Human Emotion Ontology (HEO) · Schema version 1.0.0
This report summarizes the major theoretical families that the ontology draws on, identifies where they genuinely agree and disagree, and explains the synthesis adopted in the data files (data/emotions.json, data/relationships.json).
There is no single agreed taxonomy of human emotion. Different research traditions answer the most basic questions differently:
eleven (Panksepp's systems), fifteen or more (Ekman's expanded list), twenty (Geneva Emotion Wheel families), twenty-two (OCC), twenty-seven (Cowen & Keltner's empirically derived categories), or two (valence and arousal).
(basic emotion theory), a point in a two- or three-dimensional space (dimensional models), the output of an appraisal of goal-relevance (appraisal theory), or a culturally learned categorization of bodily feeling (constructionism).
universality for a small set of expressive and experiential patterns (Ekman & Friesen 1971; Cordaro et al. 2018, 2019), and strong cultural variation in the boundaries of most emotion categories and in display rules (Gendron et al. 2014; Jack et al. 2012; Wierzbicka 1999).
The ontology therefore treats no single framework as canonical. Instead it models emotion science as it actually is: a set of partially overlapping, partially contradictory models, and preserves disagreements explicitly (see 04-disagreement-registry.md) rather than manufacturing a false consensus.
Core claim. A small set of emotions are biologically basic: evolutionarily old, hardwired, expressed in recognizable ways across cultures, and the "atoms" out of which other emotions combine.
anger, fear, sadness, enjoyment, disgust, and surprise are recognized in posed faces across literate cultures. Ekman (1999) later expanded the list with amusement, contentment, embarrassment, guilt, pride in achievement, relief, satisfaction, sensory pleasure, shame, and contempt.
disgust, anger, anticipation) arranged in four opposite pairs on a wheel, with intensity levels and dyads: combinations of two primaries that yield complex emotions (e.g., joy + trust = love; fear + surprise = awe).
the mammalian brain — SEEKING, FEAR, RAGE, LUST, CARE, PANIC/GRIEF, PLAY — the strongest evidence that something emotion-like is built into the nervous system.
Strengths. Evolutionary logic (Nesse 1990; Tooby & Cosmides 2008); cross-species continuity (Panksepp); a defensible account of the smallest set of emotions with expressive and physiological markers (Keltner et al. 2019).
Weaknesses. The exact membership of the "basic" set varies by author; facial expression data are weaker for cultures with little Western media contact (Gendron et al. 2014); meta-analytic reviews find limited autonomic specificity (Lindquist et al. 2012; but see Kreibig 2010); and most emotion words in any language do not refer to basic emotions at all (Ortony & Turner 1990).
Core claim. Emotions are not categories but points in a continuous space defined by a few dimensions.
(activation–deactivation), with emotion words arranged in a circle.
dimension, dominance/submissiveness (PAD: Pleasure–Arousal–Dominance).
neurophysiological state of feeling good/bad, activated/deactivated, which underlies all emotional episodes. Emotional episodes are core affect made meaningful by categorization in context.
Positive Affect and Negative Affect (the basis of the evaluative space idea).
Strengths. Parsimony; strong psychometric support; predicts similarity relations among emotion words; handles moods and subtle feelings that discrete categories cannot; PAD adds control, which matters for anger vs. fear (Smith & Ellsworth 1985).
Weaknesses. Explains similarity but not distinctness (why fear and anger feel different despite similar valence/arousal); flattens content (what the emotion is about); dominance/certainty dimensions are appraisal-like, blurring the dimensional/discrete boundary.
Core claim. Emotions arise from an evaluation of events relative to one's goals and well-being. Different appraisal patterns produce different emotions.
challenge determine the emotion and its coping implications.
anticipated effort, certainty, attentional activity, self-other responsibility/ control, and situational control — that distinguish emotion families empirically (e.g., fear = low control + high uncertainty; anger = high control + other-blame).
(novelty, pleasantness, goal/need relevance, coping potential, norm compatibility) synchronizes five components: cognitive appraisal, physiological arousal, motor expression, motivation, and subjective feeling.
types as valenced reactions to (a) consequences of events, (b) actions of agents, and (c) aspects of objects — the framework most used in AI/affective computing.
