Sequence-to-Sequence Learning for Task-oriented Dialogue with Dialogue State Representation

Proceedings of the 27th International Conference on Computational Linguistics, pages 3781–3792 Santa Fe, New Mexico, USA, August 20-26, 2018. 3781 Sequence-to-Sequence Learning for Task-oriented Dialogue with Dialogue State Representation Haoyang Wen, Yijia Liu, Wanxiang Che*, Libo Qin, Ting Liu Research Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China {hywen, yjliu, car, lbqin, tliu}@ir.hit.edu.cn Abstract Classic pipeline models for task-oriented dialogue system require explicit modeling the dialogue states and hand-crafted action spaces to query a domain-specific knowledge base. Conversely, sequence-to-sequence models learn to map dialogue history to the response in current turn with- out explicit knowledge base querying. In this work, we propose a novel framework that leverages the advantages of classic pipeline and sequence-to-sequence models. Our framework models a di- alogue state as a fixed-size distributed representation and use this representation to query a knowl- edge base via an attention mechanism. Experiment on Stanford Multi-turn Multi-domain Task- oriented Dialogue Dataset shows that our framework significantly outperforms other sequence- to-sequence based baseline models on both automatic and human evaluation. Title and Abstract in Chinese 面向任务型对话中基于对话状态表示的序列到序列学习 面向任务型对话中，传统流水线模型要求对对话状态进行显式建模。这需要人工定义对 领域相关的知识库进行检索的动作空间。相反地，序列到序列模型可以直接学习从对话 历史到当前轮回复的一个映射，但其没有显式地进行知识库的检索。在本文中，我们提 出了一个结合传统流水线与序列到序列二者优点的模型。我们的模型将对话历史建模为 一组固定大小的分布式表示。基于这组表示，我们利用注意力机制对知识库进行检索。 在斯坦福多轮多领域对话数据集上的实验证明，我们的模型在自动评价与人工评价上优 于其他基于序列到序列的模型。 1 Introduction Task-oriented dialogue system attracts more and more attention with the success of commercial systems such as Siri, Cortana and Echo. It helps users complete specific tasks with natural language. Figure 1 shows a typical example of a task-oriented dialogue, where an agent provides with location information for a user. The requirements for the agents to accomplish users’ demands usually involve querying the knowledge base (KB), like acquiring address from location information KB in Figure 1. Typical machine learning approaches model the problem as a partially observable Markov Decision Process (POMDP) (Williams and Young, 2007; Young et al., 2013), where a pipeline system is intro- duced. The pipeline system consists of four components: natural language understanding (NLU, Tur and De Mori, 2011) , dialogue state tracking (Williams et al., 2013; Williams, 2012), dialogue policy learning (Young et al., 2010) and natural language generation (Wen et al., 2015). Taking the utterance in Figure 1 for example, NLU maps the utterance “Address to the gas station” into semantic slot “POI type”. Dialogue state tracker keeps the probability of “gas station” close to 1 against other values of slot “POI type”. Given a semantic frame as a dialogue state, which is the combina- tion of distributions of these slots, dialogue policy learning generates the next pre-defined system action, This work is licenced under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ * Email corresponding. 3782 Address Distance POI type POI Traffic info 638 Amherst St 3 miles grocery store Sigona Farmers Market car collision nearby 269 Alger Dr 1 miles coffee or tea place Cafe Venetia car collision nearby 5672 barringer street 5 miles certain address 5672 barringer street no traffic 200 Alester Ave 2 miles gas station Valero road block nearby 899 Ames Ct 5 miles hospital Stanford Childrens Health moderate traffic 481 Amaranta Ave 1 miles parking garage Palo Alto Garage R …

Content truncated to 500 words