Redução de características baseada em grupos semânticos aplicados à classificação de textos

Abstract

The classification of texts is a technique that aims to organize and categorize information, from textual documents present on digital platforms. Usually each of the words that constitute the documents are considered as a characteristic. This approach to simple textual representations is called Bag of Words. Although these characteristics are important for classifying documents, most of them are irrelevant and/or redundant, which causes problems such as high dimensionality, making categorization costly in terms of memory and execution. In order to reduce the large demand for computational resources, dimensionality reduction techniques are applied, such as the selection and transformation of characteristics. Characteristic selection is widely used in the literature because it has a lower computational cost compared to other techniques . In this technique, characteristics are selected without presenting modifications in the original characteristics, that is, the selection of a subset that contains only the most relevant characteristics of the original set occurs. In the transformation of characteristics occurs the formation of a new set of characteristics, this new set being smaller than the original set, containing new words caused by the combination or transformation of the original words. It is important to stress that both methods have some kind of loss of information. The objective of this work is to propose a new dimensionality reduction method that minimizes the loss of characteristic information from the creation of semantically related groups of words using clustering algorithms and Word Embeddings. With this, it is possible to reduce the amount of characteristics maintaining the semantics of each word. In this work the reduction occurred through the creation of semantic groups. Initially, the words in the databases were vectorized using Word2Vec and Glove methods. After the words were vectorized, the clustering algorithms were applied, creating smaller groups of characteristics in relation to the original groups. The method was applied in widely used databases in the literature, reaching good results, especially in more unstructured data, such as Web pages, news, Twitter posts, among others.