Learning Dialog Policies from Weak Demonstrations

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 122–132 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 122 Zero-Shot Transfer Learning with Synthesized Data for Multi-Domain Dialogue State Tracking Giovanni Campagna Agata Foryciarz Mehrad Moradshahi Monica S. Lam Computer Science Department Stanford University Stanford, CA, USA {gcampagn,agataf,mehrad,lam}@cs.stanford.edu Abstract Zero-shot transfer learning for multi-domain dialogue state tracking can allow us to handle new domains without incurring the high cost of data acquisition. This paper proposes new zero-short transfer learning technique for dia- logue state tracking where the in-domain train- ing data are all synthesized from an abstract di- alogue model and the ontology of the domain. We show that data augmentation through syn- thesized data can improve the accuracy of zero-shot learning for both the TRADE model and the BERT-based SUMBT model on the MultiWOZ 2.1 dataset. We show training with only synthesized in-domain data on the SUMBT model can reach about 2/3 of the ac- curacy obtained with the full training dataset. We improve the zero-shot learning state of the art on average across domains by 21%. 1 Introduction Automated conversational agents can reduce the costs of customer support, a necessary service in just about every business. However, training a goal-directed dialogue agent for a domain often requires acquiring annotated dialogues to cover all possible conversation flows. Commonly, this is done using the Wizard-of-Oz technique (Kel- ley, 1984), where two crowdsource workers con- verse with each other, while also annotating the state at each turn. This technique has been em- ployed to construct several datasets (Hemphill et al., 1990; Wen et al., 2017; Yu et al., 2019). Recently, it has been used to build the MultiWOZ dataset (Budzianowski et al., 2018), a large corpus of dialogues across 7 domains. Unfortunately, not only is the initial acquisi- tion expensive, annotating dialogues correctly has proven to be challenging due to human errors, de- lays in annotation, inconsistent conventions, and normalization issues (Eric et al., 2019; Zhou and S: START U: Can you help with information regarding a food place? I need to book at 15:45. S: SEARCHREQUEST restaurant(book time = “15:45”) A: How about the restaurant with name La Tasca and Italian food? U: Can you find something which serves seafood? S: SEARCHREQUEST restaurant(book time = “15:45”, food = “seafood”) A: What date are you looking for? U: Thursday please. S: SEARCHREQUEST restaurant(book time = “15:45”, food = “seafood”, book day = “thursday”) A: How about the Copper Kettle? It is a food place with seafood food. U: What is the price range and the area? S: SLOTQUESTION restaurant(book time = “15:45”, food = “seafood”, book day = “thursday”, price range =?, area =?) A: The Copper Kettle is a moderately priced restaurant in the north of the city. Would you like a reservation? U: No, thanks. S: CLOSE restaurant(book time = “15:45”, food = “seafood”, book day = “thursday”) A: Can I help with you anything else? U: Thank you, that will be it …

Content truncated to 500 words