The longstanding problem of automatic table interpretation still illudes us. Its solution would not only be an aid to table processing applications such as large volume table conversion, but would also be an aid in solving related problems such as information extraction and semi-structured data management. In this paper, we offer a conceptual modeling solution for the common special case in which so-called sibling pages are available. The sibling pages we consider are pages on the hidden web, commonly generated from underlying databases. We compare them to identify and connect nonvarying components (category labels) and varying components (data values). We tested our solution using more than 2,000 tables in source pages from three different domains-car advertisements, molecular biology, and geopolitical information. Experimental results show that the system can successfully identify sibling tables, generate structure patterns, interpret tables using the generated patterns, and automatically adjust the structure patterns, if necessary, as it processes a sequence of hidden-web pages. For these activities, the system was able to achieve an overall F-measure of 94.5%.
展开▼
