Introduction

Buying new chairs can be complicated. Many constraints have to be kept in mind, including your financial situation, the style and color of the furniture you already own and possibly also the taste of your partner. But once you have made a tentative choice (say, the chair in Figure 6.1 on the left), there is one final hurdle: you have to inform the seller of your desire to buy it. Furniture stores tend to contain many chairs, so somehow you need to refer to your chair of choice, for example I'd like to buy the wooden chair with the thin legs and solid seat, the red one with the open back. It is hardly surprising that many people in this situation resort to pointing (that one). Of course, it would be helpful to know that salespeople might refer to this chair as the red Keystone chair, because that would allow you to adapt to their way of referring.

This problem illustrates the importance of reference in everyday interactions: people can only exchange information about objects when they agree on how to refer to those objects. How this agreement may arise, and how we can model this in natural language generation, is the topic of this chapter. We argue that two possibly competing forces play a role. On the one hand, speakers may have inherent preferences for certain properties when referring to objects in a given domain. On the other, they may also have a tendency to adapt to the references produced by their dialogue partner. We describe how preferences can be determined, and how they interact with adaptation. We model this tradeoff using a graph-based referring expression generation algorithm (Krahmer et al., 2003).