Article Preview
Top1. Introduction
In the past few years, billions of RDF triples have been published as Linked Data describing numerous entities. An entity usually involves multiple aspects and its property-values may focus on different aspects. For instance, latitude and longitude reveal the spatial coordinates of a location, while street, city and zip code present the address information of that location. Therefore, it is natural to cluster properties into meaningful groups based on the aspects that they intend to describe. Property clustering is useful for many applications such as entity browsing (Rutledge, van Ossenbruggen, & Hardman, 2005), entity summarization (Gunaratna, Thirunarayan, & Sheth, 2015), entity coreference resolution (Hu & Jia, 2015), query completion, etc. It can be used to present the entity information in a more formatted and understandable fashion, which significantly enhances the capability of users to consume the large-scale Linked Data (Hearst, 2006).
Take, for example, the case of entity browsing. Many state-of-the-art systems support users to manually cluster properties (Quan & Karger, 2004). But due to the limited energy and knowledge of the users, this type of manual operations is only effective at a small scale. In consideration of an open domain like Linked Data, automated property clustering is needed to solve the scalability issue, but its performance is still far from satisfactory. One reason is the fact that, when browsing entities, the data is multi-sourced and the vocabularies involved are barely predictable, i.e. probably use any vocabularies, thus it is difficult to identify useful patterns among properties in advance and make use of them to guide clustering. Another reason is that the properties used by these entities are largely heterogeneous, which makes identifying similar aspects more difficult and less accurate.
Another example is the case of context-based entity summarization. As a typical Linked Data application scenario, when a user is browsing some plain text, a tool automatically identifies the Linked Data entities mentioned in the text and shows the summaries of them according to the browsing context, e.g. the surrounding text of an entity that may represent some entity aspect. In this case, the properties in the entity data would be clustered according to the aspects they describe, and the ones that match the browsing context may be selected and shown in the summary.
Several property relatedness measures have been proposed for clustering in existing studies. However, there are few studies investigating how these measures work in an open domain environment, i.e. facing a large scale of multi-sourced, heterogeneous data. Besides, how to combine different clustering results for improving performance has not been deeply investigated. In this paper, we empirically studied property clustering in Linked Data. In order to achieve clustering, we firstly measured the relatedness of properties from five perspectives using 13 different measures: lexical similarity between property names (five measures), semantic relatedness between property names (three measures), distributional relatedness between properties (two measures), domain/range relatedness between properties (two measures), and overlap of property values. We then employed seven widely-used clustering algorithms of different characteristics. Furthermore, to combine various relatedness measures and clustering results, we developed two combination methods based on linear combination and consensus clustering (Ailon, Charikar, & Newman, 2008). In summary, the studied property relatedness measures, clustering algorithms, combination methods and their notations are listed in Table 1.