Nummenmaa, Glerean, Hari & Hietanen (2014) — Bodily maps of emotions
The founding paper of the bodily maps programme, and the source of emBODY. Five experiments, 773 subjects, one instrument: colour the body regions whose activity you feel increasing (left silhouette) and decreasing (right silhouette) while a stimulus is presented. Mass-univariate pixelwise t-tests against zero over ~50,000 body pixels, FDR-corrected, then LDA classification and hierarchical clustering.
This wiki has been citing it second-hand since the Volynets ingest. It should be read first-hand, because the claim the wiki flagged as load-bearing and unverifiable turns out to be the paper’s actual centre of gravity.
The claim the wiki has been chasing
Four pages here (embody, cultural-universality-of-emotion, lauri-nummenmaa, bodily-sensation-maps) record the same complaint: the programme’s answer to the word-cue objection rests on “Nummenmaa et al. (2014) reported that word-cued maps match those from nonverbal emotion induction,” a claim “carried by reference” and “cited forward rather than retested,” and it “deserves direct replication.”
Two corrections follow from reading the source.
It is not one finding. It is three experiments, designed against exactly this objection and described as such (“When people recall bodily sensations associated with emotion categories described by words, they could just report stereotypes”). Experiment 2 used vignettes; experiment 3 used silent 10-second films; experiment 4 used facial photographs. The concordances are rs = 0.71-0.82, and the LDA accuracies from induction exceed the word accuracies.
The design is better than the objection assumes. The vignettes were pilot-normed on 72 raters and deliberately never named the emotion, the feeling, or any bodily sensation — “You sit by the kitchen table. The dishwasher is turned on” (neutral); “While visiting the hospital, you see a dying child who can barely keep her eyes open” (sad). The films were shown without sound. The silhouettes carry no internal organs, specifically “to avoid triggering purely conceptual associations between emotions and specific body parts (e.g., love–heart).” These are the choices of people who saw the objection coming.
So the wiki’s framing needs downgrading in force and sharpening in target. The complaint was never that the 2014 evidence is thin — the wiki couldn’t see it. The complaint that survives is narrower and still correct: later papers inherit the conclusion without inheriting its scope. What experiments 2–4 license is that Finnish undergraduates produce concordant maps under verbal and nonverbal induction. Volynets et al. cite it to license word-cued maps standing in for felt experience across 101 countries. The inference travels further than the datum.
What the induction experiments do and do not retire
| version of the word-cue objection | survives experiments 2–4? |
|---|---|
| The maps are an artifact of reading an emotion word | no — dead |
| The maps are Finnish-language figurative usage (“heartache”) | no — dead, via the Swedish control |
| The maps are West European cultural knowledge | weakened — the Taiwanese sample is n = 36 |
| Subjects report where their concept says the feeling belongs, however cued | yes — untouched |
The last row is the one that matters, and the paper knows it. A vignette about a dying child is not a word, but it is still a token of a shared emotion category, and a subject who recognises it as a sad scenario can paint from the same conceptual knowledge a word would have cued. The authors’ response is not evidential; it is an argument from origins: “where then do these conceptual associations originate and why are they so similar across people with very different cultural and linguistic backgrounds? A plausible answer would again point in the direction of a biological basis.”
That is a fair argument and it is not a result. It is also, precisely, the argument Ferré et al. (2024) makes tractable a decade later from the other side — and the direction of fit remains unsettled. See bodily-sensation-maps.
Experiment 5, which the programme rarely mentions
87 (or 72 — the paper says both) naive observers were shown the averaged basic-emotion maps and asked to name them. Mean accuracy 46% against 14% chance. The per-emotion breakdown is the interesting part:
| recognition | |
|---|---|
| neutral | 99% |
| anger | 58% |
| surprise | 54% |
| disgust | 43% |
| sadness | 38% |
| happiness | 22% |
| fear | 8% — below the 14% chance level, ns |
Fear’s fingerprint is not legible to humans. An LDA on 30 principal components tells fear apart; a person looking at the map cannot. This is worth holding onto, because the programme’s strong claim (claim 2 on lauri-nummenmaa) is that these maps determine the qualia of felt emotion. If the map of fear constituted what fear feels like, the people who have felt fear should recognise it. The obvious defence — recognising a heatmap is a different task from having an experience — is available and is also a concession: it puts distance between the map and the feeling exactly where the constitutive claim needs them fused.
