Análise comparativa de algoritmos de classificação de texto

Abstract

Text mining focuses on extracting information from unstructured data, such as natural language texts. Unlike traditional data mining, which works with structured data in tables, text mining deals with information that often does not follow a defined pattern. Text classification is an important task that aims to automatically categorize large volumes of textual data into predefined classes. Although classical algorithms such as Naive Bayes and Support Vector Machines (SVM) are widely used, they have limitations in capturing complex semantic nuances, as they analyze words in isolation and do not consider the full context. BERT (Bidirectional Encoder Representations from Transformers offers a powerful alternative to Natural Language Processing (NLP) by capturing the context of words bidirectionally, which improves the quality of textual representations. This advanced capability has made BERT a more interesting choice for several NLP tasks. Based on this, this work compares the performance of BERT with classical algorithms in text classification tasks, using four public databases (two in English and two in Portuguese) for analysis in different linguistic and complexity contexts with binary and multiclass classifications. The results show that BERT outperforms traditional algorithms in accuracy and F1 measure, although classical methods still perform well in simple tasks with less data. In addition to these metrics, it is relevant to consider the computational cost of BERT, which is significantly higher, with longer training times and greater resource demand. Therefore, the choice between BERT and classical algorithms must balance performance and available computational resources.