Please use this identifier to cite or link to this item: http://hdl.handle.net/10609/149053
Title: Modeling of emotional subjectivity in affective-based predictive systems
Author: Hayat, Hassan  
Director: Lapedriza, Agata  
Ventura, Carles  
Abstract: One of the goals of affective computing is to develop affective technologies that can understand humans emotionally and make their life better. Human emotions are highly subjective in nature. This is why systems that consider affective along with subjective information play a significant role not only in mimicking an individual's cognitive process but also in an individual's interaction with others. This thesis targets emotional subjectivity in affect-related tasks. In particular, this thesis studies subjectivity from two different perspectives: (I) subjectivity in the annotations, and (II) subjectivity according to personality traits. Regarding annotations, in supervised machine learning, affective systems are trained and tested on annotated datasets. Usually, these annotations are the aggregation of multiple subjective annotations which basically represent each annotator's subjective emotional perception. The common practice to get aggregated annotations is by computing the average score and majority voting of multiple subjective annotations. These aggregated labels lose subjective information. Systems that are trained and tested based on these aggregated annotations have poor generalization capabilities for predicting subjective emotional perception. To tackle this problem, we proposed a Multi-Task (MT) learning approach that has the capability to learn each subjective emotional perception available in the annotations separately. The results show that our MT approach (that considers all subjective annotations separately) has more generalization capabilities as compared to approaches that are trained only on aggregated annotations. The second part of the thesis presents the study in the context of dialogues. Concretely, we studied the problem of predicting subjective emotional responses for the upcoming utterance with respect to each speaker in the conversation. We developed a Multi-Task (MT) learning approach that has the capability to predict multiple subjective emotional responses in the conversation using the personality information of each speaker. The results show that separate modeling of each speaker's emotional responses using joint modeling (i.e. Multi-Task learning) is better than combined modeling of all speakers' emotional responses.
Keywords: predictive modeling
statistical learning
deep learning
subjective emotional perception
affective computing
Document type: info:eu-repo/semantics/doctoralThesis
Issue Date: 19-Apr-2023
Publication license: http://creativecommons.org/licenses/by-nc-nd/3.0/es/  
Appears in Collections:Tesis doctorals

Files in This Item:
File Description SizeFormat 
Tesi_Hassan_Hayat.pdfHayat_dissertation43,28 MBAdobe PDFThumbnail
View/Open
Share:
Export:
View statistics

This item is licensed under aCreative Commons License Creative Commons