Data Mining in Document Structure

A document retroconversion method is described in this paper. Based on a training step using data mining on document samples recognized by OCR and converted in ALTO, we generate a decision tree, thanks to the ”Improved CHAID ” algorithm. The decision tree is then converted into syntactic rules which are used for the structure extraction. The programming is operated thanks to XSLT and Java. The experiments applied on 100 documents belonging to three conference proceedings reached a recall between 0.73 and 1.00 and a precision between 0.84 and 1.00 on 8 logical elements