Discovering Types in RDF Datasets
An increasing number of linked datasets is published on the Web, expressed in RDF(S)/OWL. Interlinking, matching or querying these datasets require some knowledge about the types and properties they contain. This work presents an approach, relying on a cl
- PDF / 280,780 Bytes
- 5 Pages / 439.37 x 666.142 pts Page_size
- 76 Downloads / 183 Views
Abstract. An increasing number of linked datasets is published on the Web, expressed in RDF(S)/OWL. Interlinking, matching or querying these datasets require some knowledge about the types and properties they contain. This work presents an approach, relying on a clustering algorithm, which provides the types describing a dataset when this information is incomplete or missing. Keywords: Type extraction · Clustering · Semantic web · Linked data
1
Introduction
An increasing number of linked datasets is published on the Web. Understanding these datasets is crucial in order to exploit them. Having some knowledge about the content of a dataset, such as the types it contains, is crucial for users and applications as it will enable many tasks, such as creating links between datasets or querying them. Linked datasets are not always complete with respect to type information. Even when they are automatically extracted from a controlled source, type information can be missing: in DBpedia (extracted from Wikipedia), 63.7 % of type information is provided [8]. Our goal is to infer the types describing an RDF(S)/OWL dataset. Our main contribution is a deterministic and automatic approach relying on a clustering algorithm to extract types, where several types can be assigned to an entity. Our approach does not require any schema related information in the dataset. We have implemented our algorithms and we present some experimental evaluation results to demonstrate the effectiveness of the approach.
2
Type Discovery
In order to infer the types from a dataset, our approach relies on grouping entities according to their similarity. A group of similar entities corresponds to a type definition. The similarity between two given entities is evaluated considering their respective sets of both incoming and outgoing properties. Our main requirements are the following: (i) the number of types is not known in advance, (ii) an entity can have several types, and (iii) the datasets may contain noise. The most suitable grouping approach is density-based clustering, c Springer International Publishing Switzerland 2015 F. Gandon et al. (Eds.): ESWC 2015, LNCS 9341, pp. 77–81, 2015. DOI: 10.1007/978-3-319-25639-9 15
78
K. Kellou-Menouer and Z. Kedad
introduced by [2], because it is robust to noise, deterministic and it finds classes of arbitrary shape. In addition, unlike the algorithms based on k-means and k-medoid, the number of classes is not required. Our density-based algorithm has two parameters: the maximum radius of neighborhood ε and the minimum number of neighbors for an entity M inP ts. ε represents the minimum similarity value for two entities to be considered as neighbors. We use Jaccard similarity to measure the closeness between two property sets describing two entities. M inP ts is the minimum number of similar entities required to form a core [2]: an entity is not assigned to a class if it is considered as noise, i.e. if it is neither a core itself nor the neighbor of a core. In order to speed up the clustering process, we
Data Loading...
