Domain Speci c Understanding of Spatial Expressions

This paper explores the notion of spatial expression understanding relative to a domain. A model for spatial expression understanding is proposed, and ways in which domain speci c information can be incorporated at di erent stages of the model are discussed. We describe the domain of newspaper captions and show that spatial information in this domain is primarily expressed in the open class and by projective spatial prepositions. A discussion of types of open class spatial information and the role of a priori knowledge about a domain in interpreting this information is presented. Many of the above observations have been incorporated into PICTION, a system that extracts spatial and visual information from newspaper captions for collateral-based image understanding. The representation and processing of spatial information in PICTION is presented as an application of the proposed model. The paper concludes with a discussion of our approach, and the applicability of the model to other domains.