Análise multimodal em blogs brasileiros

Abstract

The use of social media is increasingly present in our lives. It is through images, texts and videos that humans try to communicate on social networks and expose their opinions in the face of everyday events. Due to the increased volume of data transmitted over the Internet, it is difficult to perform a human analysis of the media without the use of computer resources. Scientific communities, with various motivations, such as: analyzing feelings in text, in images, detecting opinions in blogs, among others, feel challenged to discover characteristics to be extracted from these contents, being an example of the analysis of emotions in blogs. Although the area of classification of feelings through texts and images is under development, there are still several challenges. The main challenge is to build algorithms and methods that can infer subtle and subjective feelings as humans perceive them. This paper presents the corpus Cross-media Brazilian Blog, a dataset that was built based on BlogSet-BR. In addition, it was built the Ground Truth of these data (based on the opinions of subjects) about the feelings perceived in the texts and images of these blogs, which in this work become available for use. Some technologies used to predict sentiment in text and images have been tested in the Cross-Media Brazilian Blog corpus and compared with Ground Truth. In addition to the analyzes performed on the texts, a research was conducted specifically on contradictory posts, i.e. when the image is positive and the text is negative, or vice versa, when present on the same blog. Results indicate that methodologies for detecting feelings in blogs can be customized to detect conflicting posts and be able to better identify feelings in social media posts.