A framework for emotion and sentiment predicting supported in ensembles

• Autor(a) principal: Novais, Rui Miguel Boneco
• Data de Publicação: 2022
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10400.1/19247
• topic: Emoções faciais; Emoções da fala; Emoções do texto; Sentimento; Ensembles; Aprendizagem automática; Domínio/Área Científica::Engenharia e Tecnologia::Outras Engenharias e Tecnologias

Humans are prepared to comprehend each other’s emotions through subtle body movements or facial expressions; using those expressions, individuals change how they deliver messages when communicating between them. Machines, user interfaces, or robots need to empower this ability, in a way to change the interaction from the traditional “human-computer interaction” to a “human-machine cooperation”, where the machine provides the “right” information and functionality, at the “right” time, and in the “right” way. This dissertation presents a framework for emotion classification based on facial, speech, and text emotion prediction sources, supported by an ensemble of open-source code retrieved from off-the-shelf available methods. The main contribution is integrating outputs from different sources and methods in a single prediction, consistent with the emotions presented by the system’s user. For each different source, an initial aggregation of primary classifiers was implemented: for facial emotion classification, the aggregation achieved an accuracy above 73% in both FER2013 and RAF-DB datasets; For the speech emotion classification, four datasets were used, namely: RAVDESS, TESS, CREMA-D, and SAVEE. The aggregation of primary classifiers, achieved for a combination of three of the mentioned datasets results above 86 % of accuracy; The text emotion aggregation of primary classifiers was tested with one dataset called EMOTIONLINES, the classification of emotions achieved an accuracy above 53 %. Finally, the integration of all the methods in a single framework allows us to develop an emotion multi-source aggregator (EMsA), which aggregates the results extracted from the primary emotion classifications from different sources, such as facial, speech, text etc. We describe the EMsA and results using the RAVDESS dataset, which achieved 81.99% accuracy, in the case of the EMsA using a combination of faces and speech. Finally, we present an initial approach for sentiment classification.