naturally in order to accomplish tasks more efficiently. Traditional systems have a rather complex and modularized pipeline, consisting of a language understanding (LU) module, a dialogue 1The source code is available at: https://github/ com/MiuLab/TC-Bot. manager (DM), and a natural language generation (NLG) component (Rudnicky et al., 1999; Zue et al., 2000; Zue and Glass, 2000). Recent advances of deep learning have inspired many applications of neural models to dialogue systems. Wen et al. (2017) and Bordes et al. (2017) introduced a network-based end-to-end trainable task-oriented dialogue system, which treated dialogue system learning as the problem of learning a mapping from dialogue histories to system responses, and applied an encoder-decoder model to train the whole system. However, the system is trained in a supervised fashion: not only does it require a lot of training data, but it may also fail to find a good policy robustly due to lack of exploration of dialogue control in the training data. Zhao and Eskenazi (2016) first presented an end-to-end reinforcement learning (RL) approach to dialogue state tracking and policy learning in the DM. This approach is shown to be promising when applied to the task-oriented dialogue prob- lem of guessing the famous person a user thinks of. In the conversation, the agent asks the user a series of Yes/No questions to find the correct an- swer. However, this simplified task may not gen- eralize to practical problems due to the following: 1. Inflexible question types — asking request questions is more natural and efficient than Yes/No questions. For example, it is …

Content truncated to 500 words