Generating Relational Descriptions Involving Mutual Disambiguation

This paper discusses the generation of relational referring expressions in which target and landmark descriptions are allowed to help disambiguate each other. Using a corpus of referring expressions in a simple visual domain - in which these descriptions

  • PDF / 280,929 Bytes
  • 11 Pages / 439.363 x 666.131 pts Page_size
  • 32 Downloads / 170 Views

DOWNLOAD

REPORT


bstract. This paper discusses the generation of relational referring expressions in which target and landmark descriptions are allowed to help disambiguate each other. Using a corpus of referring expressions in a simple visual domain - in which these descriptions are likely to occur - we propose a classification approach to decide when to generate them. The classifier is then embedded in a REG algorithm whose results outperform a number of naive baseline systems, suggesting that mutual disambiguation is fairly common in language use, and that this may not be entirely accounted for by existing REG algorithms. Keywords: Natural Language Generation, Relational Referring Expressions, Underspecification.

1

Introduction

In Natural Language Generation (NLG), the computational task of referring expression generation (REG) consists of producing a set of semantic properties to uniquely distinguish an intended referent from other objects in the same context. Consider for instance Figure 1, which illustrates a simple context set containing five objects: a sphere and two cones on the left side, and a cube and a second sphere on the right side.

Fig. 1. A visual domain conveying simple geometric objects

Consider the goal of referring to the target r = Obj2. This may be accomplished, for instance, by making use of atomic properties of r, as in (a-b) below. A. Gelbukh (Ed.): CICLing 2014, Part I, LNCS 8403, pp. 492–502, 2014. c Springer-Verlag Berlin Heidelberg 2014 

Generating Relational Descriptions Involving Mutual Disambiguation

493

(a)The green cone (b)The small green cone Alternatively, we may also make use of relational properties, i.e., we may refer to a second object - hereby called a landmark - as in (c-d) below, and possibly in combination with some atomic properties as well. (c)The cone next to the sphere (d)The small cone on the right side of the large one The generation of relational descriptions as in (c-d) has been the focus of a number of studies in REG [1,2,3,4,5] and, of particular interest to our present discussion, there is the issue of how much information is desirable - or necessary to convey in order to describe each individual objects (i.e., target and landmark). Algorithms such as [1] implicitly assume that target and landmark descriptions are allowed to disambiguate each other, as in previous example (c), in which both ‘cone’ and ‘sphere’ would be ambiguous had we interpreted each description in isolation. This contrasts, for instances, studies such as [5], in which mutual disambiguation is shown to disrupt the search for the target in more complex situations of reference. In this paper we ask when relational descriptions involving mutual disambiguation should be produced as opposed to more (e.g., fully) distinguishing alternatives. To this end, we will focus on the particular case of mutual disambiguation in which the landmark description is left underspecified. Using a corpus of referring expressions in which mutual disambiguation is ubiquitous, we propose a classifier approach to decide when to gener