Using robustness to learn to order semantic properties in referring expression generation

  title={On the Robustness of Standalone Referring Expression Generation Algorithms Using RDF Data},
  author={Duboue, Pablo Ariel and Dom{\i}nguez, Martin Ariel and Estrella, Paula},
  booktitle={WebNLG 2016},
  address = {Edinburgh, UK},

This paper, written jointly with the large scale social and semantic networks research group at FaMAF was presented at IBERAMIA 2016 in San Jose, Costa Rica.

Continuing with the research line in GRE, in this paper we sought to learn a default ordering for referring expressions using the concept of robustness, meaning "how well does this ordering works out in the presence of errors in the training data". More salient, often-defined and highly discriminant characteristics of referents are expected to work well even if the speaker is unsure about some of more rare properties of the referent.

At its core, it tries to drill down on whether the default ordering identified in DR1995 is is psycholinguistically plausible. Our hypothesis is that descriptions that capture characteristics less likely to change will be preferred by the community of speakers.

We found that robustness was not enough to learn a full ordering from scratch by it had a strong correlation with human-written orderings.

The slides of the paper are available contact me if you want a preprint.