<?xml version="1.0" standalone="yes"?>
<Paper uid="W06-1410">
  <Title>Algorithms for Generating Referring Expressions: Do They Do What People Do?</Title>
  <Section position="3" start_page="0" end_page="63" type="intro">
    <SectionTitle>
1 Introduction
</SectionTitle>
    <Paragraph position="0"> The generation of referring expressions (henceforth GRE) -- that is, the process of working out what properties of an entity should be used to describe it in such a way as to distinguish it from other entities in the context -- is a recurrent theme in the natural language generation literature. The task is discussed informally in some of the earliest work on NLG (in particular, see (Winograd, 1972; McDonald, 1980; Appelt, 1981)), but the first formally explicit algorithm was introduced in (Dale, 1989); this algorithm, often referred to as the Full Brevity (FB) algorithm, has served as a starting point for many subsequent GRE algorithms. To overcome its limitation to one-place predicates, Dale and Haddock (1991) introduced a constraint-based procedure that could generate referring expressions involving relations; and as a response to the computational complexity of 'greedy' algorithms like FB, Reiter and Dale (Reiter and Dale, 1992; Dale and Reiter, 1995) introduced the psycholinguistically motivated Incremental Algorithm (IA). In recent years there have been a number of important extensions to the IA. The Context-Sensitive extension (Krahmer and Theune, 2002) is able to generate referring expressions for the most salient entity in a context; the Boolean Expressions algorithm (van Deemter, 2002) is able to derive expressions containing boolean operators, as in the cup that does not have a handle; and the Sets algorithm (van Deemter, 2002) extends the basic approach to references to sets, as in the red cups. Some approaches reuse parts of other algorithms: the Branch and Bound algorithm (Krahmer et al., 2003) uses the Full Brevity algorithm, but is able to generate referring expressions with both attributes and relational descriptions using a graph-based technique. There are many other algorithms described in the literature: see, for example, (Horacek, 1997; Bateman, 1999; Stone, 2000; Gardent, 2002). Their general aim is to produce naturalistic referring expressions, often explicitly by means of an attempt to follow the same kinds of principles that we believe people might be following when they produce language -- such as the Gricean maxims (Grice, 1975). However, the algorithms have rarely been tested against real data from human referring expression generation.1 In this paper, we present a data set containing human-produced referring expressions in a limited domain. Focussing specifically on the algorithms  directly concerned with the kinds of properties people select, but with phenomena such as how people group entities together (Funakoshi et al., 2004; Gatt, 2006), or with multi-modal referring expressions where the linguistic part is not necessarily distinguishing by itself (van der Sluis and Krahmer, 2004).</Paragraph>
    <Paragraph position="1">  presented in (Dale, 1989), (Dale and Haddock, 1991) and (Reiter and Dale, 1992), we explore how well these algorithms perform in the same context. There are significant differences between the referring expressions produced by humans, and those produced by the algorithms; we explore these differences and consider what it means for work in the generation of referring expressions.</Paragraph>
    <Paragraph position="2"> The remainder of this paper is structured as follows. In Section 2, we introduce the data set of human-produced referring expressions we use; in Section 3, we introduce the representational framework we use to model the domain underlying this data; in Section 4 we introduce the three algorithms considered in this paper; in Section 5 we discuss the results of using these algorithms on the data that represents the model of our domain; in Section 6 we discuss the differences between the output of the algorithms and the human-produced data; and in Section 7 we draw some conclusions and suggest some steps towards addressing the issues we have identified.</Paragraph>
  </Section>
class="xml-element"></Paper>