Verificação de autoria em mensagens de texto utilizando grafos e aprendizagem de máquina.

Abstract

The search for feature extraction in texts is an area of interest in machine learning due to its numerous related possibilities, among them authorship verification is a relevant topic due to its applications and high complexity. In this context, this article uses data from Discord server chat messages with the purpose of automatically verifying the authorship of messages through supervised training. The process begins with preprocessing that aims to reduce noise and bias in the data, and then explores the learning model's ability to generalize by identifying unknown texts and classifying them as either authored or not. Graphs are used as feature extractors in text messages, leveraging artificial neural networks as machine learning models for classification. Words become nodes, and their edges capture the intensity related to the distance between terms in the sentence, resulting in the construction of a graph that represents an individual's vocabulary and aims to capture relevant characteristics in the text. By achieving good accuracies for true positives and true negatives when adjusting the activation threshold, the models can achieve satisfactory results with reduced training cost, allowing for easier exploration of new parameters.