AFIS: Annotation-Free Induction of Full Schema for Detail-Pages

Web data extraction is an essential task for web data integration. Most researches focus on data extraction from list-pages by detecting data-rich section and record boundary segmentation. The problem of data alignment in records is small scale since only a couple data attributes need to be aligned. However, for detail-pages which contain all-inclusive product information in each page, the number of data attributes need to be aligned is much larger. In this paper, we formulate the data extraction problem as alignment of leaf nodes from DOM Trees. We propose AFIS, an Annotation-Free Induction of full Schema for detail-pages. AFIS applies Divide-and-Conquer and Longest Increasing Sequence (LIS) algorithms to mine landmarks from input. The experiments show that AFIS outperforms Road-Runner, FivaTech and TEX (with precision 0.994, recall 0.987, and F1 0.990) in terms of selected (data) columns. For full schema evaluation (all data columns), AFIS also represents the highest average performance (with precision 0.946, recall 0.930, and F1 0.937) compared with TEX and RoadRunner.