The Question of Expressiveness in the Generation of Referring Expressions

TL;DR

This paper explores the limits of expressiveness in generating referring expressions, introducing a formal framework, analyzing complexity, and proposing methods to combine different approaches for improved generation.

Contribution

It formalizes the concept of $"+L"$-referring expressions, analyzes complexity bounds, and discusses combining different generation frameworks.

Findings

Defined $"+L"$-referring expressions with formal semantics

Provided new complexity bounds for generation algorithms

Discussed combining different expressive frameworks

Abstract

We study the problem of generating referring expressions modulo different notions of expressive power. We define the notion of $\+L$-referring expression, for a formal language $\+L$ equipped with a semantics in terms of relational models. We show that the approach is independent of the particular algorithm used to generate the referring expression by providing examples using the frameworks of \cite{AKS08} and \cite{Krahmer2003}. We provide some new complexity bounds, discuss the issue of the length of the generated descriptions, and propose ways in which the two approaches can be combined.