When consuming Linked Data from multiple sources, or from a data source after deduplication of entities, con icting, missing or outdated values must be dealt with during data fusion in order to increase the usefulness and quality of the data. In this poster, we argue that the nature of Linked Data in RDF requires a more sophisticated approach to data fusion than the current Linked Data fusion tools provide. We demonstrate where they fall short on a real case of public procurement data fusion when dealing with property dependencies and fusion of structured values, and we propose new data fusion extensions to address these problems.
We demonstrate the need for new data fusion capabilities on a real scenario with public procurement data extracted from an XML-based API and RDFized using the Uni edViews framework. The extracted data needs to be deduplicated and fused in order to obtain high-quality data before further analytical processing.
Fig. 1 shows how a business entity (BE) is represented in RDF. It has a legal name, address, and o cial identi er, which may be marked as syntactically invalid. The extracted data contains many copies of the same BE because of duplication in the source dataset. Simple merge of matched BEs would result in data con icts due to misspellings and errors in the dataset or mismatches in the generated owl:sameAs links. Our goal is to fuse BEs so that each has a single legal name, address, and identi er, choosing the best possible values.
Property dependencies. We encounter the rst problem with the state-ofthe-art Linked Data fusion tools when fusing addresses. The tools resolve each property independently which can result in the selection of, e.g., a town from one address in the input and a postal code from another one. Such result could be incorrect, however, because the postal code is related to the town. We need to introduce dependency between properties to obtain a correct fused result.
Fusing structured values. Both address and identi er can be regarded as structured values of a BE. We will refer to the main resource (e.g., BE) as a parent resource and to the resource representing the structured value (e.g., address) as a dependent resource. Currently, structured values need to be fused separately. One way of achieving this is generating owl:sameAs links among structured values based on their properties, e.g., match addresses based on similarity of street, and town. This approach has two drawbacks: it doesn't guarantee that a BE will have only a single address after fusion, and the error of automatically generated owl:sameAs links accumulates. Another way is generating owl:sameAs links between dependent resources that belong to the same parent resource. This approach may lead to errors when two parent resources point to the same dependent resource, e.g., two di erent BEs point to the same address. All addresses for the two BEs would incorrectly be merged in such case.
We argue that a smarter approach considering structured values in resource
descriptions could (
In this section, we propose how to extend data fusion to improve on the issues demonstrated in Section 2. Let there be a set U (RDF URI references), a set B (blank nodes) and a set L (literals). A triple (s; p; o) 2 (U [B) U (U [B [L) is an RDF triple and we refer to its components as subject, predicate, and object, respectively. Let g 2 U be a graph name. We regard a triple (s; p; o) that belongs to a named graph g as a quad (s; p; o; g). 3.1
Property Dependencies Independent fusion of properties is not always su cient, as we demonstrated in Section 2. What we want is to keep the values of dependent properties together in the fused result if the values occurred together in the input data. Let us call a set of input quads sharing the same subject s and graph name g before resolution of identity con icts an input group IGs;g. Furthermore, let d(p1; p2) denote that there is a dependency between properties p1 and p2.
De nition 1. A fused result R from input quads I satis es property dependencies if and only if 8p1; p2 2 U such that d(p1; p2): all quads (s; p; o; g) 2 R such that p = p1 _ p = p2 are derived 5 from the same input group in I.
We chose to de ne input groups based on subject and graph because it covers
two common scenarios: (
Here is how a basic data fusion algorithm can be extended to produce results satisfying property dependencies. The input of the algorithm includes these dependencies { we assume it is given as an equivalence relation d. We also assume the input resource description contains all quads for all mutually dependent properties. The extended algorithm consists of the following high level steps: 1. Find equivalence classes C of the equivalence relation d. 2. For every class of dependent properties C 2 C: (a) Let IC be all input quads with one of the properties in C. (b) For every nonempty input group Is;g in IC , let Os;g be the fused result of the basic data fusion algorithm applied on Is;g. (c) Select one set OC from all sets Os;g of fused quads according to some fusion tool speci c criterion and add OC to the result. 3. Fuse input quads with properties that do not have any dependency using the basic data fusion algorithm.
It is straightforward to prove that the extended algorithm indeed produces results satisfying property dependencies. The criterion used in step (2c) can depend on the implementing fusion tool. In LD-FusionTool, which can assess the quality of fused quads, we select Os;g such that the average quality of the fused result is maximal. 5 By derived we mean \selected from" for the so called deciding fusion functions such as Latest, or \computed from" for mediating fusion functions such as Average. Current Linked Data tools fuse resource descriptions composed of triples having the respective resource as its subject. Further triples describing structured values are not included (e.g., street is not included for a BE). This leaves a space for improvement as demonstrated in Section 2. We propose the inclusion of dependent resources reachable from the parent resource through specially annotated properties, in analogy to [4]. For resource r with resource description R, we fuse a property p annotated with fusion function DependentResource as follows:
This approach produces a single fused dependent resource (e.g., a single
address of a BE), and takes advantage of the locality of owl:sameAs links to avoid
incorrect merge of dependent resources with multiple parents. A unique URI is
generated in step (
Our practical experience with fusion of public procurement data shows that the graph nature or RDF has its speci cs that need to be addressed. State-of-the-art Linked Data fusion tools do not cover two common patterns in RDF: fusion of structured values, and dependencies between properties.
We answer this challenge with new data fusion features. We introduce the concepts of dependent properties and dependent resources, and propose how to appropriately extend data fusion. The extensions have been implemented in LDFusionTool and successfully used to ful ll the goals of our motivational scenario.
The new data fusion features show a new direction in Linked Data fusion { taking advantage of the broader context in the RDF graph. This can be further leveraged not only in con ict resolution, but also in quality assessment. Acknowledgement. This work was supported by a grant from the EU's 7th Framework Programme number 611358 provided for the project COMSODE.