Cross-Media Sentiment Analysis in Brazilian Blogs

Abstract

The use of social media is becoming highly present in our lives. It is through images, texts, and videos that human beings try to communicate in social networks and expose their opinions in the face of everyday events. Due to the increased volume of data transmitted over the Internet, it becomes difficult to do a human analysis of this content. For this reason, it is necessary to automate the task of classifying feelings’ polarity. Although the area of classification of feelings in images and texts is well developed and applied in social network context, the classification of feelings from images together with texts is still under development. A challenge is to build algorithms and methods that can infer feelings just like humans perceive it. Firstly, we present Cross-media Brazilian Blog corpus, the dataset we built based on BlogSet-BR, which goal is to have a data ground truth (based on subjects opinions) concerning feeling perceived in texts and images, when analyzed separately as well as when presented together. Therefore, we tested some available technologies to detect sentiment polarities in texts and images and compare with the ground truth. In addition, we pursue a research specifically on the contradiction posts, i.e. when image is positive and text is negative from the same blog. Results indicate that subjectivity affects emotional judgments because there are variances between cultures.