The objectives of this thesis are focused on research in machine learning for\udcoreference resolution. Coreference resolution is a natural language processing\udtask that consists of determining the expressions in a discourse that mention or\udrefer to the same entity.\udThe main contributions of this thesis are (i) a new approach to coreference\udresolution based on constraint satisfaction, using a hypergraph to represent\udthe problem and solving it by relaxation labeling; and (ii) research towards\udimproving coreference resolution performance using world knowledge extracted\udfrom Wikipedia.\udThe developed approach is able to use entity-mention classi cation model\udwith more expressiveness than the pair-based ones, and overcome the weaknesses\udof previous approaches in the state of the art such as linking contradictions,\udclassi cations without context and lack of information evaluating pairs. Furthermore,\udthe approach allows the incorporation of new information by adding\udconstraints, and a research has been done in order to use world knowledge to\udimprove performances.\udRelaxCor, the implementation of the approach, achieved results in the\udstate of the art, and participated in international competitions: SemEval-2010\udand CoNLL-2011. RelaxCor achieved second position in CoNLL-2011.
展开▼
