Multimodal approach for tension levels estimation in news videos

摘要

In this paper, we present a novel multimodal approach to estimate tension levels in news videos. The news media constitute a particular type of discourse and has become a central part of the modern-day lives of millions of people. In this context, it is important to study how the news industry affects human life and how it works. To support such a study, our approach estimates tension levels (polarities) along the news narrative, revealing the communication patterns used. To achieve this goal, we combine audio and visual cues extracted from news participants (e.g., reporters and anchors), by using methods for: (1) emotion recognition from facial expressions, (2) field size estimation and (3) extraction of audio features (e.g., chroma and spectral features), as well as textual cues obtained from the (4) sentiment analysis of the speech transcriptions. Experimental results with a dataset containing 960 annotated news videos from three Brazilian and one American TV newscasts show that our approach achieves an overall accuracy as high as 64.17% in the tension levels classification task. Those results demonstrate the high potential of our approach to be used by media analysts in several applications, especially, in the journalistic domain.