Study at a glance
- Authors: Pilar Ferré, Marc Guasch, Hans Stadthagen-González, José Antonio Hinojosa, Isabel Fraga, Javier Marín, Miguel Ángel Pérez-Sánchez
- Year: 2024
- Journal: Emotion
- DOI: 10.1037/emo0001300
- N: 651
- Design: Normative rating study; 651 native Spanish speakers rated 1,286 candidate emotion words on nine variables (>= 20 raters per word per variable); stepwise multiple regression and exploratory factor analysis
Key findings
- Of the five components of emotion posited by the multicomponential consensus (appraisal/evaluation, action tendency, body expression, physiological reaction/interoception, feeling), only feeling and interoception substantially predict how prototypically ‘emotional’ a word is judged to be.
- Interoception alone accounts for R2 = .417 of prototypicality variance — entered first by the stepwise model and by far the largest single increment; feeling adds .105 on top of it. By standardized beta the order reverses (feeling .494, interoception .363), which is why the paper’s abstract names feeling first.
- Body expression, evaluation and action — the visible, outward components — do not make a word count as an emotion. Action (beta = -.064) and morality (beta = -.137) are negative predictors, and body expression drops out of the model entirely.
- Feeling is not reducible to valence + arousal: feeling’s partial correlation with valence was nonsignificant (-.043) and with arousal small (-.11), against the constructionist expectation that feeling just is core affect.
- EFA on 13 variables yields two near-independent factors (interfactor r = .186): ‘socioemotional polarity’ (evaluation .938, valence .886, morality .872) and ‘emotional experience’ (interoception .768, feeling .762, prototypicality .760). Prototypicality loads on the second, not the first.
- The 1,283-word semantic space shows no discrete clusters — a continuous, roughly circular scatter — and a preponderance of negative over positive words, with positive words more homogeneous in emotional charge.
- Converges with Connell et al. (2018), who independently found interoception outranked all five exteroceptive modalities in distinguishing emotion words from nonemotion abstract words.
Limitations
- Interoception is the least reliable variable in the study by the authors’ own validity check. Cross-questionnaire agreement on control words was r(22) = .654 for interoception versus .824-.955 for all eight other variables, and the men/women rating correlation was .463. The construct carrying the paper’s headline claim is the one raters agreed on least.
- Ratings are semantic judgments about words, not measurements of anyone’s body. The study establishes what the concept of emotion contains, not what emotions do — a distinction the discussion does not always hold.
- The discussion misreads its own morality coefficient. Table 4 gives morality beta = -.137, but the text concludes it is ‘more probable that an abstract word refers to an emotion when it has to do with mental activities, judgments, and moral rules.’ The sign says the opposite for morality (thought, beta = .123, is the half that holds).
- Table 4 prints thought as beta = .123 with t = -4.915 — a sign mismatch; the partial correlation (+.140) indicates the t is the typo. The first regression is reported as F(7, 1270) despite entering 12 predictors.
- The feeling-is-not-core-affect argument rests on partial correlations controlling for all other variables, including evaluation, which correlates .712 with valence — a suppression risk the authors do not address.
- Sample is 85.3% women, mean age 22.2, almost entirely Spanish university students; single language, so the lexicon’s structure may be Spanish-specific. The authors concede the gender imbalance and report no gender effects in LMM analyses.
- Not preregistered; design and analysis exploratory.
Connections
- Methods: Emotion-word rating norms (multidimensional semantic norming)
- Concepts: Emotion prototypicality, Interoception, Core affect, Bodily sensation maps (bodily fingerprints of emotion), Basic emotions and natural kinds, Cognitive appraisal (of emotion), Interoceptive taxonomy (Farb et al. Box 1)
- Debates: Should emotion terms refer to feelings or to motive states?, Are emotions culturally universal?, Do discrete emotions map onto distinct brain regions or networks?
- Related studies: Nummenmaa et al. (2014), Volynets et al. (2020), Lyons et al. (2021), Scarantino (2018), LeDoux (2012), Lindquist et al. (2012)
A psycholinguistic normative study that arrives at this wiki from an unexpected direction and answers a question none of its other sources ask: what do ordinary speakers mean by “emotion”?
