Beyond the Granularity: Multi-Perspective Dialogue Collaborative Selection for Dialogue State Tracking

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2320 - 2332 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Beyond the Granularity: Multi-Perspective Dialogue Collaborative Selection for Dialogue State Tracking Jinyu Guo1, Kai Shuang1∗, Jijie Li1, Zihan Wang2 and Yixuan Liu1 1State Key Laboratory of Networking and Switching Technology, Beijing University of Posts and Telecommunications 2Graduate School of Information Science and Technology, The University of Tokyo {guojinyu, shuangk, lijijie, liuyixuan}@bupt.edu.cn zwang@tkl.iis.u-tokyo.ac.jp Abstract In dialogue state tracking, dialogue history is a crucial material, and its utilization varies be- tween different models. However, no matter how the dialogue history is used, each existing model uses its own consistent dialogue history during the entire state tracking process, regard- less of which slot is updated. Apparently, it requires different dialogue history to update different slots in different turns. Therefore, us- ing consistent dialogue contents may lead to insufficient or redundant information for dif- ferent slots, which affects the overall perfor- mance. To address this problem, we devise DiCoS-DST to dynamically select the relevant dialogue contents corresponding to each slot for state updating. Specifically, it first retrieves turn-level utterances of dialogue history and evaluates their relevance to the slot from a com- bination of three perspectives: (1) its explicit connection to the slot name; (2) its relevance to the current turn dialogue; (3) Implicit Mention Oriented Reasoning. Then these perspectives are combined to yield a decision, and only the selected dialogue contents are fed into State Generator, which explicitly minimizes the dis- tracting information passed to the downstream state prediction. Experimental results show that our approach achieves new state-of-the-art per- formance on MultiWOZ 2.1 and MultiWOZ 2.2, and achieves superior performance on mul- tiple mainstream benchmark datasets (includ- ing Sim-M, Sim-R, and DSTC2).1 1 Introduction Task-oriented dialogue systems have recently at- tracted growing attention and achieved substan- tial progress. Dialogue state tracking (DST) is a core component, where it is responsible for interpreting user goals and intents and feeding ∗Corresponding author. 1Code is available at https://github.com/guojinyu88/DiCoS-master S1:Good morning! How can I help you? S2:The alpha-milton guest house is in the north and moderately priced. Would you like to book a stay? U2:I need something cheaply priced. U1:I'm looking to stay at a guesthouse while I'm in town. I don't need internet access, so don't worry about that. S3:We have 9 guesthouses that match your search. Would you like to narrow it down? S4:May i suggest the Worth House? It is a cheap, 4 star hotel in northern Cambridge. U4: How long can I book it from Monday for 7 people? U3:I don't care as long as it's a guesthouse located in the north for cheap. S5: You can book it for 3 nights. S6:Your booking for 3 nights was a success! Is there anything else I can help you with? U6:Thank you. Are there any cheap restaurants near the hotel as well? U5: That's fine. Please book it for me. S7:Just to clarify, are you looking …

Content truncated to 500 words