Generalizable and Explainable Dialogue Generation via Explicit Action Learning

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3981–3991 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3981 Generalizable and Explainable Dialogue Generation via Explicit Action Learning Xinting Huang,1 Jianzhong Qi,1 Yu Sun,2 Rui Zhang1∗ 1The University of Melbourne, 2Twitter Inc. {xintingh@student., jianzhong.qi@, rui.zhang@}unimelb.edu.au, ysun@twitter.com Abstract Response generation for task-oriented dia- logues implicitly optimizes two objectives at the same time: task completion and lan- guage quality. Conditioned response genera- tion serves as an effective approach to sepa- rately and better optimize these two objectives. Such an approach relies on system action an- notations which are expensive to obtain. To alleviate the need of action annotations, latent action learning is introduced to map each utter- ance to a latent representation. However, this approach is prone to over-dependence on the training data, and the generalization capability is thus restricted. To address this issue, we pro- pose to learn natural language actions that rep- resent utterances as a span of words. This ex- plicit action representation promotes general- ization via the compositional structure of lan- guage. It also enables an explainable genera- tion process. Our proposed unsupervised ap- proach learns a memory component to sum- marize system utterances into a short span of words. To further promote a compact action representation, we propose an auxiliary task that restores state annotations as the summa- rized dialogue context using the memory com- ponent. Our proposed approach outperforms latent action baselines on MultiWOZ, a bench- mark multi-domain dataset. 1 Introduction Task-oriented dialogue systems complete tasks for users, such as making a hotel reservation or find- ing train routes, in a multi-turn conversation (Gao et al., 2018; Sun et al., 2016, 2017). The gener- ated system utterances should not only be naturally sound, but more importantly be informative, i.e., to proceed the dialogue towards task completion. To fulfill this requirement, conditioned response gen- eration is widely adopted based on system actions ∗Rui Zhang is the corresponding author. Table 1: Conditioned Response Generation Example Dialog Context Utterance User: I need a train that departs bishops stortford on wednesday, please. Dialogue state annotation Train:{departure=bishops stortford, day=wednesday} Action Selection System action annotation train-inform:{choice=five}; train-request:{leaveat} Latent action Natural language action [0,0,0,1,0] {‘option’, ‘five’ ‘request’,‘time’ } Response Generation alright I found five options available. when would you like to leave by? (Wen et al., 2017; Chen et al., 2019). The response generation process is decoupled into two consecu- tive steps, where an action is first selected and then an utterance is generated conditioned on this action. One can optimize each step towards its goal, i.e., in- formative and naturally sound, without impinging the other (Yarats and Lewis, 2018). However, such approaches rely on action annotations (as in Table 1), which require domain knowledge and extensive efforts to obtain. To deal with the absence of action annotations, latent action learning has been introduced (Zhao et al., 2018; Yarats and Lewis, 2018). System ut- terances are represented as low-dimensional latent variables by an auto-encoding task (Zhao et al., 2019), and …

Content truncated to 500 words