Semantic similarity of natural language spatial relations

Abstract

Communication problems between humans and machines are often the reason for failures or wrong computations. While machines use well-defined languages and rules in formal models to compute information, humans prefer natural language expressions with only vaguely specified semantics. Similarity comparisons are a central construct of the human way of thinking. For instance, humans are able to act sensible in completely new situations by comparing them to similar experiences in the past. Similarity is used for reasoning on unknown information. It is necessary to overcome the differences in representing and processing information to avoid error-prone communication. A machine being able to understand natural language and detect the semantic similarity between expressions would be the key to eliminate human-machine communication problems. This paper addresses human-machine communication about spatial configurations in natural language. We propose a computational model to capture the semantics of natural language spatial relations and to compare them with respect to their semantic similarity. The semantic description is based on an approach developed by Shariff, Egenhofer and Mark which describes natural language spatial relations via a combination of several formal spatial relations. The semantic similarity measure is inspired by Gärdenfors' conceptual spaces: we model the formal relations as quality dimensions of a geometric space, describe the natural language expressions as regions in the multidimensional space and determine their similarity via spatial distances.