Training Neural Response Selection for Task-Oriented Dialogue Systems

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5392–5404 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5392 Training Neural Response Selection for Task-Oriented Dialogue Systems Matthew Henderson, Ivan Vuli´c, Daniela Gerz, Iñigo Casanueva, Paweł Budzianowski, Sam Coope, Georgios Spithourakis, Tsung-Hsien Wen, Nikola Mrkši´c, and Pei-Hao Su PolyAI Limited London, United Kingdom matt@poly-ai.com ivan@poly-ai.com Abstract Despite their popularity in the chatbot liter- ature, retrieval-based models have had mod- est impact on task-oriented dialogue systems, with the main obstacle to their application be- ing the low-data regime of most task-oriented dialogue tasks. Inspired by the recent suc- cess of pretraining in language modelling, we propose an effective method for deploying response selection in task-oriented dialogue. To train response selection models for task- oriented dialogue tasks, we propose a novel method which: 1) pretrains the response selec- tion model on large general-domain conversa- tional corpora; and then 2) fine-tunes the pre- trained model for the target dialogue domain, relying only on the small in-domain dataset to capture the nuances of the given dialogue domain. Our evaluation on six diverse appli- cation domains, ranging from e-commerce to banking, demonstrates the effectiveness of the proposed training method. 1 Introduction Retrieval-based dialogue systems conduct conver- sations by selecting the most appropriate system response given the dialogue history and the input user utterance (i.e., the full dialogue context). A typical retrieval-based approach to dialogue en- codes the input and a large set of responses in a joint semantic space. When framed as an ad-hoc re- trieval task (Deerwester et al., 1990; Ji et al., 2014; Kannan et al., 2016; Henderson et al., 2017), the system treats each input utterance as a query and retrieves the most relevant response from a large re- sponse collection by computing semantic similarity between the query representation and the encod- ing of each response in the collection. This task is referred to as response selection (Wang et al., 2013; Al-Rfou et al., 2016; Yang et al., 2018; Du and Black, 2018; Weston et al., 2018; Chaudhuri et al., 2018), as illustrated in Figure 1. Input Candidate Responses Is that place affordable? Absolutely, call me any time! There is no place like home. The restaurant serves Japanese food. I would say that the prices are reasonable. This was their second warning. It was so unfortunate to concede the goal. Figure 1: The conversational response selection task: given the input sentence, the goal is to identify the rele- vant response from a large collection of candidates. Formulating dialogue as a response selection task stands in contrast with other data-driven di- alogue modeling paradigms such as modular and end-to-end task-based dialogue systems (Young, 2010; Wen et al., 2017b; Liu and Perez, 2017; Li et al., 2017; Bordes et al., 2017). Unlike stan- dard task-based systems, response selection does not rely on explicit task-tailored semantics in the form of domain ontologies, which are hand-crafted for each task by domain experts (Henderson et al., 2014a,b; Mrkši´c …

Content truncated to 500 words