Análise de modelos de processamento de linguagem natural para classificação de notícias da UFRN

Abstract

Artificial Intelligence (AI) is transforming various sectors, reshaping what we understand as “technology” on a daily basis. Its application in natural language analysis is of utmost importance, with this field evolving exponentially every day. In this developmental context, this work aims to analyze and compare different approaches to Natural Language Processing (NLP) and text classification applied to the news found on the website of the Federal University of Rio Grande do Norte (UFRN). The classification is performed among the themes of “sciences,” “events,” “job openings,” and “announcements,” while also providing interfaces with the best models. The adopted methodology encompasses everything from the choice of technologies to the implementation of interfaces. In this sense, Data Version Control (DVC) was used for data storage and versioning during the development of the models, GitHub for code versioning, Weight and Biases for versioning and generating graphs related to model training, the data used alongside the interfaces, OpenAI’s API for embedding the texts, and Gradio, in conjunction with Hugging Face, to build and deploy the models. Finally, the results show that most models achieved high accuracy; however, the one utilizing the XGBoost algorithm delivered the best performance, successfully classifying natural language texts with precision. Considering that AI is continuously growing and new approaches are constantly emerging, the discussion on which is the best approach for a given context is indispensable.