Redes neurais recorrentes para a classificação de estruturas retóricas

Abstract

Rhetoric is the art of using language effectively and persuasively. One important factor about rhetoric is its structure, i.e. the organization of how such arguments are presented in the text. This Master?s work intend to evaluate if recurrent neural networks are able to contribute in the classification of rhetorical structure. To do this, it was necessary first to create a corpus with a large number of abstracts of scientific articles with their rhetorical structures annotated so that it was possible to train the proposed neural network architecture. The proposed recurrent neural network (RNN) uses LSTM (Long Short-Term Memory) layers to avoid the problems that are common in RNNs (vanishing and exploding gradients). An encoder-decoder layer was also used to obtain intermediate representations of the abstracts sentences so that it was possible to transform the word sequence into the abstract sentence into a fixed-length vector representing that sequence. For comparison, two baselines were implemented using the CRF and SVM machine learning algorithms, and with features that were used in other works in the literature. The results obtained by the proposed neural network were satisfactory. When comparing the best results of the network with the best results of the baselines, it was possible to observe 3.18% gain in the accuracy of the classification of rhetorical structures in abstracts as a whole. In the sentence classification, gains of 1.02% in accuracy and 0.88