Análise exploratória e experimental sobre detecção inteligente de fake news

• Ano de defesa: 2020
• Autor(a) principal: Silva, Caio Vinícius Meneses
• Orientador(a): Rodrigues Júnior, Methanias Colaço
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Programa de Pós-Graduação: Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Notícias falsas; Eleições; Processamento eletrônico de dados; Mineração de texto
• Palavras-chave em Inglês: Fake news; Elections; Text mining
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO
• Link de acesso: https://ri.ufs.br/jspui/handle/riufs/14136

Context: The evolution of the media has contributed to the spread of false news, especially after the emergence of digital social networks. However, this practice is not a recent phenomenon in human history. Reports from the First World War period show the use of misleading advertising by the press, which culminated in new standards of objectivity and journalistic balance. In digital social media, this phenomenon, now called fake news, has found a new environment conducive to spreading worldwide, making it impossible to manually check this immense volume of data. In this context, work in several areas has been carried out in order to try to minimize the damage caused by the proliferation of fake news. Objective: The purpose of this work was to evaluate the effectiveness of the most used methods to check text correspondence, in the task of automatic detection of fake news about the Brazilian presidential elections of 2018, comparing the evidence found with the results obtained from a mapping of the state of art published in this research. Method: Initially, a systematic mapping was carried out to identify and characterize the main approaches, techniques and algorithms used, in computing, to detect false news. Finally, a controlled experiment was carried out, in vitro, using as perspective one of the works found in the literature, whose context has a strong relationship with this study: the American elections of 2016. In this way, the effectiveness of the methods was evaluated, comparing the results and contexts of the two works. Results: For the state of the art, it was identified that the main algorithms used in the task of detecting false news are LSTM (17.14%), Naive-Bayes and Similarity Algorithm (11.43% each). With the execution of the entire experimental process, it was evidenced that the TF-IDF and BM25 methods obtained statistically similar averages of accuracy, respectively, 79.86% and 79.00%. Finally, the Word2Vec and Doc2Vec methods also obtained, respectively, the worst averages, 75.69% and 72.39%. Conclusions: After analyzing the state of the art, gaps related to work in the Big Data context and the need for replication of existing studies, in the form of more controlled experiments, became evident. With the experimental evaluation, it was found that the effectiveness of the methods evaluated were similar to the effectiveness of the work used as a control. In addition, considering the universe of checked news available, the analyzed period and a margin of error of approximately 3.5%, the disclosure of fake news by the followers of both candidates evaluated in the experiment was evidenced. Followers of candidate Jair Bolsonaro (PSL) were responsible for 62.25% of tweets related to fake news, against 37.75% of followers of candidate Fernando Haddad (PT). With regard to accounts deleted from the social network in a short period of time, 59.96% were followers of the PSL candidate and 40.04% of followers of the PT candidate. The dissemination of fake news does not always imply intention, and may only imply greater engagement by some.