Note also that neutral at 99% is doing quiet work in the 46% average. Neutral is the one map that is mostly blank.
Where the discreteness actually is
The paper’s own topographic summary is more modest than its reputation:
- Shared, not specific: head activation appears in every emotion; upper-chest activation in most.
- Specific: digestive/throat sensations in disgust; whole-body enhancement in happiness (uniquely); limb deactivation in sadness; upper-limb activation in the approach pair (anger, happiness).
- Weak: the seven nonbasic emotions show “a much smaller degree of bodily sensations and spatial independence,” with fear/anxiety and sadness/depression nearly overlapping.
The mean mismatched-emotion correlation of rs = 0.52 across experiments belongs in the same paragraph. These maps share most of their structure; the emotion-specific part is a real but comparatively small residual on top of a large common component. “Discrete fingerprints” is a description of what a classifier can extract, not of what the maps look like.
That is not a criticism of the result — it is the reason Volynets et al.’s inter-nation rs > 0.82 needs its floor stated: maps of different emotions already correlate at 0.52 by default.
The autonomic reconciliation — a claim worth naming
Buried in the discussion is a specific proposal about autonomic-specificity-of-emotion, and it is Nummenmaa’s clearest statement in this wiki on that debate:
individuals are poor at detecting specific physiological states… physiological data have not revealed consistent emotion-specific patterns of bodily activation, with some recent reviews pointing to high unspecificity (Barrett 2006) and others to high specificity (Kreibig 2010). Our data reconcile these opposing views by revealing that even though changes in specific physiological systems would be difficult to access consciously, net sensations arising from multiple physiological systems during different emotions are topographically distinct.
The structure of the claim: individual autonomic channels are neither reliably emotion-specific nor consciously accessible, but the compound of skeletomuscular, visceral and autonomic sensation — “which the individuals cannot separate” — is emotion-specific at the level of felt topography.
This is more interesting than the wiki’s current “adjacent, not aligned” filing of Nummenmaa on that debate, and it should be filed as a position rather than a near miss. It is also not supported by these data. Nothing here measures the compound; it measures reports of the compound. The reconciliation is a hypothesis about why felt maps could be specific while channels are not — the same hypothesis a constructionist would call “the concepts are specific while the channels are not,” with a different noun. See autonomic-specificity-of-emotion, autonomic-specificity.
Against the Lyons replication
Reading the original clarifies two things the Lyons et al. (2021) page had to leave hanging.
The deactivation discrepancy is sharper than recorded. Deactivation is not incidental here. Sadness and depression are the two maps carried primarily by it (Fig. 2, near-solid blue), and “sensations of decreased limb activity were a defining feature of sadness” is the paper’s own phrasing. Lyons et al. found “almost no deactivation reached significance” in healthy controls under picture induction. That is not a minor between-study wobble; it is the disappearance of the defining feature of one of the six maps under a change of cue. Recorded on embody and bodily-sensation-maps.
Anger is the programme’s recurring failure, and it is now three-for-three. Nummenmaa et al. dropped anger and surprise from the movie experiment outright, citing “the inherent difficulties associated with eliciting anger and surprise with movie stimuli.” Lyons et al. got below-chance classification for anger under picture induction and dropped it from their reanalysis. The one place anger works is word cues — where it is also the best-recognised map in experiment 5 (58%). An emotion that classifies well when named and fails whenever it must be induced is the single cleanest test case the paradigm has for the word-cue objection, and no source in this wiki has treated it as one.
Assessment
The strongest paper in the wiki’s bodily-maps folder, and the one whose reputation is least accurate. It is usually cited for cultural universality (which rests on 36 Taiwanese subjects here) and for discrete fingerprints (which are a residual on a large common component). What it actually establishes, and establishes well, is that felt bodily topography is reproducible across radically different elicitation methods within a population — words, vignettes, silent film, and photographs of strangers’ faces converge at rs = 0.71–0.82.
That is a real and non-obvious result, and it is the load-bearing one for everything downstream. It also does not answer the question the downstream papers use it to answer.