Detecção automática de memes preconceituosos em português brasileiro.

Abstract

With the exponential growth of social media platforms, various types of content are shared daily, with memes standing out — multimodal artifacts typically composed of text and image. Memes serve as a fast form of communication to express ideas, critiques, humor, among others. However, their use has also promoted the spread of harmful content. Although significant advances have been made in recent years in the automatic detection of such content, most state-of-the-art solutions focus on memes in the English language. Due to their multimodal nature, the automatic analysis of memes is a complex problem, since understanding a meme's content is only possible when text and image are analyzed together, and external knowledge related to the semantic and cultural context in which it is embedded is also necessary. This thesis aimed to propose an approach for the automatic detection of prejudiced content in Portuguese-language memes, specifically Brazilian Portuguese. The proposed method is based on the application of knowledge transfer techniques through fine-tuning with reparameterization in an open-source Multimodal Large Language Model (MLLM). For training and validation purposes, a dataset of Portuguese-language memes was built and labeled. The results of the binary classification task (prejudiced meme versus non-prejudiced meme) demonstrated that detecting prejudiced memes in Brazilian Portuguese is a challenging task, even for state-of-the-art models, while also demonstrating the effectiveness of the adopted strategy by achieving competitive performance, with an F1 score of 0.81, when compared to models trained and evaluated on English-language memes. Additionally, an indirect comparison of the proposed approach with methods designed for other languages showed superior performance to approaches aimed at Tamil, Bengali, and Spanish. Relevant insights were also observed regarding the linguistic and cultural context of Brazilian Portuguese. Thus, this research contributed a solution to the problem of automatic identification of harmful content on social media in Brazilian Portuguese. Furthermore, the results presented promote discussion about the linguistic and cultural specificities of the target language in the studied problem.