Semi‐supervised learning for affect modelling

Abstract

The current dominant approach in Affective Computing when creating an Affect Model is to make use of a fully supervised learning algorithm, which would require a fully labelled corpus. Because of this, the field can be limited by the amount of labelled data that can be found, and if one does not exist, creating and annotating a new corpora is a tedious process. Creating a new dataset also comes with other issues, such as results being not accurate due to the annotator’s fatigue, which could heavily alter the validity of the annotations. The goal of this final year project is to explore a new alternative. This paper aims to investigate whether making use of algorithmically dissimilar semi‐supervised approaches for training an affect model could produce results that are equivalent to a set of models created using a fully supervised approach, following the current dominant approach. The aim of this paper is not to attain state‐of‐the‐art results, but rather to analyse the potential benefits of using such an approach in contrast to the more dominant approach. For this paper, five semi‐supervised algorithms will be implemented, and a number of tests will be conducted on them, using two publicly available affect corpora. The results of these tests would then be compared to a second set of tests conducted using three fully supervised approaches, following the standard approach, with the aim of them acting as a benchmark. Through the evaluation process, we could see that the semi‐supervised models would overall perform similar (if not sometimes better) when compared to the fully supevised counterparts, despite using less labelled data.