MÉTRICAS E MODELOS DE NLP PARA COMPREENDER E REDUZIR A PROPAGAÇÃO DE DESINFORMAÇÃO

Abstract

This work investigates how Natural Language Processing (NLP) algorithms can be used to combat online misinformation, a phenomenon that threatens the quality of information and democracy. The aim is to analyze the potentialities of pysentimiento, an NLP library for Spanish, that offers functionalities such as sentiment analysis, emotions, hate speech and irony. The hypothesis is that these functionalities can contribute to the detection and reduction of the spread of fakenews on social media. To test this hypothesis, the work compares the performance off our deep neural network models, based on RoBERTa, on the task of sentiment analysis of tweets in Spanish. The results show that the model Robertuito-Base-Uncased outperforms the other models in terms of loss, f1, runtime and samples per second, being the most suitable for our proposal. The work concludes that pysentimiento is a promising tool for the fight against online misinformation, and suggests possible applications and improvements for future works.