Year: 2025
Type: dissertation
Abstract: The exponential growth of textual data in digital environments has driven the development of efficient techniques for automatic text classification. In this context, the appropriate choice of a vectorization technique is a determining factor for the performance of classification models. This work aims to analyze and compare the performance trends of different text vectorization approaches applied to the classification task, considering frequency-based methods such as Bag of Words and TF-IDF, models that explore se- mantic relations such as Word2Vec, Glove, and Doc2Vec, and recent transformer-based and deep learning techniques such as BERT, ALBERT, ROBERTA, E5, Instructor-XL, and GPT2. To this end, the IMDb Reviews and 20 Newsgroups textual datasets are used, along with natural language preprocessing techniques, covering distinct domains in order to ensure the generalization of the results and to allow an analysis of differences between datasets. The methodology involves the standardized application of each vector- ization technique to the preprocessed textual data, followed by the training of supervised classification models. The resulting vectors are used as input to traditional classification models, including Random Forest, LinearSVC, and Logistic Regression, enabling a uni- form comparative analysis among the different textual representations. The techniques are compared based on quantitative metrics, including primarily accuracy, F1-score, and processing time. The study aims to identify the advantages, trends, and limitations of each approach, considering factors such as predictive performance, computational complexity, and suitability to different types of textual data. The results are intended to support researchers and practitioners in selecting more effective textual representation methods for specific applications, promoting greater efficiency in automated text analysis systems, as well as mapping the aptitude of different approaches across the chosen datasets.
Author: Mello, Gustavo José da Silveira
Language: Portuguese
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7155582004