A methodology for topographic clustering of structured text documents

Sets of texts are structured through a more or less refined hierarchy of sections, subsections and paragraphs; this structure contains information that should be exploited to handle these data and in particular, to enrich the comparison of texts, as a complement to the vector description of their contents. We propose a kernel-based methodology that follows this principle for a topographic clustering task and define a hierarchical kernel which compares paragraphs using the available hierarchical decomposition and in particular the provided titles.