A computational model of reference
Ori Garin

Abstract:

This paper is an interdisciplinary study toward a theory of
information status and representation of objects. The motivation is a
computational problem: in Natural Language Generation, the task of
Generating Referring Expressions (GRE) tacitly assumes that the
referent is part of the user's focus of attention and must be
distinguished from other entities in that context. The generation of
Hearer-New entities is hopelessly flawed due to a problem which
Kronfeld [Kronfeld, 1990] identifies as the standard-name assumption:
every domain entity is mutually known to all discourse
participants. Under this assumption, referring means finding a
description which identifies the internal presentation of the intended
referent. Hearer-new referents do not fit this model.

Kronfeld addresses the external perspective of reference, which is
crucial for overcoming the standard-name assumption, and provides a
theory of referring in which the basis of a successful referring act
is mutual individuation: the mutual belief of interlocutors that they
are both thinking about the same object. The current paper is focused
on the epistemic aspect of Kronfeld's model, aiming at a realistic
internal representation of domain entities. To this aim, several
epistemic distinctions are made. Finally, applications to GRE and
linguistics are discussed.

Keywords: