Cross-Media Sentiment Analysis on German Blogs

Abstract

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.

Investigating sentiments in Brazilian and German Blogs

“…The sentiment analysis performed by Dal Molin et al [2019] takes up the topic of cross-media retrieval, investigating images and texts of Brazilian blog posts and associated contradictions. In a recent paper Zahn et al [2021], we provided the same study on German blogs and provide some comparison with Brazilian Blogs Dal Molin et al [2019]. In the field of elements that make up an image, in the present work, we researched the faces to assess the emotions that are transmitted and verify whether the facial expression could contribute to alleviating contradictions between the image and text domains.…”

“…They build a base in BlogSet-BR called Cross-Media Brazilian Blog (CBB), which is based on people's opinions, containing the sentiments perceived in texts and images, when analyzed separately and also when presented together. Zahn et al [2021] also performed this study, but in German. They create the Cross-Media German Blog (CGB) corpus which is based on the compilation of German blog posts collected on DWDS.de, published by the Berlin-Brandenburgischen Akademie der Wissenschaften.…”

“…In the present work, we used the corpus called Cross-Media German Blogs (CGB), presented in Zahn et al [2021], composed of 950 blog posts. CGB was created with an algorithm based on a dataset obtained in DWDS.de.…”

“…In this paper, we extend the discussion of sentiment contradiction by including a new feature in the image, i.e., the study of the sentiment of faces. While in previous work proposed by Zahn et al [2021] and Dal Molin et al [2019], we discussed the contradictions of the multimodal information, including text and images, in this paper we are interested in the study of a relevant component in the images: the human faces. Our main question here is to understand if the polarity of the image has any agreement with the emotion expressed in the faces present in the images.…”

“…Our main question here is to understand if the polarity of the image has any agreement with the emotion expressed in the faces present in the images. As mentioned before, in this paper we use the two datasets that contain images and texts in the German language (CGB) Zahn et al [2021] and the Cross-Media Brazilian Blog (CBB) Dal Molin et al [2019], however in order to answer our main question, we work only with images in order to detect faces and expressed emotions for comparison with images polarities as defined in the ground truth.…”