A Framework for Semantic Similarity Measures to enhance Knowledge Graph Quality

Abstract

Precisely determining similarity values among real-world entities becomes a building block for data driven tasks, e.g., ranking, relation discovery or integration. Semantic Web and Linked Data initiatives have promoted the publication of large semi-structured datasets in form of knowledge graphs. Knowledge graphs encode semantics that describes resources in terms of several aspects or resource characteristics, e.g., neighbors, class hierarchies or attributes. Existing similarity measures take into account these aspects in isolation, which may prevent them from delivering accurate similarity values. In this thesis, the relevant resource characteristics to determine accurately similarity values are identified and considered in a cumulative way in a framework of four similarity measures. Additionally, the impact of considering these resource characteristics during the computation of similarity values is analyzed in three data-driven tasks for the enhancement of knowledge graph quality. First, according to the identified resource characteristics, new similarity measures able to combine two or more of them are described. In total four similarity measures are presented in an evolutionary order. While the first three similarity measures, OnSim, IC-OnSim and GADES, combine the resource characteristics according to a human defined aggregation function, the last one, GARUM, makes use of a machine learning regression approach to determine the relevance of each resource characteristic during the computation of the similarity. Second, the suitability of each measure for real-time applications is studied by means of a theoretical and an empirical comparison. The theoretical comparison consists on a study of the worst case computational complexity of each similarity measure. The empirical comparison is based on the execution times of the different similarity measures in two third-party benchmarks involving the comparison of semantically annotated entities. Ultimately, the impact of the described similarity measures is shown in three data-driven tasks for the enhancement of knowledge graph quality: relation discovery, dataset integration and evolution analysis of annotation datasets. Empirical results show that relation discovery and dataset integration tasks obtain better results when considering semantics encoded in semantic similarity measures. Further, using semantic similarity measures in the evolution analysis tasks allows for defining new informative metrics able to give an overview of the evolution of the whole annotation set, instead of the individual annotations like state-of-the-art evolution analysis frameworks.