Web tables have been proven to constitute valuable sources of information for applications, ranging from Web search, to data discovery in spreadsheet software and KB augmentation [1]. A requirement for those applications is to understand the semantics of Web tables and potentially match their contents with existing URIs in the Web of Data, a process known as Web table annotation [4]. Recent works on Web table annotation follow an iterative approach between instance- and schema-level re nements, until convergence [6, 7]. In this work, we annotate Web tables using ontology matching. As this eld has solid tools and benchmarks3, we design a framework that provides the required input to any ontology matching tool, resulting in Web table annotations. Moreover, our blocking enables even the less scalable ontology matching tools provide annotations to large-scale KBs, such as DBpedia. The contributions of our work are: { We introduce a generic and scalable framework for Web table annotation using existing ontology alignment systems. { We evaluate our framework and compare the results against state-of-the-art Web table annotation tools, with promising results. { Our framework can be extended as a benchmark for ontology matching tools.
Introduction
Model. We assume that each table row describes a real-world entity, and each
column represents a property. Each cell of the header row de nes the name of
a property, except the cell of the label column, which de nes the name of the
table's class. All the entities in the table are instances of this class. The values of
a column can be either literals, or references to other entities, corresponding to
dataype, or object properties, respectively. To make this distinction, we sample
the data types of each column, also identifying the label column, as in [
Blocking. To enable ontology matching tools that do not scale well be
applicable in this framework, and to improve the e ciency of matching tools that
do scale, we have applied a pre-processing step of candidate mappings selection,
known as blocking [
Finally, we call an ontology matching tool with the table ontology and the DBpedia ontology after blocking, as input, and return the mapping results.
Evaluation. We evaluate our approach using the instance mappings of the
T2D gold standard4 and LogMap [
Table 1. Results over T2D gold standard. Blocking results in parentheses.
Method Recall Precision F-measure DBpedia lookup 0.73 0.79 0.76 DBpedia lookup re ned 0.76 0.86 0.81 T2K 0.76 0.90 0.82
Ontology matching 0.57 (0.71) 0.89 (0.32) 0.70 (0.44)
The results show that our framework, using LogMap, suggests a good number of correct results, with high precision. In the future, we plan to improve blocking and extend our model to provide a rst alignment, which can be utilized by many ontology matching tools. Our goal is to provide an ontology matching benchmark for instance-, class- and property-mappings, that can result in a new track in the upcoming OAEI campaigns.
5 http://wiki.dbpedia.org/projects/dbpedia-lookup