GCDST: A Graph-based and Copy-augmented Multi-domain Dialogue State Tracking

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 1063–1073 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 1063 GCDST: A Graph-based and Copy-augmented Multi-domain Dialogue State Tracking Peng Wu, Bowei Zou, Ridong Jiang, Ai Ti Aw Aural & Language Intelligence Department, Institute for Infocomm Research (I2R), A∗STAR, Singapore wudapeng268@gmail.com, {zou bowei,rjiang,aaiti}@i2r.a-star.edu.sg Abstract As an essential component of task-oriented dialogue systems, Dialogue State Tracking (DST) takes charge of estimating user inten- tions and requests in dialogue contexts and ex- tracting substantial goals (states) from user ut- terances to help the downstream modules to de- termine the next actions of dialogue systems. For practical usages, a major challenge to con- structing a robust DST model is to process a conversation with multi-domain states. How- ever, most existing approaches trained DST on a single domain independently, ignoring the information across domains. To tackle the multi-domain DST task, we first construct a dialogue state graph to transfer structured fea- tures among related domain-slot pairs across domains. Then, we encode the graph infor- mation of dialogue states by graph convolu- tional networks and utilize a hard copy mech- anism to directly copy historical states from the previous conversation. Experimental re- sults show that our model improves the per- formances of the multi-domain DST baseline (TRADE) with the absolute joint accuracy of 2.0% and 1.0% on the MultiWOZ 2.0 and 2.1 dialogue datasets, respectively. 1 Introduction A task-oriented dialogue system provides funda- mental technologies for continuous interactions with a human to accomplish predefined specific goals, such as taxi reservation or hotel booking. Dialogue State Tracking (DST) is a crucial compo- nent in the task-oriented dialogue system. Users’ intentions and goals are extracted from the current utterances and the conversation history. Then, the DST model encodes the information as a set of states to help dialogue systems to determine which actions should be taken in next steps (Young and Thomson, 2013). A dialogue state generally comprises an entity Usr: I’m looking for an expensive restaurant in the centre of town. Sys: What about the Cambridge chop house? A British restaurant located in the centre of town. Usr: I need to book a table for four people at 16:45 on Friday. Sys: Booking was successful. Reference number is 10p0levh. Anything else today? Usr: I’m also looking for a place to stay. Ideally a hotel with free WIFI that is also expensive. Sys: There is the Gonville hotel. It has internet and is rated 3 stars. Would you like to book? Usr: Great, can you book it for two people, for four nights starting Friday? Sys: Your booking was successful starting Friday for four nights. Reference number is drw9. Usr: I also need a taxi to this hotel and leave at 21:45. Sys: Ok, what will your departure be? Usr: St. John’s College Sys: Okay, I have booked a taxi for you, the number is 07240037071. (price-range, expensive) Domain: Restaurant Domain: Hotel (price-range, expensive) (area, centre) (book-day, Friday) (book-day, Friday) (book-people, four) (book-people, two) (name, …

Content truncated to 500 words