Análise sintática semi-supervisionada por constituência aplicada ao português e inglês

Abstract

This paper presents results on semi-supervised for Portuguese and English syntactic analysis. Our model induces grammatical structures through context modeling, span and label this triple(context,span and label) being used to induce the most likely constituents and labels of a test sentence. Our model of semi-supervised syntactic analysis induces grammars through the constituency of words. The model also presents results for syntactic tagging, because our model induces the constituents of sentences and their labels if they exist. As a result we obtained a measure-F for the Portuguese unlabelled overall average of 17.30% for the best results of corpus Tycho Brahe. Have we obtained results for the English average of 28.95% below the best possible result. For the syntactic tagging our model got 20.79% of the reach of sentences of each corpus of Portuguese, and among those he meets gets a hit of 1.09%. For labeling applied to the English got 9.61% of the sentences of each corpus were tagged, and this obtained a settlement of 1.71% of certain labels. This work presents results relevant to unlabeled corpus and also realize that the size of the corpus both training and testing negatively influence the results, ie if we buy the WSJ corpus Tycho Brahe and taking into account its size and results we see that if we have a large amount of training and testing our model shows lower results. For results in syntactic tagging realized that our model was not very effective because the model simply selects the label most likely rather than model it as a new dimension. We also presente results of the models proposed by Klein e Manning (2004) applied to the Portuguese, what results were not found in the bibliography.