Extracting discourse elements and annotating scientific documents using the SciAnnotDoc model: a use case in gender documents

When scientists are searching for information, they generally have a precise objective in mind. Instead of looking for documents "about a topic T", they try to answer specific questions such as finding the definition of a concept, finding results for a particular problem, checking whether an idea has already been tested, or comparing the scientific conclusions of two articles. Answering these precise or complex queries on a corpus of scientific documents requires precise modelling of the full content of the documents. In particular, each document element must be characterised by its discourse type (hypothesis, definition, result, method, etc.). In this paper, we present a scientific document model (SciAnnotDoc ontology), developed from an empirical study conducted with scientists, that models the discourse types. We developed an automated process that analyses documents effectively identifying the discourse types of each element. Using syntactic rules (patterns), we evaluated the process output in terms of precision and recall using a previously annotated corpus in Gender Studies. We chose to annotate documents in Humanities, as these documents are well known to be less formalised than those in "hard science". The process output has been used to create a SciAnnotDoc representation of the corpus on top of which we built a faceted search interface. Experiments with users show that searches using with this interface clearly outperform standard keyword searches for precise or complex queries.