Towards a Progression-Aware Autonomous Dialogue Agent

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1194 - 1212 July 10-15, 2022 ©2022 Association for Computational Linguistics Towards a Progression-Aware Autonomous Dialogue Agent Abraham Sanders1, Tomek Strzalkowski1, Mei Si1, Albert Chang1, Deepanshu Dey1, Jonas Braasch1, Dakuo Wang2 1Rensselaer Polytechnic Institute, Troy, NY, USA, 2 IBM Research, USA {sandea5,tomek,sim,changa4,deyd,braasj}@rpi.edu dakuo.wang@ibm.com Abstract Recent advances in large-scale language mod- eling and generation have enabled the creation of dialogue agents that exhibit human-like re- sponses in a wide range of conversational sce- narios spanning a diverse set of tasks, from general chit-chat to focused goal-oriented dis- course. While these agents excel at generating high-quality responses that are relevant to prior context, they suffer from a lack of awareness of the overall direction in which the conversation is headed, and the likelihood of task success in- herent therein. Thus, we propose a framework in which dialogue agents can evaluate the pro- gression of a conversation toward or away from desired outcomes, and use this signal to inform planning for subsequent responses. Our frame- work is composed of three key elements: (1) the notion of a "global" dialogue state (GDS) space, (2) a task-specific progression function (PF) computed in terms of a conversation’s trajectory through this space, and (3) a plan- ning mechanism based on dialogue rollouts by which an agent may use progression signals to select its next response. 1 Introduction All human conversation serves some purpose. These may range from negotiating an agreement to explaining a topic to maintaining a social relation- ship. People are generally capable of forming an assessment, sometimes subconsciously, whether a conversation is going well or not and adjusting their behavior accordingly. Such assessment, which un- derlies most human conversation, is essential in continuous awareness of the direction where the interaction is heading and whether the parties are in sync or not, e.g., Bernieri and Rosenthal (1991). In a task-oriented interaction, the participants as- sess if progress towards a successful outcome is being made. In a negotiation, parties assess if an agreement is likely. Even in a casual conversation, people intuitively sense when to continue, when to change the subject, or when to stop. Based on such Figure 1: Our framework applied to the charity solici- tation task in Persuasion For Good (Wang et al., 2019). Given the dialogue history (center left), the system uses rollouts (Lewis et al., 2017) to simulate the outcome of two response candidates (bottom, in red). Each rollout is mapped as a path through the Global Dialogue State space (center right) where it can be compared with sim- ilar outcomes. The candidates are finally ranked using the Progression Function (top), and the best is selected. (subjective) assessment, participants adjust what to say next: whether to push forward, make a concession, soften the tone, digress, or say good- bye. A wide range of research in conversation and discourse analysis is devoted to these and related issues including (Beebe and Masterson, 2000; Cassell …

Content truncated to 500 words