Categorização de textos por aprendizagem de máquina

Abstract

Automatic categorization of texts into predefined classes has witnessed a growing interest in recent years, due to the increased availability of documents in digital format and the subsequent need to organize them. In the research community, the dominant approach to these problems is based on machine learning techniques: a general inductive process automatically builds a classifier learning from a set of pre-classified documents, the characteristics of categories. The advantages of this approach over the knowledge engineering approach (consisting of the manual definition of a classifier by domain experts) are very good effectiveness, considerable savings in terms of skilled workforce and direct portability to different domains. This research discusses the main approaches to text classification that fit the machine learning paradigm. We will discuss in detail issues relating to document representation, classifier construction, and classifier evaluation as well as the Deep Learning methods. Deep neural networks (DNNs) have revolutionized the field of Natural Language Processing (NLP). Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) are the two main types of widely exploited DNN architectures to handle various NLP tasks. CNN is supposed to be good at extracting position and variable resources and RNN in sequence unit modeling. The state of the art in many NLP tasks usually changes due to the battle of CNNs and RNNs. This work is a comparison between the classic machine learning models (Maximum Entropy Modeling, Support Vector Machine, Bootstrap Aggregating, Boosting, NNET Neural Networks, Random Forest, Scaled Discriminant Analysis, Decision Trees and Naïve Bayes) and this new approach which is state of the art using CNN and RNN networks, with the main objective of building the index of the journal IEEE Geoscience and Remote Sensing Letters, observing the performance and the performance of different models, the classification is performed from of two datasets (Title and Abstract) from IEEEGRSL articles. In contrast to the traditional methods, we introduced a recurring convolutional neural network for text classification from the Abstract of the journal’s articles by noting that classic models tend to lose accuracy when we increase the amount of data. The experimental results show that the proposed method had a satisfactory performance, but the RCNN network surpassed the classical methods in performance. However, when implementing this classification technique using Deep Learning, the hit ratios for the Abstract dataset surpassed the classic models implemented in this paper, reaching a precision of 94 % with a performance of 6 seconds.