Instance-based Hierarchical Schema Alignment in Linked Data

Abstract

Along with the development of Web of documents, there is a natural need for sharing, exchanging, and merging heterogeneous data to provide more comprehensive information and answer users with more complex questions. However, the data published on the Web are raw dumps that sacrifice much of the semantics that can be used for exchanging and integrating data. Resource Description Framework (RDF) and Linked Data are designed to expose the semantics of data by interlinking data represented with well-defined relations. With the profusion of RDF resources and Linked Data, ontology alignment has gained significance in providing highly comprehensive knowledge embedded in disparate sources. Ontology alignment, however, in Linking Open Data (LOD) has traditionally focused more on the instance-level rather than the schema-level. Linked Data supports schema-level matching, provided that instance-level matching is already established. Linked Data is a hotbed for instance-based schema matching, which is considered a better solution for matching classes with ambiguous or obscure names. In this dissertation, the author focuses on three issues in instance-based schema alignment for Linked Data: (1) how to align schemas based on instances, (2) how to scale the schema alignment, (3) how to generate a hierarchical schema structure. Targeting the first issue, the author has proposed an instance-based schema alignment algorithm called IUT. The IUT builds a unified taxonomy for the classes from two ontologies based on an instance-class matrix and obtains the relations of two classes by the common instances. The author tested the IUT with DBpedia and YAGO2, and compared the IUT with two state-of-the-art methods in four alignment tasks. The experiments show that the IUT outperforms the methods in terms of efficiency and effectiveness (e.g., costs 968 ms to obtain 0.810 F-score on intra-subsumption alignment in DBpedia). Targeting the second issue, the author has proposed a scaled version of the IUT called IUT(M). The IUT(M) decreases the computations of the IUT from two aspects based on Locality Sensitive Hashing (LSH): (1) decreasing the similarity computations for each pair of classes with MinHash functions, and (2) decreasing the number of similarity computations with banding. The author tested the IUT(M) with YAGO2-YAGO2 intra-subsumption alignment task to demonstrate that the running time of IUT can be reduced by 94% with a 5% loss in F-score. Targeting the third issue, the author has proposed a method to generate a faceted taxonomy based on object properties on Linked Data. A framework is proposed to build a sub-taxonomy in each facet with sub-data, extracted with an object property, with an Instance-based Concept Taxonomy generation algorithm called ICT. Two experiments demonstrate: (1) The ICT efficiently and effectively generates a sub-taxonomy with rdf:type in DBpedia and YAGO2 (e.g., costs 49 and 11,790 ms to build the concept taxonomies that achieve 0.917 and 0.780 on Taxonomic F-score). (2) The faceted taxonomies for Diseasome and DrugBank, efficiently generated based on multiple object properties (e.g., costs 2,032 and 2,525 ms to build the faceted taxonomies based on 6 and 16 properties), can effectively reduce the search spaces in faceted searches (e.g., obtains 1.65 and 1.03 on Maximum Resolution with 2 facets).