Lost in translation? Empirical analysis of mapping compositions for large ontologies

When three or more ontologies have been aligned, longer chains of mapped concepts start to appear. In this paper, we empirically study the nature of these composite mappings, focusing on chains of (near) equivalence links of length two. We ask human experts to evaluate samples of composite mappings, taken from large real life data sets. Based on these evaluations, we analyze the features of mappings produced by composition in three different domains (bio-medicine, cultural heritage, and library subject headings), among ontologies in multiple languages (English, Dutch, German, and French), and using existing mappings that were created by different methods (lexical and instance-based methods). We examine the quality of the composite mappings relative to the quality of the input mappings and analyze how characteristics of the input mappings and the ontologies influence the composition.