Schema extraction for tabular data on the web

Tabular data is an abundant source of information on the Web, but remains mostly isolated from the latter’s inter-connections since tables lack links and computer-accessible descriptions of their structure. In other words, the schemas of these tables — attribute names, values, data types, etc. — are not explicitly stored as table metadata. Consequently, the structure that these tables contain is not accessible to the crawlers that power search engines and thus not accessi-ble to user search queries. We address this lack of structure with a new method for leveraging the principles of table construction in order to extract table schemas. Discover-ing the schema by which a table is constructed is achieved by harnessing the similarities and differences of nearby ta-ble rows through the use of a novel set of features and a feature processing scheme. The schemas of these data ta-bles are determined using a classification technique based on conditional random fields in combination with a novel fea-ture encoding method called logarithmic binning, which is specifically designed for the data table extraction task. Our method provides considerable improvement over the well-known WebTables schema extraction method. In contrast with previous work that focuses on extracting individual re-lations, our method excels at correctly interpreting full ta-bles, thereby being capable of handling general tables such as those found in spreadsheets, instead of being restricted to HTML tables as is the case with the WebTables method. We also extract additional schema characteristics, such as row groupings, which are important for supporting information retrieval tasks on tabular data. 1