Automatic rhetorical parsing based on large amount of data.

Abstract

Discourse parsing is one of the essential tasks for Natural Language Processing area, being useful for both linguistic tasks (as the study of some particular phenomenon) and applied computational tasks (as text summarization and question answering). While for some languages there are good parsers, for Portuguese there is only a symbolic method based on discourse patterns for identifying discourse structures. This project aims at exploring discourse parsing methods, especially for Portuguese, in a hybrid approach in order to advance the state of the art. In particular, for producing good quality analyses, we aim at taking advantage of the vast amount of data in the web with the available already annotated corpora, using semi-supervised and non-supervised machine learning techniques, trying to overcome the traditional sparse data limitation. Among the several existent discourse models, the Rhetorical Structure Theory (RST) will be followed, since it is traditional in the area and used in several computational linguistic applications. Besides the practical contributions of creating and making available a new text analysis tool, we believe that this project has potential to produce good theoretical contributions, as the discourse parsing task modeling and the discourse learning strategy proposals.