LUNA: Learning Slot-Turn Alignment for Dialogue State Tracking

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3319 - 3328 July 10-15, 2022 ©2022 Association for Computational Linguistics LUNA: Learning Slot-Turn Alignment for Dialogue State Tracking Yifan Wang∗, Jing Zhao∗, Junwei Bao†, Chaoqun Duan, Youzheng Wu, Xiaodong He JD AI Research, Beijing, China {wangyifan15,zhaojing857,baojunwei,duanchaoqun1,wuyouzheng1,xiaodong.he}@jd.com Abstract Dialogue state tracking (DST) aims to predict the current dialogue state given the dialogue his- tory. Existing methods generally exploit the ut- terances of all dialogue turns to assign value for each slot. This could lead to suboptimal results due to the information introduced from irrele- vant utterances in the dialogue history, which may be useless and can even cause confusion. To address this problem, we propose LUNA, a SLot-TUrN Alignment enhanced approach. It first explicitly aligns each slot with its most rel- evant utterance, then further predicts the corre- sponding value based on this aligned utterance instead of all dialogue utterances. Furthermore, we design a slot ranking auxiliary task to learn the temporal correlation among slots which could facilitate the alignment. Comprehensive experiments are conducted on multi-domain task-oriented dialogue datasets, i.e., MultiWOZ 2.0, MultiWOZ 2.1, and MultiWOZ 2.2. The results show that LUNA achieves new state-of- the-art results on these datasets.1 1 Introduction Dialogue State Tracking (DST) refers to the task of estimating the dialogue state (i.e., user’s intents) at every dialogue turn, where the state is represented in forms of a set of slot-value pairs (Williams et al., 2016; Eric et al., 2019). DST is crucial to the suc- cess of a task-oriented dialogue system as the dia- logue policy relies on the estimated dialogue state to choose actions. Traditional DST approaches assume that all candidate slot-value pairs are pre- defined in an ontology (Mrkši´c et al., 2017; Zhong et al., 2018; Lee et al., 2019). Then, they scores all possible pairs and selecting the value with the highest score as the predicted value of a slot. ∗Equal contribution. †Corresponding author: baojunwei001@gmail.com 1Our code is available at https://github.com/ nlper27149/LUNA-dst Sys: There are lots to choose from. What type of cuisine are you looking for? User: I do not care. It needs to be on the south side and moderately priced. State: restaurant-area=south; pricerange=moderate Sys: There are 2 options, pizza hut cherry hinton and restaurant alimentum. Can I book you for those ? User: Yes please. I also need a hotel with at least 3 stars and free parking near by the restaurant. State: hotel-parking=yes; hotel-stars =3 Sys: I am sorry, there is no guest house that meets those criteria, either. Would you like to try a different rating, or a different area? User: Sure, what about in the city centre? State: hotel-area =centre; hotel-type=guest house �: hotel-area=south; hotel-type=guest house Table 1: An example of DST. “User" and “Sys" means user query and system response respectively. “State" is the golden label of dialogue state. “�" denotes the predicted states of some existing models and the state marked red is the incorrect …

Content truncated to 500 words