An Empirical Study of Vocabulary Relatedness and Its Application to Recommender Systems

Abstract

Abstract. When thousands of vocabularies having been published on the Semantic Web by various authorities, a question arises as to how they are related to each other. Existing work has mainly analyzed their similarity. In this paper, we inspect the more general notion of relatedness, and characterize it from four angles: well-defined semantic relatedness, lexical similarity in contents, closeness in expressivity and distributional relatedness. We present an empirical study of these measures on a large, real data set containing 2,996 vocabularies, and 15 million RDF documents that use them. Then, we propose to apply vocabulary relatedness to the problem of post-selection vocabulary recommendation. We implement such a recommender service as part of a vocabulary search engine, and test its effectiveness against a handcrafted gold standard.

An empirical survey of Linked Data conformance

“…39 Dodds has also raised similar concerns; cf. http://www.flickr.com/photos/ ldodds/4043803502/ ; retr.…”

“…More recently, Cheng et al [39] performed a study of 2996 Web vocabularies, spanning 261 pay-level-domains, finding 396023 classes and 59868 properties. Approximately 72% of vocabularies were found to contain no more than 25 terms.…”

“…However, others have also observed a worrying lack of licencing information for RDF documents published on the Web, where, e.g., Bizer et al put the figure at 15% of Linked Data publishers offering document-level licencing [52]. 39 …”

Resource-constrained reasoning using a reasoner composition approach

“…The selective rule loading algorithm dimensions a selected ruleset according to the amount of semantics contained in the target ontology. There are two motivating factors behind selective rule loading algorithm: (1) the diversity in the amount of semantics included in different ontologies, ranging from lightweight taxonomies using (partial) RDFS to ontologies with complicated concept descriptions using full OWL 1 or OWL 2 expressiveness [6,33], and (2) the often very fine-grained semantics carried by each RS entailment rule. Therefore, by avoiding the loading of unnecessary rules, a reasoner can construct a smaller RETE network and hence can reduce memory consumption and reasoning time.…”

Making sense from Big RDF Data: OUSAF for measuring ontology usage

“…They observed that most of the ontologies from the OWL family are OWL Domain Labels and OWL Lite ontologies. Gong et al conducted a study on roughly 3000 vocabularies, comprising 396,023 classes and 59,868 properties in total. In addition to vocabulary documents, the authors also considered 15 million instance documents to investigate the relatedness between ontologies.…”