Property Alignment On The Semantic Web.pdf

The performance of current alignment systems on property matching lags behind that on class and instance matching. This work seeks to understand the reasons for this and consider avenues for improvement. The paper contains an exploration of the performance of current alignment systems on the only commonly accepted alignment benchmark that involves matches between properties. A second benchmark involving properties from DBPedia and YAGO, scaled to be within the capabilities of most existing alignment systems, is also proposed. A basic approach focused on aligning properties is then presented and evaluated using both benchmarks to serve as a baseline against which to compare more complex matchers on the property alignment task. The results show that even a relatively simplistic approach can achieve a significantly higher F-measure than current matchers. Finally, an existing full-featured alignment system is augmented with the basic property matching approach and the difference in performance is assessed.

The paper provides quantitative support for the claim that current alignment systems perform significantly worse on properties than on classes and explores the reasons behind this performance gap. These include differences in property labeling conventions that limit the effectiveness of string similarity metrics that are traditionally employed in alignment systems on properties compared to classes, as well as the common lack of a rich property hierarchy comparable to that often found for classes, which limits the utility of structural matching approaches.

The paper shows that it is possible to achieve reasonable performance on property matching if string metrics are chosen to maximize performance on the type of labels given to properties rather than classes, and if domain and range information, which is of course not applicable to classes, is considered. This basic approach provides a useful baseline against which alignment systems can be compared on property matching tasks.

Another limiting factor in the development of strong property alignment techniques is the lack of suitable benchmarks. The OAEI Conference track is currently the only commonly used non-synthetic benchmark that involves any matches between properties, and these are a small percentage of the overall matches within the benchmark. Therefore, a new property-centric alignment benchmark based on DBPedia and YAGO is proposed and evaluated.

The central contributions of this paper are: an in-depth analysis of the performance of current alignment systems on properties, including common false positives and false negatives; the introduction of a baseline property alignment approach designed for accurate similarity computation between properties; and the proposal of a second real-world property-centric alignment test set that is scaled to be within the capabilities of most existing alignment systems.

The paper also discusses related work on property matching, including the use of string similarity metrics, the impact of different naming conventions for properties, and the variety of design decisions made when an ontology is created. It also discusses the use of ontology transformation to handle differences in design philosophy and the development of a strategy called similarity flooding to improve the performance of alignment systems.