651 Spanish speakers rated all 1,286 words from the Pérez-Sánchez et al. (2021) Emopro database on nine variables (at least 20 raters per word per variable), five of them chosen to operationalize the five components of the multicomponential consensus — evaluation (appraisal), action (action tendency), body expression, interoception (internal body sensations), and feeling — and four chosen from the concreteness/abstractness literature (sensory experience, social interaction, thought, morality). Those ratings were regressed onto the database’s existing emotion prototypicality scores: how strongly each word “refers to an emotion,” from anger (high) to boredom (low). See Emotion-word rating norms (multidimensional semantic norming), Emotion prototypicality.
The instruction that matters is one sentence long: “Rate whether the word is related to internal body sensations.” That is the whole interoception manipulation, and it carries the paper.
The result
Five components went in. Two came out.
| component | variable | partial r with prototypicality | beta |
|---|---|---|---|
| feeling | feeling | .408 | .494 |
| physiological reaction | interoception | .340 | .363 |
| — | thought | .140 | .123 |
| appraisal | evaluation | −.044 (ns) | −.119 |
| action tendency | action | −.076 | −.064 |
| body expression | body expression | −.034 (ns) | dropped |
The stepwise model reached R² = .582, and the order of entry tells a different story than the betas: interoception went in first and took .417 of the variance by itself — more than everything else combined. Feeling added .105 on top. The abstract names feeling first because its standardized beta is larger; the variance decomposition names interoception first. Both are in the paper and they are not in conflict, but the wiki should hold the stronger version: the single best predictor of whether a Spanish speaker thinks a word names an emotion is whether they think it names something felt inside the body.
What did not predict emotionhood is as interesting. Body expression — the face, the voice, the posture, the entire Ekman evidence base — contributes nothing once feeling and interoception are in the model. Action tendency is a small negative predictor. Evaluation, the appraisal component, is negative too. Words like respeto (respect, high evaluation), impulso (impulse, high action) and pereza (laziness, high body expression) are simply not good examples of emotions to the people who use the word.
Why this belongs in an interoception wiki
Three things, in ascending order of consequence.
1. Independent convergence on interoception as the modality of emotion. Connell, Lynott & Banks (2018) — a different lab, a different language, a different question (the perceptual grounding of abstract concepts) — found that interoception ratings outranked hearing, taste, touch, smell and vision in distinguishing emotion words from nonemotion words, with the pattern reversed for concrete words. Ferré et al. reproduce that from the prototypicality side. Two independent routes to the claim that interoception is the sensory modality of the emotion concept, which is a stronger position than either paper alone.
2. It gives Should emotion terms refer to feelings or to motive states? its first usage data. LeDoux’s argument for reserving “emotion” for feelings is sociological: common sense identifies emotions with feelings so firmly that any other usage will be misread. Scarantino replies that common sense is “all over the map” — folk psychology routinely treats emotions as motive causes (“Maria slapped Lucas because she was angry”) — and the wiki’s debate page notes that this reply is offered without usage data. Here is usage data, and it does not favour Scarantino: feeling is the top predictor, action is negative, and the folk category tracks LeDoux’s explanandum rather than Scarantino’s. See that page for what this does and does not settle — reply (2), that common sense should not legislate scientific definitions, is untouched, and it was always the strongest of the three.
3. It sharpens the word-cue objection to Bodily sensation maps (bodily fingerprints of emotion) — and cuts both ways doing it. This is the finding with the most leverage and the one that needs the most care.
The word-cue objection, sharpened in both directions
emBODY’s central interpretive risk is that showing someone an emotion word and asking where they feel it may tap shared concepts rather than shared bodily experience (Are emotions culturally universal?). Ferré et al. establish, without intending to, that the Spanish concept of emotion is constituted in part by interoceptive content — being about internal body sensations is most of what makes a word an emotion word at all.
