Dual Dynamic Memory Network for End-to-End Multi-turn Task-oriented Dialog Systems

Proceedings of the 28th International Conference on Computational Linguistics, pages 4100–4110 Barcelona, Spain (Online), December 8-13, 2020 4100 Dual Dynamic Memory Network for End-to-End Multi-turn Task-oriented Dialog Systems Jian Wang1 Junhao Liu2 Wei Bi Xiaojiang Liu Kejing He1∗ Ruifeng Xu3 Min Yang2∗ 1South China University of Technology 2Shenzhen Key Laboratory for High Performance Data Mining, Shenzhen Institutes of Advanced Technology, Chinese Academy of Sciences 3Harbin Institute of Technology (Shenzhen) cs wangjian@mail.scut.edu.cn, {jh.liu, min.yang}@siat.ac.cn {victoriabi, kieranliu}@gmail.com kjhe@scut.edu.cn, xuruifeng@hit.edu.cn Abstract Existing end-to-end task-oriented dialog systems struggle to dynamically model long dialog context for interactions and effectively incorporate knowledge base (KB) information into dialog generation. To conquer these limitations, we propose a Dual Dynamic Memory Network (DDMN) for multi-turn dialog generation, which maintains two core components: dialog memory manager and KB memory manager. The dialog memory manager dynamically expands the dialog memory turn by turn and keeps track of dialog history with an updating mechanism, which encourages the model to filter irrelevant dialog history and memorize important newly coming information. The KB memory manager shares the structural KB triples throughout the whole conversation, and dynamically extracts KB information with a memory pointer at each turn. Experimental results on three benchmark datasets demonstrate that DDMN significantly outperforms the strong baselines in terms of both automatic evaluation and human evaluation. Our code is available at https://github.com/siat-nlp/DDMN. 1 Introduction Task-oriented dialog systems are designed to help users achieve specific goals with natural language, such as weather inquiry or restaurant reservation. Compared with traditional pipeline methods (Williams and Young, 2007; Young et al., 2013), end-to-end approaches recently have gained much attention (Zhao et al., 2017; Eric and Manning, 2017; Madotto et al., 2018), since they free the task-oriented dialog systems from the manually designed pipeline modules and can be automatically scaled up to new domains. Recently, sequence-to-sequence (Seq2Seq) models have dominated the study of end-to-end task- oriented dialog systems (Bordes et al., 2017). Different from typical Seq2Seq models for open-domain dialog systems, the successful conversations for task-oriented dialog systems heavily rely on both dialog history and domain-specific knowledge base (KB). To effectively incorporate KB information and perform knowledge-based reasoning, memory augmented models have been proposed (Madotto et al., 2018; Wu et al., 2019), which model the dialog history and the KB knowledge as a bag of words in a flat memory. Despite the remarkable progress of previous studies, current memory based models for multi-turn task- oriented dialog systems still suffer from the following limitations. First, existing methods concatenate dialog utterances of current turn and previous turns as a whole, which ignore previous reasoning process performed by the model and are incapable of dynamically tracking long-term dialog states. These methods introduce much noise since previous utterances as the context is lengthy and redundant (Zhang et al., 2018). Taking the dialog in Table 1 as an example, when answering the user question in 6-th turn, it is difficult for the model to infer that the name of the restaurant is “cocum” from a long concatenated dialog context. Therefore, previous …

Content truncated to 500 words