Self-Supervised Dialogue Learning

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 3857–3867 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 3857 Self-Supervised Dialogue Learning Jiawei Wu and Xin Wang and William Yang Wang Department of Computer Science University of California, Santa Barbara Santa Barbara, CA 93106 USA {jiawei wu,xwang,william}@cs.ucsb.edu Abstract The sequential order of utterances is often meaningful in coherent dialogues, and the or- der changes of utterances could lead to low- quality and incoherent conversations. We con- sider the order information as a crucial su- pervised signal for dialogue learning, which, however, has been neglected by many previ- ous dialogue systems. Therefore, in this pa- per, we introduce a self-supervised learning task, inconsistent order detection, to explic- itly capture the flow of conversation in dia- logues. Given a sampled utterance pair triple, the task is to predict whether it is ordered or misordered. Then we propose a sampling- based self-supervised network SSN to per- form the prediction with sampled triple ref- erences from previous dialogue history. Fur- thermore, we design a joint learning frame- work where SSN can guide the dialogue sys- tems towards more coherent and relevant di- alogue learning through adversarial training. We demonstrate that the proposed methods can be applied to both open-domain and task- oriented dialogue scenarios, and achieve the new state-of-the-art performance on the Open- Subtitiles and Movie-Ticket Booking datasets. 1 Introduction In recent years, dialogue systems have achieved fruitful results with neural conversation models in both open-domain generation (Ritter et al., 2011; Sordoni et al., 2015b; Li et al., 2016b, 2017; Xu et al., 2017; Zhang et al., 2018b) and task-oriented completion (Wen et al., 2015, 2017; Williams et al., 2017; Bordes et al., 2017; Su et al., 2018). These methods empower lots of real-world dia- logue applications such as Google Home and Ap- ple Siri. However, the utterance generation from di- alogue systems still faces some critical chal- lenges, including utterance blandness and incoher- ence (Gao et al., 2018). They are mainly caused by the objective function of the dialogue systems that prefer utterances with unconditionally high probability (Li et al., 2016a). We argue that in a meaningful and coherent dialogue, the change of utterance order will lead to a low-quality dia- logue. However, most existing neural-based dia- logue systems either encode the full dialogue his- tory (Li et al., 2017; Xu et al., 2017) or only the current utterance (Liu and Lane, 2018). None of them explicitly models the sequential order and studies its criticality to the dialogue learning prob- lem. In this paper, we explore the sequential order within the dialogue as the self-supervised signal to guide meaningful and coherent dialogue learning. We introduce a self-supervised learning task, in- consistent order detection, to explicitly capture the order signal of the dialogue. The task is defined as, given a target utterance pair triple, the model is required to predict whether the triple is cor- rectly ordered or not. For instance, the utterance pair triple …

Content truncated to 500 words