Identificação de desinformação em áudio dentro de grupos públicos do aplicativo whatsapp

Abstract

The spread of disinformation on social media is a growing concern, negatively impacting public perception and decision-making. So far, efforts to identify fake news have focused mainly on textual content, leaving a significant gap in the detection of disinformation in audio. This study proposes an approach to fill that gap by analyzing a corpus of audio content and applying it to various machine learning models, with the aim of effectively supervising the detection of disinformation in sound-based content. Initially, the audio files are transcribed using speech recognition technologies (speech-to-text), enabling textual analysis with techniques aimed at identifying fake news. The multimodal approach — integrating both audio and text — enhanced detection effectiveness by considering contextual and semantic aspects that might go unnoticed in purely textual analyses. Different vectorization combinations (BoW and TF-IDF) were evaluated, with variations of n-grams and preprocessing, using machine learning models such as SGD, SVM, Random Forest, among others. Performance was measured using metrics such as F1- score, precision, recall, and FPR. The best results were achieved with the SGD and Linear SVM models, particularly with TF-IDF combining unigrams, bigrams, and trigrams with preprocessing, reaching an F1-score of up to 0.579. This demonstrates a strong ability to identify both fake and truthful content, even in imbalanced class scenarios. The proposed methodology proved effective in detecting disinformation in audio, contributing relevant technological solutions to combat the spread of fake news on social media platforms.