Análise da informatividade do ambiente sintático na caracterização de entidades mencionadas no português brasileiro

Abstract

This dissertation investigates the degree of informativeness of the syntactic environment of named entities in Portuguese, deliberately abstracting away from lexical information. Drawing on the framework of Dependency Grammar under the Universal Dependencies standard, we represent sentences from two corpora, the Second HAREM Golden Collection and the Roda Viva Corpus, as dependency trees. These trees were traversed by breadth-first search algorithms taking the named entity as the central node, which allowed us to measure the syntactic distance of the elements that best characterize each entity and to compare the syntactic signatures of ten entity categories from the HAREM taxonomy. Results show that the syntactic environment is informative enough to constrain the occurrence space of an entity, but not to characterize it clearly. Categories such as Time, Organization, and Location displayed relatively discriminable syntactic profiles, whereas Person, Thing, Abstraction, and Other proved strongly dependent on lexical and semantic information. We conclude that pure syntax, even when richly abstracted into dependency relations and morphosyntactic categories, exerts a power of constraint over the occurrence of named entities, but not of definition.