Strategy-level Entrainment of Dialogue System Users in a Creative Visual Reference Resolution Task

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5768–5777 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5768 Strategy-level Entrainment of Dialogue System Users in a Creative Visual Reference Resolution Task Deepthi Karkada1, Ramesh Manuvinakurike2, Maike Paetzel-Pr¨usmann3, Kallirroi Georgila4 1Intel Corp, USA, 2Intel Labs, USA, 3University of Potsdam, Germany, 4University of Southern California, USA {deepthi.karkada, ramesh.manuvinakurike}@intel.com, paetzel-pruesmann@uni-potsdam.de, kgeorgila@ict.usc.edu Abstract In this work, we study entrainment of users playing a creative reference resolution game with an autonomous dialogue system. The language understanding module in our dialogue system leverages annotated human-wizard conversational data, openly available knowledge graphs, and crowd-augmented data. Unlike previous entrainment work, our dialogue system does not attempt to make the human conversation partner adopt lexical items in their dialogue, but rather to adapt their descriptive strategy to one that is simpler to parse for our natural language understanding unit. By deploying this dialogue system through a crowd-sourced study, we show that users indeed entrain on a “strategy-level” without the change of strategy impinging on their creativity. Our work thus presents a promising future research direction for developing dialogue management systems that can strategically influence people’s descriptive strategy to ease the system’s language understanding in creative tasks. Keywords: entrainment, dialogue systems, knowledge graphs 1. Introduction Linguistic entrainment (also known as alignment, coor- dination, priming, convergence, accommodation, and adaptation) is the phenomenon in which interlocutors start speaking more similarly to each other (Brennan and Clark, 1996; Rahimi et al., 2017). Entrainment can happen on multiple levels such as prosody, speech rate, vocabulary usage, and syntactic and stylistic patterns (Iio et al., 2009; Levitan and Hirschberg, 2011; Levi- tan, 2013). Entrainment has been associated with di- alogue and task success in human-human (Reitter and Moore, 2007; Friedberg et al., 2012) and human-agent (Fandrianto and Eskenazi, 2012; Lopes et al., 2013) interaction. So far entrainment has been studied in relatively constrained domains with a limited vocab- ulary size. However, in more complex dialogue sys- tem (DS) applications, where a large vocabulary size and variety or unpredictability of user responses would be challenging for automatic speech recognition (ASR) and natural language understanding (NLU), entrain- ment can potentially be even more beneficial. In this paper, we focus on entrainment in the context of a pedagogical game called RDG-Map (Paetzel et al., 2020). RDG-Map is a collaborative game between two players, a Director and a Matcher, engaging in a goal-oriented conversation to identify a country on the world map. The Director gives descriptions for a country so that the Matcher can locate and select it on their map. The game is time-constrained to gener- ate natural and fluid conversations similar to everyday dialogue. While directional descriptions (above USA, southwest of Zambia, etc.) and referring to the coun- tries by name are the most unambiguous for a given tar- get country, the knowledge of the average player about the location of countries on the map is very limited, which makes references to neighbouring countries dif- ficult to …

Content truncated to 500 words