MateTee: A Semantic Similarity Metric Based on Translation Embeddings for Knowledge Graphs

Large Knowledge Graphs (KGs), e.g., DBpedia or Wikidata, are created with the goal of providing structure to unstructured or semi-structured data. Having these special datasets constantly evolving, the challenge is to utilize them in a meaningful, accurate, and efficient way. Further, exploiting semantics encoded in KGs, e.g., class and property hierarchies, provides the basis for addressing this challenge and producing a more accurate analysis of KG data. Thus, we focus on the problem of determining relatedness among entities in KGs, which corresponds to a fundamental building block for any semantic data integration task. We devise MateTee, a semantic similarity measure that combines the gradient descent optimization method with semantics encoded in ontologies, to precisely compute values of similarity between entities in KGs. We empirically study the accuracy of MateTee with respect to state-of-the-art methods. The observed results show that MateTee is competitive in terms of accuracy with respect to existing methods, with the advantage that background domain knowledge is not required.