Research – Paper 98
Web tables constitute valuable sources of information for various applications, ranging from Web search to Knowledge Base (KB) augmentation. An underlying common requirement is to annotate the
rows of Web tables with semantically rich descriptions of entities published in Web KBs. In this paper, we evaluate three unsupervised annotation methods: (a) a lookup-based method which relies on the minimal entity context provided in Web tables to discover correspondences to the KB, (b) a semantic embeddings method that exploits a vectorial representation of the rich entity context in a KB to identify the most relevant subset of entities in the Web table, and (c) an ontology matching method, which exploits schematic and instance information of entities available both in a KB and a Web table. Our experimental evaluation is conducted using two existing benchmark data sets in addition to a new large-scale benchmark created using Wikipedia tables. Our results show that: 1) our novel lookup-based method outperforms state-of-the-art lookup-based methods, 2) the semantic embeddings method outperforms lookup-based methods in one benchmark data set, and 3) the lack of a rich schema in Web tables can limit the ability of ontology matching tools in performing high-quality table annotation. As a result, we propose a hybrid method that significantly outperforms individual methods on all the benchmarks.