“…The most direct originates in work that grows out of the INEX and ICDAR book structure extraction com- petitions (Kazai et al, , 2010Doucet et al, 2011Doucet et al, , 2013, in which participants are challenged to recognize the fine-grained structure present in documents (recognizing, for example, that the current article has sections entitled "Abstract," "Introduction," "Data, "References," etc.). The most successful systems recognize structure by parsing the table of contents Meunier, 2005, 2009;Wu et al, 2013;Gao et al, 2009) rather than relying on the content of the book itself. Our work primarily differs in the fundamental design choice of prescribing a fix set of categories into which we classify pages (in order to enable comparison across documents) rather than prioritizing the idiosynractic structure of a book (which is useful for generating new tables of contents).…”