Human Emotion Experiences Can Be Predicted on Theoretical Grounds: Evidence from Verbal Labeling

Abstract

In an effort to demonstrate that the verbal labeling of emotional experiences obeys lawful principles, we tested the feasibility of using an expert system called the Geneva Emotion Analyst (GEA), which generates predictions based on an appraisal theory of emotion. Several thousand respondents participated in an Internet survey that applied GEA to self-reported emotion experiences. Users recalled appraisals of emotion-eliciting events and labeled the experienced emotion with one or two words, generating a massive data set on realistic, intense emotions in everyday life. For a final sample of 5969 respondents we show that GEA achieves a high degree of predictive accuracy by matching a user’s appraisal input to one of 13 theoretically predefined emotion prototypes. The first prediction was correct in 51% of the cases and the overall diagnosis was considered as at least partially correct or appropriate in more than 90% of all cases. These results support a component process model that encourages focused, hypothesis-guided research on elicitation and differentiation, memory storage and retrieval, and categorization and labeling of emotion episodes. We discuss the implications of these results for the study of emotion terms in natural language semantics.

Nonlinear Appraisal Modeling: An Application of Machine Learning to the Study of Emotion Production

“…A repeated measurements protocol would have been required to estimate each respondent's individual bias. Thirdly, the GEA questionnaire allowed participants the opportunity to provide not just one but up to two emotion labels to characterize the emotion they experienced, which more than 70 percent of respondents did [34]. This made the final data more heterogeneous than the primary label alone would suggest.…”

“…Literature in machine learning has generally not focused on models for multiple-response multiclass prediction. 4 Note also that, in the original Scherer study [36] as well as the recent paper on the theoretical prediction model of the GEA system [34], the theoretical model was allowed to make two guesses rather than one, and calculated an unbalanced estimate of accuracy. At two guesses, the theoretical model is about as good as cluster prediction in the current study (i.e., predicting happiness, anger, shame/guilt, or distress).…”

“…Data were obtained from a recent study on appraisal and emotion [34]. In this study, 6,034 users responded to a webbased expert system called the Geneva Emotion Analyst (GEA; see [36], for an earlier application of this system).…”

“…To achieve this, we obtained a large data set ðN ¼ 5901Þ from a web-based expert system on recalled emotion episodes [34], with 25 appraisal predictors and a categorical emotion response (12 levels). The goal of the analysis was threefold:…”

“…In total, 14 different models were fitted to the training data, four of which were linear and 10 of which were nonlinear ( Table 1). The 14 models were selected to cover a wide range of different tools for machine learning, including traditional approaches (linear discriminant analysis, multinomial logistic regression), pure black-box learners (artificial neural networks, support vector machines), hierarchical models (classification tree, multivariate adaptive regression splines), and theoretical prototypes (Scherer prototypes; reapplied from [34]). A detailed discussion on the mathematical properties of these models is beyond the scope of this paper, but an extensive treatment can be found in [13].…”