Computational models of referring: A study in cognitive science by Kees van Deemter (review)

In this book, Kees van Deemter provides a wide-ranging synthesis of work from computational linguistics on the natural language generation of referring expressions. vD has been a leading member of a highly productive research community focused on empirical and computational approaches to referring expression generation (REG). Here, he places these efforts into a comprehensive theoretical and historical context, outlining a range of issues and questions important to the development of algorithms for reference generation and providing a guide for where such work is likely to go in the future. In the preface, vD characterizes reference as the ‘fruit fly’ of the study of language, given all the ways that researchers from many disciplines have used reference to explore various phenomena in language and communication. And indeed, this book makes a compelling case for linguistic reference as a central intellectual topic within the cognitive sciences, while at the same time illustrating how computational models, in particular, can be useful tools for exploring an assortment of thorny theoretical issues.

The volume is organized into four major parts. Part I orients the reader to some of the most relevant issues and questions related to the generation of referring expressions, and introduces so-called ‘classic’ approaches to modeling reference generation. ‘Reference’ is a large and diverse research area, so vD takes the opportunity in Ch. 1 to be explicit about his treatment of this topic. For example, the focus is mostly (but not exclusively) on reference production, and especially on algorithms intended to capture how human speakers refer in particular contexts (versus applications of natural language generation that prioritize utility for listeners). There is also a focus on content determination, or the conceptualization of referential expressions at the level of logical form (i.e. semantic specifications like {car, blue}). Other aspects of production, like linguistic realization, are not addressed because they involve tasks (such as specifying word order) not specific to REG. Finally, most of the models presented are concerned with ‘one-shot’ reference, which excludes forms of reference that rely on the discourse context, like demonstrative NPs or pronouns. vD points out that these contextual phenomena often involve language-dependent mechanisms that lie beyond the scope of content determination. While acknowledging the dangers of restricting attention to instances of so-called ‘literary’ reference, in which speakers generate referring expressions intended to uniquely identify referents for attentive addressees, vD argues (and successfully demonstrates throughout the book) that there are still a sufficient number of open issues that justify this focus.

The next two chapters provide important background concerning relevant philosophical, linguistic, and psycholinguistic issues within the study of reference. Ch. 2 provides a useful introduction to fundamental distinctions, such as denotation versus connotation, extensional and intensional contexts, and referential versus attributive descriptions. Generally, the focus is on concepts that have had particular implications for computational models, especially in their application to more complex cases. Then, Ch. 3 reviews key psycholinguistic work, including debates on the role of shared knowledge in the production and comprehension of referring expressions. Particular attention is devoted to evidence concerning possible Gricean constraints on the types of information that speakers mention when referring to objects. Crucially, vD contrasts the notion of discriminatory power, which emphasizes properties that successfully distinguish referents from other objects in the domain, with intrinsic preference, which captures how some object properties, such as color, appear to have psychological priority. In general, vD emphasizes the fact that human speakers are not always strictly rational (in the Gricean sense) and that computational models interested in capturing ‘humanlikeness’ must be explicit about the constraints that determine how REG algorithms consider particular object attributes.

Part II is the heart of the book, focusing on classic computational approaches to REG. Ch. 4 begins by describing early artificial-intelligence work on natural language generation, which often emphasized problems that were difficult to explore systematically. In this context, vD highlights Dale and Reiter’s (1995...