Article Preview
TopIntroduction
Data integration has been widely studied by the database community. In the web of data and especially in linked data, it becomes one of the main issues in data sharing and exploitation. However, ontology matching is one of the crucial tasks that support data integration where identifying correspondences presents a challenging problem. Instance matching (IM) is a subtask of ontology matching and seems to be an interesting solution. IM aims at discovering correspondences between instances, from various sources, that refer to the same real-world objects. Despite recent advances, IM still presents a real problem especially with the rapid growth of published data (Cukier & Mayer-Schoenberger, 2013;Zhang et al., 2008). It is a long-standing issue known as record linkage (Newcombe et al., 1959), merge/purge problem (Hernández & Stolfo, 1995) or reference reconciliation (Dong et al., 2005). For its importance, several approaches have been proposed such as ASL (Nguyen & Ichise, 2018), AIM-PC (Lu et al., 2018), RIMOM-IM (Shao et al., 2016), SERIMI (Araujo et al., 2011; 2015) and VMI (Li et al., 2013), approach of (Wang et al., 2013), FBEM (Stoermer & Rassadko, 2009), DSSim (Nagy et al., 2008) and HMatch(I) (Castano et al., 2008). Although the existing approaches work well in many cases of IM, they could fail to detect correspondences between instances that possess different descriptions.
In IM, identifying correspondences or matches is achieved by establishing links between instances from multiple sources to be integrated. The most important link is the owl:sameAs one. However, a significant number of existing owl:sameAs links on the Web of data do not adhere to their formal semantic (Halpin et al., 2010). This is due to the diverse contexts of instances descriptions. For example, when the similarity score between two instances indicates that they are different, they should be considered as non-matches. But in fact, they could refer to the same real-world object. Another issue of misidentified instances occurs when two instances referring to different objects are still connected by owl:sameAs link. For these reasons, many studies have shown that the owl:sameAs can lead to inconsistencies. Some of them propose novel constructors or predicates to replace owl:sameAs (Raad et al., 2017; Idrissou et al., 2017; Halpin et al, 2010), while other works focus on the detection of incorrect or quality sameAs statements (De Melo, 2013; Papaleo et al., 2014). Our purpose is to deal with the problem of discovering more correct links between instances with multiple descriptions and that are provided by various sources.
In this paper, we propose a novel approach dealing with the IM problem where matching instances with diverse descriptions presents the principal aim. This problem is poorly addressed in the literature. The proposed approach is based on three processes: IM-PC (IM based on Property Classification), IM-VSA (IM based on ViewSameAs) and IM-AMD (IM based on Alignment MetaData). The first process aims to maximize the utilization of discriminative information within instances where a novel link is introduced named ViewSameAs. This last allows connecting partially similar instances. The second process IM-VSA allows discovering more correspondences by reusing the alignment results of IM_PC. A novel method named ViewSameAs-based clustering is thus proposed. By using metadata about IM_VSA, the third process IM_AMD allows improving our IM solution. In fact, the alignment reuse is considered as a new and helpful technique in IM where few works used it such as RIMOM-IM (Shao et al., 2016). The contributions of this paper can be summarized as follows: