Start Small, Build Complete: Effective and Efficient Semantic Table Interpretation using TableMiner

This article introduces TableMiner, the first semantic Table Interpretation method able to annotate Web tables using an incremental, bootstrapping learning approach seeded by automatically selected ‘partial’ content from tables. The basic principle is to create initial and partial annotations of a table using some as opposed to all content in the table. The partial outcome then serves as ‘stepping stones’ to guide interpretation of remaining content in the table, followed by a process of iteratively refining results on the entire table to create final optimal annotations. To construct feature representations, TableMiner uses various types of contextual information both inside and outside tables, including pre-defined semantic markups (e.g., RDFa/Microdata annotations) within some webpages that to the best of the author’s knowledge, have never been used in Natural Language Processing tasks. Evaluated on the largest collection of datasets known to-date including four datasets with a total of more than 15,000 tables, TableMiner consistently outperforms four baselines under all experimental settings. On the twomost representative datasets covering multiple domains and various table schemata, it achieves improvement in F1 by between 1 and 42 percentage points depending on specific tasks. Compared against state-of-the-art, it also obtains higher results on similar datasets. The bootstrapping learning strategy seeded by partial table content also enables TableMiner to be very efficient. Empirically, it reduces data to be processed by up to 66% or 29% savings in CPU time when compared against classic methods that ‘exhaustively’ processes the entire table content to build features for interpretation.