Bacharelado em Sistemas de Informação (UAST)
URI permanente desta comunidadehttps://arandu.ufrpe.br/handle/123456789/2935
Siglas das Coleções:
APP - Artigo Publicado em Periódico
TAE - Trabalho Apresentado em Evento
TCC - Trabalho de Conclusão de Curso
Navegar
3 resultados
Resultados da Pesquisa
Item Analisando a opinião de usuários do Twitter em relação ao lockdown no Brasil: um estudo frente à mídia convencional(2022-05-23) Silva, Paulo Alves da; Belo, Ítalo César de Souza; Oliveira, Marcelo Iury de Sousa; http://lattes.cnpq.br/2328386382232459; http://lattes.cnpq.br/3240928244006916; Não localizadoIn the pandemic scenario, several countries sought alternatives to reduce the high rates of contagion and mortality caused by COVID-19. Among the means used was lockdown, which is characterized as the most stringent measure of social isolation. Although lockdown is considered by many to be an excellent tool for health protection, the adoption of this measure divides opinions. In this sense, sentiment analysis emerges as a mechanism that can be used to understand the population's opinion about the lockdown. The social impact caused by this measure contributes to the emergence of articles from the most diverse media vehicles, which can strongly influence the critical thinking of the population. In this context, the present work performed a sentiment analysis to understand how Twitter users reacted to the lockdown implementation in Brazil during the COVID-19 pandemic. In addition, the results of the Twitter analysis were compared with the content coming from the gray literature, focusing on the articles authored by the conventional media, which addressed information about the theme. To carry out this research, data were extracted from the social network Twitter and processed through an ensemble composed of the following machine learning algorithms: Multinomial Naïve Bayes, SVM, Random Forest, MLP neural network, and the Logistic Regression classifier. In addition, Active Learning techniques, through the uncertainty approach, were combined with the ensemble to obtain higher accuracy of the classified data. According to the results, an interesting predominance of sentiments in favor of lockdown was noticed, as well as conflicts between the data obtained by the Twitter sentiment analysis process and information published in the gray literatureItem Classificação automática de discursos de ódio em textos do twitter(2019) Nascimento, Robson Murilo Ferreira do; Souza, Ellen Polliana Ramos; http://lattes.cnpq.br/6593918610781356; http://lattes.cnpq.br/8962852253787699Hate Speech can be defined as any communication that denigrates a person by their ethnicity, race, religion, sexual orientation, nationality or other characteristics. This behavior is becoming increasingly common in social networks, where many people confuse freedom of expression with intolerance. Young people are the main users affected since they represent a portion which might be easier to be influenced by the ideology propagated by haters, which in turn spread violence, racism, xenophobia, religious intolerance, etc. The tool that might help to handle this issue is Text Mining, which is capable of capture patterns or trends of texts in natural language. This task can be defined as a method of extracting relevant information in unstructured databases or semi-structured. Given that the Twitter is one of the most used social networks in Brazil, this work aims to implement and evaluate supervised machine learning techniques in order to automatically identify hate speech in tweets. With that in mind, we build a corpus with data collected from Twitter and part of it is manually annotated and subsequently preprocessed so we can obtain the ground truth collection used for training and evaluation of the supervised models. Finally, we conduct a comparison between machine learning algorithms, namely the SVM, Naive-Bayes and Logistic Regression. Later, we identify the best model under the described domain.Item Comitê de classificador para mineração de opinião de eleitores brasileiros(2019) Souza, Guilherme Henrique Pereira de; Souza, Ellen Polliana Ramos; http://lattes.cnpq.br/6593918610781356; http://lattes.cnpq.br/2474970020729937The recent and exponential growth of social media and user-generated content on the Internet, provide a wealth of data to uncover the feelings, opinions, and experiences of users and customers. Among the various social media currently available, Twitter stands out, with more than 336 million monthly active users. Understanding what people are thinking or their opinions is critical to decision making, especially in the context in which these people express their comments voluntarily. For this type of analysis, opinion mining is used, also known as sentiment analysis, which analyzes people’s opinions, feelings and emotions about entities such as services, products, brands, individuals, problems, events and topics, be expressed in an unstructured form, as in texts, for example. The spread of social media also raises the possibility of delving deeper into the web to explore and track citizens’ political and electoral preferences. In this sense, recent studies explore social media as a device to assess the popularity of politicians during the campaign. Analyzing social media during an election campaign can be a useful supplement or supplement to traditional surveys, and among the reasons include the low cost and speed of real-time analysis. This paper aims to analyze the efficiency of a ensemble for opinion mining applications, using data extracted from Brazilian voters. In this study, an application of opinion mining was developed at the document level using ensembles, with the objective of analyzing the efficiency of the ensembles in such applications. For that, data extracted from Twitter containing opinions of Brazilian voters were used. Six pre-processing configurations and three ensemble approaches were evaluated: Voting, Bagging and Boosting. The results were obtained in terms of Accuracy. According to the experiments, configuration 3, which uses stopwords, n-gram and TF-IDF, achieved the best results, with a 90.26% accuracy for the committee that used the Boosting approach.