A Grammar Based Analysis of Column Header Categories for Web Tables

As part of a project to harvest semi-structured data from web tables, we describe an approach to extract an abstract representation of the column-header categories based on a context-free grammar for linear strings. The column-header structure is generally an XY-tessellation. The grammar provides a compact representation of infinitely many structural variations possible within column headers. Before parsing, the 2D column-header structure is converted to a linear string of its atomic cell labels and delimiters for the X and Y cuts. The acceptable strings represent a superset of admissible column-header structures from which the invalid ones are eliminated by performing geometric and lexical checks on the labels of the parse tree. Experimental results on web tables show that 80% of the headers in the sample could be processed successfully using the grammatical approach.