Parallel Interactive Networks for Multi-Domain Dialogue State Generation

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 1921–1931, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 1921 Parallel Interactive Networks for Multi-Domain Dialogue State Generation Junfan Chen BDBC and SKLSDE Beihang University, China chenjf@act.buaa.edu.cn Richong Zhang∗ BDBC and SKLSDE Beihang University, China zhangrc@act.buaa.edu.cn Yongyi Mao School of EECS University of Ottawa, Canada ymao@uottawa.ca Jie Xu School of Computing University of Leeds, United Kingdom j.xu@leeds.ac.uk Abstract The dependencies between system and user ut- terances in the same turn and across different turns are not fully considered in existing multi- domain dialogue state tracking (MDST) mod- els. In this study, we argue that the incorpo- ration of these dependencies is crucial for the design of MDST and propose Parallel Inter- active Networks (PIN) to model these depen- dencies. Specifically, we integrate an interac- tive encoder to jointly model the in-turn de- pendencies and cross-turn dependencies. The slot-level context is introduced to extract more expressive features for different slots. And a distributed copy mechanism is utilized to se- lectively copy words from historical system ut- terances or historical user utterances. Empiri- cal studies demonstrated the superiority of the proposed PIN model. 1 Introduction Spoken dialogue system (SDS) is an application that can help users complete their goals efficiently. An SDS usually has a logic engine, called dialogue manager, which involves two main sub-tasks for determining how the system will respond to the users: dialogue state tracking and dialogue pol- icy learning. The task we discuss in this paper is dialogue state tracking, which allows the system maintaining an internal representation of the state of the dialogue as the dialogue progress (Young et al., 2010). Dialogue state tracking involving a single do- main has been extensively studied and achieved much progress. As a more challenging task, Multi- domain dialogue state tracking (MDST) has been introduced in (Ramadan et al., 2018) and attracts ∗Corresponding author much attention in the research community. In- stead of only predicting the (slot, value) pair, in MDST, a model is expected to predict the (domain, slot, value) triplets for each slot in each domain. This task is a great challenge not only because of the large ontology involving 30 slots and exceeding 4500 values (Wu et al., 2019), but also the mixed-domain nature of the dialogues and some complex cases involving cross-turn inference. u1: I want a cheap european restaurant. s1: Can I help you? s2: There is a Curry Garden, and hotel? u2: I need a hotel with free-wifi. s3: There is a Ashley hotel. Anything else? u3: A taxi from the retaurant to the hotel. Figure 1: The dependencies between the system utter- ances and user utterances in a multi-domain dialogue. The red lines imply cross-turn dependencies and the blue lines imply in-turn dependencies. Several models have been proposed for MDST task and proven to be successful (Mrksic et al., 2015; Ramadan et al., 2018; Goel et al., 2019; Eric et al., 2020; Lee et al., 2019; Wu et al., 2019). Among …

Content truncated to 500 words