Mejorando la extracción automática de relaciones biomédicas usando diferentes características lingüísticas de los textos

Abstract

Extracting biomedical relations from texts is a relatively new, but rapidly growing researchfield in natural language processing (NLP). Due to the increasing number of biomedicalresearch publications and the key role of databases of biomedical relations in biological andmedical research, extracting biomedical relations from scientific articles and text resourcesis of utmost importance.Drug-drug interactions (DDI) are, in particular, a widespread concern in medicine, and thus,extracting this kind of interactions automatically from texts is of high demand in BioNLP. Adrug-drug interaction usually occurs when one drug alters the activity level of another drug.According to the reports prepared by the U. S. Food and Drug Administration (the FDA) andother acknowledged studies [1], over 2 million life-threatening DDIs occur in the UnitedStates every year. Many academic researchers and pharmaceutical companies havedeveloped relational and structural databases, where DDIs are recorded. Nevertheless,most up-to-date and valuable information is still found only in unstructured research textdocuments, including scientific publications and technical reports.In this thesis, three complementary, linguistically driven, feature sets, are studied: negation,clause dependency, and neutral candidates. The ultimate aim of this research is to enhancethe performance of the DDI extraction task by considering the combinations of theextracted features with well-established kernel methods...