A Multimodal Approach to Identify Bias in Digital Social Media

• Grant number: 20/05173-4
• Support type: Regular Research Grants
• Duration: February 01, 2021 - January 31, 2023
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Cooperation agreement: MCTI/MC
• Principal researcher: Altigran Soares da Silva
• Grantee: Altigran Soares da Silva
• Home Institution: Instituto de Computação. Universidade Federal do Amazonas (UFAM). Ministério da Educação (Brasil). Manaus , SP, Brazil
• Assoc. researchers: André Luiz da Costa Carvalho ; Eduardo Freire Nakamura ; Fabiola Guerra Nakamura ; Tiago Eugenio de Melo

Abstract

An important but often overlooked problem in social network analysis is the presence of biases, whether intentional or not, introduced in the content published on social media. By making use of resources such as selective omission of information and choice of words, different sources or authors can transmit biased impressions about the same fact, in addition to contributing to the spread of misinformation through the individual or collective manipulation of unprepared individuals or without sufficient knowledge about the subject. In addition, the presence of bias can impact the way content consumers perceive events, political decisions and discussions related to different topics. Identifying the political or ideological bias in social content is a difficult task even for humans, given the high level of subjectivity involved. Proposals in the recent literature present methods mainly focused on textual content through sentiment analysis, with results that are, unfortunately, still below expectations. In this project, we will investigate a new multimodal approach to this problem. Specifically, we will use and combine, through data fusion techniques, several other aspects that, although successfully used in several other social network analysis problems, have been little explored in this specific problem. Among these aspects we highlight: relationships between content sources through citations; named entities often mentioned by sources; abstract topics implicitly and tacitly present in the content produced by the sources; in addition to the analysis of feelings to identify emphasis. Our hypothesis is that these aspects, when properly explored and correctly combined, can improve the results of the state of the art in detecting ideological bias in news portals and social networks. The project's research team was formed in order to bring together specialists in methods and techniques related to each of these aspects, and who also have extensive research experience on social network analysis and massive data processing (Big Data). Through experimental procedures involving collections of real data obtained from the Web or data generated synthetically with similar properties, we intend to evaluate the effectiveness of the methods, techniques and algorithms developed during the research and also their efficiency and scalability for real instances of the problem in question. (AU)