Open Access

Lukas Stappen

• This thesis extends the emerging field of multimodal sentiment analysis of real-life videos, taking two components into consideration: the emotion and the emotion's target. The emotion component of media is traditionally represented as a segment-based intensity model of emotion classes. This representation is replaced here by a value- and time-continuous view. Adjacent research fields, such as affective computing, have largely neglected the linguistic information available from automatic transcripts of audio-video material. As is demonstrated here, this text modality is well-suited for time- and value-continuous prediction. Moreover, source-specific problems, such as trustworthiness, have been largely unexplored so far. This work examines perceived trustworthiness of the source, and its quantification, in user-generated video data and presents a possible modelling path. Furthermore, the transfer between the continuous and discrete emotion representations is explored in order toThis thesis extends the emerging field of multimodal sentiment analysis of real-life videos, taking two components into consideration: the emotion and the emotion's target. The emotion component of media is traditionally represented as a segment-based intensity model of emotion classes. This representation is replaced here by a value- and time-continuous view. Adjacent research fields, such as affective computing, have largely neglected the linguistic information available from automatic transcripts of audio-video material. As is demonstrated here, this text modality is well-suited for time- and value-continuous prediction. Moreover, source-specific problems, such as trustworthiness, have been largely unexplored so far. This work examines perceived trustworthiness of the source, and its quantification, in user-generated video data and presents a possible modelling path. Furthermore, the transfer between the continuous and discrete emotion representations is explored in order to summarise the emotional context at a segment level. The other component deals with the target of the emotion, for example, the topic the speaker is addressing. Emotion targets in a video dataset can, as is shown here, be coherently extracted based on automatic transcripts without limiting a priori parameters, such as the expected number of targets. Furthermore, alternatives to purely linguistic investigation in predicting targets, such as knowledge-bases and multimodal systems, are investigated. A new dataset is designed for this investigation, and, in conjunction with proposed novel deep neural networks, extensive experiments are conducted to explore the components described above. The developed systems show robust prediction results and demonstrate strengths of the respective modalities, feature sets, and modelling techniques. Finally, foundations are laid for cross-modal information prediction systems with applications to the correction of corrupted in-the-wild signals from real-life videos.…