Generating Relational References: What Makes a Difference?

Generating Relational References: What Makes a Difference? Jette Viethen Centre for Language Technology Macquarie University Sydney, Australia jviethen@ics.mq.edu.au Robert Dale Centre for Language Technology Macquarie University Sydney, Australia rdale@ics.mq.edu.au Abstract When we describe an object in order to en- able a listener to identify it, we often do so by indicating the location of that object with re- spect to other objects in a scene. This requires the use of a relational referring expression; while these are very common, they are rela- tively unexplored in work on referring expres- sion generation. In this paper, we describe an experiment in which we gathered data on how humans use relational referring expressions in simple scenes, with the aim of identifying the factors that make a difference to the ways in which humans construct referring expressions. 1 Introduction The generation of referring expressions—the pro- cess of determining how best to describe an ob- ject in order to enable the hearer to identify it—is a widely explored topic in natural language genera- tion research. We expect the first practical applica- tions of the algorithms developed in this area to be those where the location of objects within scenes is a requirement: examples of such scenarios are the description of entities such as buildings and other landmarks in automatically-generated route descrip- tions (see Dale et al. (2005)) and the description of locations in ‘omniscient room’ scenarios, where an intelligent agent might try to tell you where you left your RFID-tagged keys. In these scenarios, it is very likely that the referring expressions gener- ated will need to make use of the spatial relation- ships that hold between the intended referent and other entities in the domain; but, surprisingly, the generation of relational references is a relatively un- explored task. The few algorithms that address this task (Dale and Haddock (1991), Gardent (2002), Krahmer and Theune (2002), Varges (2005), Kelle- her and Kruiff (2005, 2006)) typically adopt fairly simple approaches: they only consider spatial rela- tions if it is not possible to fully distinguish the target referent from the surrounding objects in any other way, or they treat them in exactly the same as non- relational properties. As acknowledged by some of this work, this creates additional problems such as infinite regress and the inclusion of relations with- out regard for the properties of the landmarks that are associated with them. To be able to develop algorithms that meet the re- quirements of applications like those just mentioned, we first need to have a better understanding of how humans use relational referring expressions. In this paper, we provide a detailed analysis of a set of ex- perimental data that was gathered in a context where human subjects were encouraged to use relational expressions in order to identify intended referents. In Section 2 we describe our data gathering ex- periment in some detail, explaining the rationale be- hind our development of the test data we used. In Section 3 we provide a summary of the corpus of …

Content truncated to 500 words