Recent years have seen a trend towards empirically motivated and more data-driven approaches in the field of referring expression generation (REG). Much of this work has focussed on initial reference to objects in visual scenes. While this scenario of use is one of the strongest contenders for real-world applications of referring expression generation, existing data sets still only embody very simple stimulus scenes. To move this research forward, we require data sets built around increasingly complex scenes, and we need much larger data sets to accommodate their higher dimensionality. To control the complexity, we also need to adopt a hypothesis-driven approach to scene design. In this paper, we describe GRE3D7, the largest corpus of humanproduced distinguishing descriptions available to date, discuss the hypotheses that underlie its design, and offer a number of analyses of the 4480 descriptions it contains.
|Title of host publication||Proceedings of the UCNLG+Eval|
|Subtitle of host publication||Language Generation and Evaluation Workshop|
|Publisher||Association for Computational Linguistics (ACL)|
|Number of pages||11|
|Publication status||Published - 2011|
|Event||Workshop on Language Generation and Evaluation - Edinburgh|
Duration: 31 Jul 2011 → 31 Jul 2011
|Workshop||Workshop on Language Generation and Evaluation|
|Period||31/07/11 → 31/07/11|
Bibliographical noteCopyright the Publisher 2011. Version archived for private and non-commercial use with the permission of the author/s and according to publisher conditions. For further rights please contact the publisher.
Viethen, J., & Dale, R. (2011). GRE3D7: A Corpus of distinguishing descriptions for objects in visual scenes. In Proceedings of the UCNLG+Eval: Language Generation and Evaluation Workshop (pp. 12-22). Association for Computational Linguistics (ACL).