A Framework for Obtaining Structurally Complex Condensed Representations of Document Sets in the Biomedical Domain

Resumen

In this paper, we present a framework for obtaining structurally complex condensed representations of documents sets, which will be used as a base for summarization, answering complex questions, etc. This framework includes a method for extracting a ranked list of facts, triples of the form entity - relation - entity, which relies on dependency parsing-based extraction patterns and language modeling; and methods for constructing a bipartite graph encoding the information contained in the set of facts and determining an appropriate traversing order on that structure. We evaluate the components of our framework on a subcollection extracted from MEDLINE, obtaining promising results.