Learning to Plan and Realize Separately for Open-Ended Dialogue Systems

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 2736–2750 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 2736 Learning to Plan and Realize Separately for Open-Ended Dialogue Systems Sashank Santhanam1, Zhuo Cheng1, Brodie Mather2, Bonnie Dorr2, Archna Bhatia2, Bryanna Hebenstreit2, Alan Zemel3, Adam Dalton2, Tomek Strzalkowski4 and Samira Shaikh1 University of North Carolina, Charlotte, NC, USA1 Institute for Human and Machine Cognition (IHMC), Ocala, FL, USA2 State University of New York, Albany, NY, USA3 Rensselaer Polytechnic Institute, Troy, NY, USA4 {ssantha1, samirashaikh}@uncc.edu1 Abstract Achieving true human-like ability to conduct a conversation remains an elusive goal for open-ended dialogue systems. We posit this is because extant approaches towards natural language generation (NLG) are typically con- strued as end-to-end architectures that do not adequately model human generation processes. To investigate, we decouple generation into two separate phases: planning and realization. In the planning phase, we train two planners to generate plans for response utterances. The realization phase uses response plans to pro- duce an appropriate response. Through rigor- ous evaluations, both automated and human, we demonstrate that decoupling the process into planning and realization performs better than an end-to-end approach. 1 Introduction Recent advancements in the area of generative mod- eling have helped increase the fluency of generative models. However, several issues persist: coher- ence of output and the semblance of mere repeti- tion/hallucination of tokens from the training data (Moryossef et al., 2019; Wiseman et al., 2017). One reason could be that the generation task is typically construed as an end-to-end system. This is in contrast to traditional approaches, which in- corporate a sequence of steps in the NLG system, including content determination, sentence planning, and surface realization (Reiter, 1994; Reiter and Dale, 2000). A review of literature from psycholin- guistics and cognitive science also provides strong empirical evidence that the human language pro- duction process is not a monolith (Dell, 1985; Bock, 1996; Bock et al., 2007; Kennison, 2018). Prior approaches have indeed incorporated con- tent planning into the NLG system, for example data-to-text generation problems (Puduppully et al., 2019; Moryossef et al., 2019) as well as classic Figure 1: Example conversation between two speakers A & B where the response for the speaker B is gener- ated based on the response plan from two learned plan- ners: Context Attention and Pseudo Self Attention. works that include planning, based on speech acts (Cohen and Perrault, 1979) (for an in-depth review c.f. (Garoufi, 2014)). Our work closely follows these prior approaches, with one crucial difference: our planners are not based on dialogue acts or speech acts. Consider the example in Fig. 1. An input utter- ance by Person B, a statement (Unfortunately no.), followed by a question (What do they do?), can be effectively responded to using plans, learned and generated, prior to the realization phase. The realization output can then include the mention of provides relief, consistent with the generated plan (PERFORM [provides [relief]]). Dialogue acts (Stolcke et al., 2000) (e.g., state- ments, questions), by their …

Content truncated to 500 words