Populating a Linked Data Entity Name System

Book excerpt: Resource Description Framework (RDF) is a graph-based data model used to publish data as a Web of Linked Data. RDF is an emergent foundation for large-scale data integration, the problem of providing a unified view over multiple data sources. An Entity Name System (ENS) is a thesaurus for entities, and is a crucial component in a data integration architecture. Populating a Linked Data ENS is equivalent to solving an Artificial Intelligence problem called instance matching, which concerns identifying pairs of entities referring to the same underlying entity. This publication presents an instance matcher with 4 properties, namely automation, heterogeneity, scalability and domain independence. Automation is addressed by employing inexpensive but well-performing heuristics to automatically generate a training set, which is employed by other machine learning algorithms in the pipeline. Data-driven alignment algorithms are adapted to deal with structural heterogeneity in RDF graphs. Domain independence is established by actively avoiding prior assumptions about input domains, and through evaluations on 10 RDF test cases. The full system is scaled by implementing it on cloud infrastructure using MapReduce algorithms. Resource Description Framework (RDF) is a graph-based data model used to publish data as a Web of Linked Data. RDF is an emergent foundation for large-scale data integration, the problem of providing a unified view over multiple data sources. An Entity Name System (ENS) is a thesaurus for entities, and is a crucial component in a data integration architecture. Populating a Linked Data ENS is equivalent to solving an Artificial Intelligence problem called instance matching, which concerns identifying pairs of entities referring to the same underlying entity. This publication presents an instance matcher with 4 properties, namely automation, heterogeneity, scalability and domain independence. Automation is addressed by employing inexpensive but well-performing heuristics to automatically generate a training set, which is employed by other machine learning algorithms in the pipeline. Data-driven alignment algorithms are adapted to deal with structural heterogeneity in RDF graphs. Domain independence is established by actively avoiding prior assumptions about input domains, and through evaluations on 10 RDF test cases. The full system is scaled by implementing it on cloud infrastructure using MapReduce algorithms.