A Contextual Hierarchical Attention Network with Adaptive Objective for Dialogue State Tracking

This paper was retracted. For more information, see https://aclanthology.org/2020.acl-main.563. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6322–6333 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 6322 A Contextual Hierarchical Attention Network with Adaptive Objective for Dialogue State Tracking Yong Shan12†, Zekang Li12, Jinchao Zhang3, Fandong Meng3, Yang Feng12∗ Cheng Niu3, Jie Zhou3 1 Key Laboratory of Intelligent Information Processing Institute of Computing Technology, Chinese Academy of Sciences (ICT/CAS) 2 University of Chinese Academy of Sciences 3 Pattern Recognition Center, WeChat AI, Tencent Inc, China {shanyong18s,lizekang19g,fengyang}@ict.ac.cn {dayerzhang,fandongmeng,chengniu,withtomzhou}@tencent.com Abstract Recent studies in dialogue state tracking (DST) leverage historical information to deter- mine states which are generally represented as slot-value pairs. However, most of them have limitations to efficiently exploit relevant context due to the lack of a powerful mech- anism for modeling interactions between the slot and the dialogue history. Besides, ex- isting methods usually ignore the slot imbal- ance problem and treat all slots indiscrimi- nately, which limits the learning of hard slots and eventually hurts overall performance. In this paper, we propose to enhance the DST through employing a contextual hierarchical attention network to not only discern relevant information at both word level and turn level but also learn contextual representations. We further propose an adaptive objective to alle- viate the slot imbalance problem by dynami- cally adjust weights of different slots during training. Experimental results show that our approach reaches 52.68% and 58.55% joint ac- curacy on MultiWOZ 2.0 and MultiWOZ 2.1 datasets respectively and achieves new state- of-the-art performance with considerable im- provements (+1.24% and +5.98%). 1 1 Introduction Recently, task-oriented dialogue systems have at- tracted increasing attention in both industry and academia due to their broad application for help- ing users accomplish tasks through spoken interac- tions (Young, 2002; Young et al., 2013; Gao et al., 2019a). Dialogue state tracking (DST) is an essen- tial part of dialogue management in task-oriented dialogue systems. Given current utterances and dialogue history, DST aims to determine the set of †Joint work with Pattern Recognition Center, WeChat AI, Tencent Inc. ∗Yang Feng is the corresponding author. 1Code is available at https://github.com/ictnlp/CHAN-DST User: Hello, I’m looking for a resraurant with fair prices. State: price range=moderate Sys: OK. There are Golden Wok Chinese restaurant and Nirala which serves Indian food, which one do you like? User: Are they both have a reasonable price ? State: price range=moderate Sys: Of course. User: Please tell me the address of Golden Wok. State: price range=moderate; food=chinese Table 1: An example dialogue. At the last turn, it is nec- essary to capture relevant information in dialogue his- tory to correctly predict the value of slot “food”, which is underlined. “User” and “Sys” represent user utter- ance and system response respectively, and the italic text means dialogue states. goals that a user tries to inform at each turn which are represented as slot-value pairs (Williams et al., 2013; Henderson et al., 2014a). As Table 1 shows, the dialogue state is …

Content truncated to 500 words