Transforming Large Collections of Scientific Publications to XML

We describe an experiment transforming large collections of L ^a T_EX documents to more machine-understandable representations. Concretely, we are translating the collection of scientific publications of the Cornell e-Print Archive (arχiv) using LaTeXML, a L ^a T_EX to XML converter currently under development. While the long-term goal is a large body of scientific documents available for semantic analysis, search indexing and other experimentation, the immediate goals are tools for creating such corpora. The first task of our arXMLiv project is to develop LaTeXML bindings for the (thousands of) L ^a T_EX classes and packages used in the arχiv collection, as well as methods for coping with the eccentricities that T_EX encourages. We have created a distributed build system that runs LaTeXML over the collection, in part or entirely, while collecting statistics about missing bindings and other errors. This guides debugging and development efforts, leading to iterative improvements in both the tools and the quality of the converted corpus. The build system thus serves as both a production conversion engine and software test harness. We have now processed the complete arχiv collection through 2006 consisting of more than 400,000 documents (a complete run is a processor-year-size undertaking), continuously improving our success rate. We are now able to convert more than 90% of these documents to XHTML+MathML. We consider over 60% to be successes, converted with no or minor warnings. While the remaining 30% can also be converted, their quality is doubtful, due to unsupported macros or conversion errors.