A Sequence-to-Sequence Approach to Dialogue State Tracking

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1714–1725 August 1–6, 2021. ©2021 Association for Computational Linguistics 1714 A Sequence-to-Sequence Approach to Dialogue State Tracking Yue Feng† ∗ Yang Wang‡ Hang Li‡ †University College London, London, UK ‡ByteDance AI Lab, Beijing, China † yue.feng.20@ucl.ac.uk ‡{wangyang.127, lihang.lh}@bytedance.com Abstract This paper is concerned with dialogue state tracking (DST) in a task-oriented dialogue sys- tem. Building a DST module that is highly effective is still a challenging issue, although significant progresses have been made recently. This paper proposes a new approach to dia- logue state tracking, referred to as Seq2Seq- DU, which formalizes DST as a sequence-to- sequence problem. Seq2Seq-DU employs two BERT-based encoders to respectively encode the utterances in the dialogue and the descrip- tions of schemas, an attender to calculate atten- tions between the utterance embeddings and the schema embeddings, and a decoder to gen- erate pointers to represent the current state of dialogue. Seq2Seq-DU has the following ad- vantages. It can jointly model intents, slots, and slot values; it can leverage the rich rep- resentations of utterances and schemas based on BERT; it can effectively deal with cate- gorical and non-categorical slots, and unseen schemas. In addition, Seq2Seq-DU can also be used in the NLU (natural language understand- ing) module of a dialogue system. Experimen- tal results on benchmark datasets in different settings (SGD, MultiWOZ2.2, MultiWOZ2.1, WOZ2.0, DSTC2, M2M, SNIPS, and ATIS) show that Seq2Seq-DU outperforms the exist- ing methods. 1 Introduction A task-oriented dialogue system usually consists of several modules: natural language understand- ing (NLU), dialogue state tracking (DST), dialogue policy (Policy), and natural language generation (NLG). We consider DST and also NLU in this paper. In NLU, a semantic frame representing the content of user utterance is created in each turn ∗The work was done when the first author was an intern at ByteDance AI Lab. <start>, “FindFlight”, <sep>, “Depart”, #4, #4, <sep>, “Arrive”, #6, #7,<sep>, “Seating_class”, #Economy, <sep>, “Depart_Date”, #17,#18, <end> <start>, “FindFlight”, <sep>, “Depart”, #4, #4, <sep>, “Arrive”, #6, #7, <sep>, “Seating_Class”, #Economy, <end> Slots: “Depart”: Starting city for the trip. “Arrive”: Ending city for the trip. “Seating_Class”: Seating class for the booking. Possible_Values: ["Economy", "Business", "First Class”] “Depart_Date”: Start date for the trip. Intents: “FindFlight”: Search for one-way flights to a destination. “ReserveFlight”: Reserve a one-way flight. Find economy flights from Beijing to Los Angeles. Sure, what dates are you looking for? Flying on May 2. Ok, I found a Delta flight for 3500 dollars. User System State Service: “Flight”: Find your next flight. Schema Figure 1: An example of dialogue state tracking. Given a dialogue history that contains user utterances and sys- tem utterances, and descriptions of schema that contain all possible intents and slot-value pairs, a dialogue state for the current turn is created which is represented by intents and slot-value pairs. There are slot values ob- tained from the schema (categorical) as well as slot val- ues …

Content truncated to 500 words