LiQuate-Estimating the Quality of Links in the Linking Open Data Cloud

摘要

During the last years, RDF datasets from almost any knowledge domain have been published in the Linking Open Data (LOD) cloud. The Linked Open Data guidelines establish the conditions to be satisfied by resources in order to be included as part of the LOD cloud, as well as connected to previously published data. The process of publication and linkage of resources in the LOD cloud relies on: i) data cleaning and transformation into existing RDF formats, ii) storage of the data into RDF storage systems, and iii) data interlinking. Because of data source heterogeneity, generated RDF data may be ambiguous and links may be incomplete with respect to this data. Users of the Web of Data require linked data to meet high quality standards in order to develop applications that can produce trustworthy results, but data in the LOD cloud has not been curated; thus, tools are necessary to detect data quality problems. For example, researchers that study Life Sciences datasets to explain phenomena or identify anomalies, demand that their findings correspond to current discoveries, and not to the effect of low data quality standards of completeness or redundancy. In this paper we propose LiQuate, a system that uses Bayesian networks to study the incompleteness of links, and ambiguities between labels and between links in the LOD cloud, and can be applied to any domain. Additionally, a probabilistic rule-based system is used to infer new links that associate equivalent resources, and allow to resolve the ambiguities and incompleteness identified during the exploration of the Bayesian network. As a proof of concept, we applied LiQuate to existing Life Sciences linked datasets, and detected ambiguities in the data, that may compromise the confidence of the results of applications such as link prediction or pattern discovery. We illustrate a variety of identified problems and propose a set of enriched intra- and inter-links that may improve the quality of data items and links of specific datasets of the LOD cloud