That fact is genuinely bidirectional, and the wiki should not pretend otherwise:
- The constructionist reading. If bodily content is baked into the concept, then a subject cued with an emotion word already has an interoceptive expectation supplied by the lexicon before any introspection occurs. Cross-cultural concordance of word-cued maps is then doubly overdetermined by shared conceptual knowledge, and the objection on emBODY (topographical bodily sensation mapping) gets stronger rather than weaker.
- The Nummenmaa reading. Concepts have to come from somewhere. The most natural explanation for why the folk concept of emotion is interoceptive is that emotions are felt in the body, and the lexicon has recorded that fact. On this reading Ferré et al. is a semantic fossil of exactly the bodily experience Volynets et al. measure directly. Ferré et al. themselves lean here, citing Nummenmaa’s programme approvingly.
Nothing in this paper adjudicates between them — it is a rating study and cannot. But it does move the burden: the word-cue objection can no longer be waved off as an abstract possibility, because the semantic content it postulates has now been measured, and it is large. Recorded on emBODY (topographical bodily sensation mapping) and Are emotions culturally universal?.
Note the one thing that does help Nummenmaa here, from elsewhere in the wiki: Lyons et al. (2021) found BSMs fracture by diagnosis, and shared lexical knowledge should not do that. The Spanish lexicon does not know who is medicated.
Against core affect, mildly
A smaller result with a target. Core affect holds that feeling is valence plus arousal. If that were right, the feeling ratings here should be largely recoverable from the valence and arousal norms — and they are not: feeling’s partial correlation with valence was nonsignificant (−.043) and with arousal only −.11. The authors read this as evidence that feeling is not reducible to the two dimensions.
Two cautions the paper does not raise. These are partial correlations controlling for every other variable including evaluation, which correlates .712 with valence — a live suppression risk. And a null partial correlation between feeling and valence is not obviously what a constructionist would predict anyway: core affect is valence and arousal jointly, and “feeling” ratings collapsed across the whole lexicon would be expected to be orthogonal to polarity, since sadness and joy are both intensely felt. The result is suggestive against the reduction, not decisive, and the EFA arguably says more: prototypicality loads .760 on emotional experience and not at all on socioemotional polarity, so whatever makes a word an emotion is nearly independent of whether it is a good or bad one.
Two defects in the source, flagged not propagated
The morality coefficient is misread by its own discussion. Table 4 reports morality at beta = −.137 — controlling for everything else, more moral content makes a word less prototypically emotional. The discussion states the reverse: “it is more probable that an abstract word refers to an emotion when it has to do with mental activities, judgments, and moral rules.” That sentence is true of thought (beta = .123) and false of morality, which it bundles in. The wiki records morality as a negative predictor, per the table.
Table 4’s thought row is internally inconsistent: beta = .123 with t = −4.915. Since the partial correlation of thought with prototypicality is +.140 and the discussion treats thought as a positive contributor, the negative t is the misprint. Minor, but this is the second APA-journal table in this folder with a sign/transposition error — compare the verified transposition in Lyons et al. (2021).
The reliability problem nobody flags
The most serious internal weakness is not in the tables but in the validity section, and the authors pass over it without comment.
Every variable’s reliability was checked by correlating control-word means across questionnaire versions. Eight of nine variables landed between r(22) = .824 and .955. Interoception came in at .654 — the lowest by a wide margin, and the only one below .8. The men/women agreement correlation tells the same story (interoception .463; evaluation .755). Its ICC is respectable (.80), so this is not a fatal objection, but the pattern is consistent: raters agree least about interoception, and interoception is the variable the paper’s title, abstract and headline R² all rest on.
There is a charitable reading that is also more interesting than a complaint. Low inter-rater agreement about which words are “related to internal body sensations” is what you would expect if interoceptive experience genuinely varies between people — which is the premise of the entire individual-differences literature on Interoceptive sensitivity and the Interoceptive taxonomy (Farb et al. Box 1). The noise may be the signal. But the paper does not make that argument, and a variable that raters agree on least explaining variance best is a combination that deserves stating out loud.