Two-stage Web Record Extraction

To extract structured data from the Web is a challenging subtask of information extraction. In the Web, the structured data are usually presented as lists, records or tables. Present methods require to identity the boundaries of data regions before separating them into records. Because records do not always have the same count of items or occur in consecutive sections, these methods often fail to handle such complicated or noisy pages. In this paper, we propose a fully automatic method called Two-Stage Web Record Extraction (TSWRE) to extract records from an open domain corpus. This approach uses a bottom-up analysis that starts with sequences of visually similar attribute sequences. It first identifies attribute sequences based on distinct tag paths of the ordered DOM tree of the document. The method exploits the position interleave characteristics of the attribute sequences to estimate how likely the sequences belong to the same records. Empirical experiments show that our method achieves promising performance compared to existing methods and is scalable to a large corpus.