Generation and Extraction Combined Dialogue State Tracking with Hierarchical Ontology Integration

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2241–2249 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2241 Generation and Extraction Combined Dialogue State Tracking with Hierarchical Ontology Integration Xinmeng Li and Qian Li and Wansen Wu and Quanjun Yin College of Systems Engineering, National University of Defense Technology xml.nudt@gmail.com, liqian9510@outlook.com, {wuwansen14, yinquanjun}@nudt.edu.cn Abstract Recently, the focus of dialogue state tracking has expanded from single domain to multi- ple domains. The task is characterized by the shared slots between domains. As the scenari- o gets more complex, the out-of-vocabulary problem also becomes more severe. Curren- t models are not satisfactory for addressing the challenges of ontology integration between domains and out-of-vocabulary problems. To address the problem, we explore the hierarchi- cal semantics of the ontology and enhance the interrelation between slots with masked hierar- chical attention. In state value decoding stage, we address the out-of-vocabulary problem by combining generation method and extraction method together. We evaluate the performance of our model on two representative datasets, MultiWOZ in English and CrossWOZ in Chi- nese. The results show that our model yield- s a significant performance gain over curren- t state-of-the-art state tracking model and it is more robust to out-of-vocabulary problem compared with other methods. 1 Introduction Dialogue state tracking (DST) is in charge of up- dating the belief state in task-oriented dialogue system (Gao et al., 2019a). Traditional discrim- inative DST models assume that the task ontolo- gy is well defined in advance, that is to say, al- l states and their values are known to the model. They usually rely on hand-crafted features or task- specific lexicon (Henderson et al., 2014; Mrkši´c and Vuli´c, 2018). An inconvenience is that they are time-consuming and hard to expand to new tasks. To overcome it, the open vocabulary-based models are proposed to decode the state value ac- cording to the dialogue context (Xu and Hu, 2018; Jin et al., 2018; Lei et al., 2018; Goel et al., 2019). In recent years, the research frontier for task- oriented dialogue systems has expanded from s- ingle domain to multiple domains (Budzianowski Figure 1: The hierarchically represented ontology for MultiWOZ dataset. Each of the states can be repre- sented by an arrow line connecting through Domain, Intent and Slot. et al., 2018; Zhu et al., 2020; Cheng et al., 2020). There come new challenges demanding promp- t solution. Firstly, current model does not suffi- ciently consider the interrelation between slots in multi-domain scenario. For example, user asks “I also want to find an attraction near the restauran- t”, which implies that the hotel need to have the same area with the restaurant. The implicit rela- tion between the area slot of hotel and restaurant is the key to exactly track user’s intent (Wu et al., 2019). Prior work simply used the summed (Wu et al., 2019) or concatenated (Zhang et al., 2019; Kim et al., 2020) embedding of the domain and slot as the states representation for the decoder. Secondly, …

Content truncated to 500 words