The Impact of Common Ground on Referring Expressions in Human-Robot interaction

In a world where conversational Artificial Intelligence (AI) is playing an increasingly large role in our daily lives, it is important to understand how conversations between humans and AI or robots can become more personal, and stay useful over a prolonged period of time. The way two humans talk about other people is shaped by their degree of shared knowledge, or common ground. These referring expressions tend to become shorter and more personal over time, as codified, shared ways of making reference called conventions start to form. This process could make conversations with AI such as robots more personal, but there has been little research on this topic. In this dissertation, I examine how common ground plays in a role in making reference in this Human-Robot Interaction (HRI), addressing the research gap that exists in this space, and providing insights into both human-human and human-robot use of referring expressions in longitudinal interactions. My investigation has three main goals: examining how useful data can be obtained to study referring expressions in HRI; examining how referring expressions in HRI are influenced by the common ground, conventions, and contextual factors such as ambiguity; and examining the role of the robot in shaping, using and understanding referring expressions and conventions. I carry out this investigation through a number of experiments using data sets, field experiments with a physical robot, and online experiments with a chatbot. My approach draws from fields in linguistics related to (human-human) common ground and pragmatics, as well as fields in Natural Language Processing related to Natural Language Understanding and Referring Expression Generation. An important element in my research is the introduction of an inner circle of familiar entities, for which common ground is built up, and an outer circle of unfamiliar entities. By comparing the use of referring expressions for the inner and outer circle, I create a contrastive analysis which allows me to draw conclusions about the impact of common ground on referring expressions. This contrast is implemented in an experimental framework called SPOTTER, a game in which referring expressions to the inner and outer circle are elicited. I use this framework to analyse the use of referring expressions and the role of the robot in shaping them. One of the main findings of this dissertation is that, although present, the role of common ground is less pronounced in human-robot reference-making compared to human-human interactions. However, I do find a large influence of the contextual setting within each round of the SPOTTER game, as shaped by the outer circle. This influence is shown by longer, more detailed referring expressions when ambiguity is high, as well as through longer referring expressions for the outer circle compared to the inner circle. The findings also suggest that the robot's use of referring expressions has a large impact on which referring expressions humans use. However, though both humans and robots form conventions, they rarely truly converge on a shared convention. Naturally, the findings are influenced by design choices for the experimental framework and the conversational model, as well as by the complicated nature of field studies in HRI. The dissertation lays the groundwork for further investigations of reference-making and common ground in HRI, providing experimental tools and models.