Multimodal Sentiment Analysis for Automatic Estimation of Polarity Tension of TV News in TV Newscasts Videos

• Moisés H. R. Pereira UNIBH
• Flávio L. C. Pádua CEFET-MG
• Adriano C. M. Pereira UFMG
• Giani David-Silva CEFET-MG
• Fabrício L. C. Benevenuto UFMG

Resumo

This paper presents a multimodal approach to perform contentbased sentiment analysis in TV newscasts videos in order to assist in the automatic estimation of polarity tension of TV news. The proposed approach aims to contribute to the semiodiscoursive study relative to the construction of ethos of those TV shows. In order to achieve this goal, it is proposed the application of computational methods of state-of-the-art that, through the processing of newscasts’ videos of interest, perform the automatic emotion recognition in facial expressions. Moreover, they extract modulations in the participants’ speech (e.g., news anchors, reporters, among others) and apply sentiment analysis techniques in their text obtained from closed caption, therefore making possible to estimate the emotional tension level in the enunciation of the TV news. In order to evaluate the accuracy and the applicability of the system, we use an actual dataset composed by 358 videos from three Brazilian newscasts. The experimental results are promising, which indicate the potential of the approach to support the analysis of TV newscasts discourse.