Superando o estado da arte na etiquetagem morfossintática por meio de regras de pós-etiquetagem

Abstract

According to Kveton e Oliva (2002) tagging errors, when a corpus is used in order to statistically train a natural language processing system, constitute deviations from regularities which the system is expected to learn, leading to a false model of the analyzed language.These authors propose a correction method for these errors in a post-tagging process, taking into account the identification of impossible n-grams in the language to be modeled. In this paper we try to systematize the errors made by the Brazilian Portuguese morphosyntactic tagger Aelius, built through the NLTK (BIRD, KLEIN and LOPER, 2009). The departing point was the compilation of computer mediated communication corpus. This genre is characterized by an abundance of abbreviations, spelling variation and standard language deviations. After the Aelius annotation was done, we listed the errors made and classified them. This was the basis for a set of rules for automatic tagging correction. Our goal was to reach a 99% accuracy, above the state of the art 97 (GÜNGÖR,2010,p.207), and in order to reach that we implement a system of Python rules to correct the errors created by the tagging system.