Contribution
A data-driven mapping between discrete emotions and the full componential space, from reports of experienced (not semantic) emotion. It extends a conference paper of the same group.
Key insight
Discrete emotion ratings are recoverable from componential descriptors. The feeling items alone carry little more than valence and arousal, and the other components add what discriminates emotions of similar valence and arousal.
Assumptions
- That self-report after a clip captures appraisal, motivation, physiology and expression. The authors concede that self-report of physiology and expression “may not be very accurate and objective recording are more desirable” (p. 12).
- That a passive viewer’s appraisals can be probed with CoreGRID items. Items such as “it was caused by my own behavior” were removed for having no variance in that setting (§III.A).
Key results
- Hierarchical clustering of the emotion centroids (equation 1, weighted by each discrete-emotion rating) splits positive from negative, then happiness, serenity, distress, anger and surprise.
- Six varimax principal components explain about 59% of the variance; the first (30%) is action tendency, the second (14%) pleasant feeling.
- Expression is strong for most emotions but weak for sadness, surprise, disgust and calm. Appraisal is the most informative component for surprise and disgust, physiology for fear and anxiety.
- Nonlinear classifiers (random forest, RBF SVM) did not beat logistic regression significantly.
- Adding personality scores to the components did not improve prediction.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Emotional experience needs more than valence and arousal to describe | moderate | six PCA dimensions; agrees with GRID work |
| C2 | All components together predict discrete emotions better than any one | moderate | cross-validated ordinal regression, Table II |
| C3 | Personality shapes componential responses more than feelings | weak | uncorrected partial correlations, p < 0.01 |
Concepts
- componential profile: the weighted mean of CoreGRID ratings for a discrete emotion.
- MAE^M: macro-averaged mean absolute error, for ordinal ratings with imbalanced classes.
Connections
Cites Meuleman and Rudrauf (LIT-1440) as the only prior study of the same kind, and contrasts its active induction with this study’s passive one. Its film set and items were reused in the fMRI study of the same group (LIT-1444, NOTE-1107).
Bearing on the record
A per-episode measure of synchronization? No, and candidly the paper is further from one than the owner’s query hoped. It never uses the word. Each episode is a whole clip, about two minutes, rated once after viewing, and the authors name the problem themselves: “First, the emotional assessments for each clip is done at global level which neglects the presence of different events in the clip with potentially different emotional” content (p. 11). The coherence it reports, clusters and principal components, is computed across 1,567 surveys. It is the structure of the population’s ratings, not coordination within anyone’s episode.
Signed, or which tendency prevails? Partly signed: its first principal component is action tendency and its second pleasant versus unpleasant feeling. But nothing says which tendency won a conflict. Items such as “want the situation to continue” and “feel the urge to stop what was happening” are rated separately and could both be high, which is the ambivalent case QUESTION-268 needs separated from neutrality. The paper does not look at such co-occurrences.
Anything for QUESTION-266? Little. The feeling items (“feel good”, “feel bad”) are retrospective reports, which the question rules out for validating its third layer, and they are not a hedonic marker independent of choice. Its sequel (NOTE-1107) is where this design produces something the record can use.
Filed because the owner asked, and because the sequel cannot be read without it.
Limitations
Passive, vicarious induction; global ratings per clip; crowdsourced online data; all components self-reported; no correction for multiple comparisons in the personality analysis. The paper’s own second hypothesis (personality effects beyond valence and arousal) is tested only by correlation.
Open questions
None for the record beyond those its sequel takes up.