Social interactions have changed in recent years. People post their thoughts, opinions and feelings on social media platforms more often. Due to the increase in the amount of data on the internet, it is impracticable to carry out the sentiment analysis manually, requiring automation of the process. In this work, we present the corpus Cross-Media German Blog (CGB) which consists of German blogs with feelings in the domain of images, texts and posts (Ground Truth), classified according to human perceptions. We apply existing Machine Learning technologies and lexicons to the corpus to detect the feelings (negative, neutral or positive) of the images and texts and compare the results with the GT. We examined contradictory posts, when the image and text classified by humans in the same post had diverging feelings. The comparison of this article with the analysis of sentiment among the media of Brazilian blogs finds its justification for performance results in cultural differences, since, throughout this work, Brazil is classified as indulgent and Germany as a restrained country.

No abstract

The sense of strangeness (or discomfort) perceived in certain virtual characters, discussed in Uncanny Valey (UV) theory, can be a key factor in our perceptual and cognitive discrimination. Understanding how this strangeness happens is essential to avoid it in the process of modeling virtual humans. In this paper, we investigate the relationship between images features and the discomfort that human beings can perceive. We extract image features based on Hu Moments (Hum) and Histogram Oriented Gradient (Hog). The saliency detection is also extracted in the specific parts of the virtual face. Finally, a model using Support Vector Machine (SVM) to provide binary classification is suggested. The results indicate accuracy of around 80% in the image estimation process comparing with subjective classification. As a contribution, some areas may benefit from this study for avoiding the creation of characters that may cause strangeness, such as the games, conversational agents and cinema industry.

No abstract

Social interactions have changed in recent years. People post their thoughts, opinions and sentiments on social media platforms more often, through images and videos, providing a very rich source of data about population of different countries, communities, etc. Due to the increase in the amount of data on the internet, it becomes impossible to perform any analysis in a manual manner, requiring the automation of the process. In this work, we use two blog corpora that contain images and texts. Cross-Media German Blog (CGB) corpus consists of German blog posts, while Cross-Media Brazilian Blog (CBB) contains Brazilian blog posts. Both blogs have the Ground Truth (GT) of images and texts feelings (sentiments), classified according to human perceptions. In previous work, Machine Learning and lexicons technologies were applied to both corpora to detect the sentiments (negative, neutral or positive) of images and texts and compare the results with ground truth (based on subjects perception). In this work, we investigated a new hypothesis, by detecting faces and their emotions, to improve the sentiment classification accuracy in both CBB and CGB datasets. We use two methodologies to detect polarity on the faces and evaluated the results with the images GT and the multimodal GT (the complete blog using text and image). Our results indicate that the facial emotion can be a relevant feature in the classification of blogs sentiment.

No abstract