Data-Efficient Goal-Oriented Conversation with Dialogue Knowledge Transfer Networks

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 1741–1751, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 1741 Data-Efficient Goal-Oriented Conversation with Dialogue Knowledge Transfer Networks Igor Shalyminov†, Sungjin Lee‡, Arash Eshghi†, and Oliver Lemon† †Heriot-Watt University, UK ‡Amazon Alexa AI, USA †{is33, a.eshghi, o.lemon}@hw.ac.uk, ‡sungjinl@amazon.com Abstract Goal-oriented dialogue systems are now be- ing widely adopted in industry where it is of key importance to maintain a rapid prototyp- ing cycle for new products and domains. Data- driven dialogue system development has to be adapted to meet this requirement — therefore, reducing the amount of data and annotations necessary for training such systems is a cen- tral research problem. In this paper, we present the Dialogue Knowl- edge Transfer Network (DiKTNet), a state- of-the-art approach to goal-oriented dialogue generation which only uses a few example dialogues (i.e. few-shot learning), none of which has to be annotated. We achieve this by performing a 2-stage training. Firstly, we perform unsupervised dialogue representation pre-training on a large source of goal-oriented dialogues in multiple domains, the MetaLWOz corpus. Secondly, at the transfer stage, we train DiKTNet using this representation to- gether with 2 other textual knowledge sources with different levels of generality: ELMo en- coder and the main dataset’s source domains. Our main dataset is the Stanford Multi- Domain dialogue corpus. We evaluate our model on it in terms of BLEU and Entity F1 scores, and show that our approach signifi- cantly and consistently improves upon a se- ries of baseline models as well as over the previous state-of-the-art dialogue generation model, ZSDG. The improvement upon the lat- ter — up to 10% in Entity F1 and the average of 3% in BLEU score — is achieved using only 10% equivalent of ZSDG’s in-domain training data. 1 Introduction Machine learning-based dialogue systems, while still being a relatively new research direction, are experiencing increasingly wide adoption in indus- try. Large-scale dialogue assistant platforms such as Google Assistant, Amazon Alexa, and Apple Siri provide a unified conversational user inter- face (CUI) for third-party applications and ser- vices. Furthermore, products like Google Di- alogflow, Wit.ai, Microsoft LUIS, and Rasa offer means for rapid development of a dialogue sys- tem’s core modules. In addition, with the re- cently adopted technique of training dialogue sys- tems end-to-end data-efficiency of such systems becomes the key question in their adoption in practical applications. Currently, while being ex- tremely flexible and requiring little to no program- ming of in-domain business logic (see e.g. Ultes et al. (2018); Wen et al. (2017); Rojas-Barahona et al. (2017)), such systems have too high data consumption — including both collection and an- notation effort — in order for them to be used in rapidly paced industrial product cycles. There- fore, approaches to training such systems with ex- tremely limited data (i.e. zero-, one- and few-shot training) are a priority research direction in the di- alogue systems area. In …

Content truncated to 500 words