Reflect, Not Reflex: Inference-Based Common Ground Improves Dialogue Response Quality

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 10450 - 10468 December 7-11, 2022 ©2022 Association for Computational Linguistics Reflect, Not Reflex: Inference-Based Common Ground Improves Dialogue Response Quality Pei Zhou Hyundong Cho Pegah Jandaghi Dong-Ho Lee Bill Yuchen Lin Jay Pujara Xiang Ren Department of Computer Science and Information Sciences Institute University of Southern California {peiz,dongho.lee,yuchen.lin,jpujara,xiangren}@usc.edu, {jcho,jandaghi}@isi.edu Abstract Human communication relies on common ground (CG), the mutual knowledge and be- liefs shared by participants, to produce coher- ent and interesting conversations. In this paper, we demonstrate that current response gener- ation (RG) models produce generic and dull responses in dialogues because they act reflex- ively, failing to explicitly model CG, both due to the lack of CG in training data and the stan- dard RG training procedure. We introduce Re- flect, a dataset that annotates dialogues with explicit CG (materialized as inferences approxi- mating shared knowledge and beliefs) and solic- its 9k diverse human-generated responses each following one common ground. Using Reflect, we showcase the limitations of current dialogue data and RG models: less than half of the re- sponses in current data is rated as high quality (sensible, specific, and interesting) and models trained using this data have even lower qual- ity, while most Reflect responses are judged high quality. Next, we analyze whether CG can help models produce better quality responses by using Reflect CG to guide RG models. Sur- prisingly, we find that simply prompting GPT3 to “think” about CG generates 30% more qual- ity responses, showing promising benefits to integrating CG into the RG process.1 1 Introduction Human communication is a collaborative ef- fort (Grice, 1975; Allwood, 1976; Bohm et al., 2004) where participants strive to achieve com- mon ground (CG), consisting of mutual beliefs and common knowledge (Stalnaker, 1978; Clark and Schaefer, 1989; Clark and Brennan, 1991). Conver- sational AI systems, while able to produce fluent texts, often generate generic and dull dialogue re- sponses (Serban et al., 2017; Zhao et al., 2017), potentially because they do not explicitly model 1Link to our data and code will be provided on our project page https://inklab.usc.edu/Reflect/. Oh no, I spilled the food I prepared for dinner They might be feeling bad and need help cleaning it up Sorry to hear that! (okay.. they don’t seem so interested ) Don’t worry! How about let’s clean it up and order from your favorite pasta place? Aww, thank you! I feel much better now! Respondin� “reflexively” produces dull and �eneric conversations Makin� in�erences to reach common �round produces more meanin��ul and en�a�in� conversations! Figure 1: A motivating example. We aim to help RG models produce more human-like responses instead of generic ones. We argue that integrating common ground by making inferences is crucial. CG in communication (as illustrated in Figure 1). Specifically, existing models mostly follow a dia- logue history → response training paradigm since such data can be easily obtained in the wild, skip- ping an important middle step that builds common ground, …

Content truncated to 500 words