Transferência de conhecimento com aprendizagem profunda para classificação de textos em língua portuguesa

Abstract

Approaches based on Machine Learning present computational solutions to various real-world problems that are difficult to solve through traditional programming. Among the several techni- ques of this area, we highlight the methods of the field of Supervised Learning, which are capable of generating models based on labeled data. In the classification task, these models are functions that map an object of the domain of interest to a class. In the learning process, classes are known in advance and identified by the data label. Some approaches within this field, with varying degrees of complexity, can be used for Classification of Texts. In this domain, models based on Deep Neural Networks (DNN), when compared to other techniques, present state-of-the-art results. A limitation of DNNs for generating robust models is the requirement for availability of a large amount of labeled data. When you do not have access to a data set large enough for a satisfactory learning, one way around this problem is to use Knowledge Transfer techniques. This is a prominent approach in the area of Artificial Intelligence, whose purpose is to use a generated model for a source problem as a starting point for learning a target model. In this work, we apply this technique to improve the classification performance of texts in Portuguese language. It investigates, through computational experiments, different models of DNNs and characterizes the conditions that guarantee the transfer of knowledge.