Emotion Prediction with Weighted Appraisal Models - Validating a Psychological Theory of Affect

Abstract

Appraisal theories are a prominent approach for the explanation and prediction of emotions. According to these theories, the subjective perception of an emotion results from a series of specific event evaluations. To validate and extend one of the most known representatives of appraisal theory, the Component Process Model by Klaus Scherer, we implemented four computational appraisal models that predicted emotion labels based on prototype similarity calculations. Different weighting algorithms, mapping the models input to a distinct emotion label, were integrated in the models. We evaluated the plausibility of the models structure by assessing their predictive power and comparing their performance to a baseline model and a highly predictive machine learning algorithm. Model parameters were estimated from empirical data and validated out-of-sample. All models were notably better than the baseline model and able to explain part of the variance in the emotion labels. The preferred model, yielding a relatively high performance and stable parameter estimations, was able to predict a correct emotion label with an accuracy of 40.2% and a correct emotion family with an accuracy of 76.9%. The weighting algorithm of this favored model corresponds to the weighting complexity implied by the Component Process Model, but uses differing weighting parameters.