Strengths. Explains content and object-directedness; handles the fine distinctions among similar emotions (regret vs. disappointment, shame vs. guilt); computationally tractable; strongly supported by experimental work (Smith & Ellsworth 1985; Scherer 2009).
Weaknesses. Appraisals and emotions are hard to separate causally (appraisal can follow the emotion); individual and cultural variation in appraisal; risk of infinite regress (an emotion about an emotion about…).
Core claim. Emotions are not natural kinds with dedicated brain circuits; they are conceptual categories that the brain constructs from two more basic ingredients — core affect (valence/arousal) and conceptual knowledge — applied to interoceptive and sensory input (Barrett 2006, 2017; Russell 2003; Lindquist et al. 2012).
Strengths. Explains why emotion categories vary across languages and individuals; accounts for the weak one-to-one mapping between emotion and physiology/expression; motivates emotion differentiation and granularity research (people with richer emotion concepts regulate better; Barrett et al. 2001; Kashdan et al. 2015).
Weaknesses. Can undersell the strong cross-cultural and cross-species evidence for a small set of evolved affective responses (Panksepp 2011); the theory's falsifiability and treatment of facial-expression evidence are actively debated (see 04-disagreement-registry.md).
distinct categories of emotional experience bridged by continuous gradients** — e.g., anxiety→fear→horror→disgust, calmness→aesthetic appreciation→awe. This is the strongest current evidence that emotional experience is both categorically structured and continuously graded: the discrete-vs-dimensional dichotomy is partly false.
emotion words are culture-specific lexicalizations; "sadness," "anger," and "disgust" are not language-neutral labels for universal states.
languages, showing the lexicon expands far beyond English terms.
self-representation and attention to others' evaluations (Lewis 1993; Tangney & Dearing 2002; Tracy & Robins 2004, 2007). Guilt motivates repair; shame motivates hiding — a behavioral distinction with strong empirical support.
respond to the welfare of others and moral violations (Haidt 2003; Rozin et al. 1999).
theory (2001) and work on awe (Keltner & Haidt 2003), gratitude (McCullough et al. 2001), compassion (Goetz et al. 2010), and kama muta / being moved (Fiske et al. 2017) established them as a distinct field with distinct functions.
The ontology adopts a multi-level, multi-representation model that treats these theories as describing different levels of the same phenomenon rather than rivals:
valence, arousal, and dominance (plus approach/avoidance direction). This is the best-supported lowest level (Russell & Barrett 1999; Mehrabian & Russell 1974) and is modeled by the dimensions field on every node.
into discrete, nameable states using evolved predispositions (basic emotions, Panksepp's systems) and learned, language-shaped concepts (constructionism, Wierzbicka). This level is modeled by the node set itself.
emotion is a response to (loss, threat, injustice, goal blockage) — is modeled by appraisal.* nodes and appraised_as edges.
schadenfreude) are modeled as nodes whose parents and relationships encode their interpersonal structure.
and co-occur. This is modeled by transitions_to, amplifies, inhibits, co_occurs_with, often_follows/often_precedes edges.
What the field agrees on (high-confidence anchors of the ontology):
dimension (Russell 1980; Watson et al. 1988).
plus (with more debate) interest/anticipation, contempt, shame, embarrassment — have special status as expressive, cross-culturally recognizable, and neurobiologically grounded (Ekman 1999; Izard 1977; Plutchik 1980; Panksepp 1998; Cordaro et al. 2018).
constructs with different appraisals, behaviors, and consequences (Tangney & Dearing 2002; Parrott & Smith 1993; Öhman 2008).
(Larsen et al. 2001; Berrios et al. 2015; Cacioppo & Berntson 1994).
approach/aggression, sadness → withdrawal, disgust → avoidance, guilt → reparation, joy → affiliation/approach (Frijda 1986; Nesse 1990).
evaluation, even if a small core is broadly shared (Wierzbicka 1999; Lutz 1988; Gendron et al. 2014; Cordaro et al. 2019).
What the field disagrees on (preserved in the ontology): see 04-disagreement-registry.md. The most consequential disagreements — basic vs. constructed, how many categories, whether blends are real combinations or linguistic conveniences — are encoded as (a) multiple parents/groupers for contested nodes, (b) confidence and evidence_status fields, and (c) notes recording the dispute on the node itself.
The ontology keeps the following constructs as separate categories rather than collapsing them into "emotions":
| Construct | Modeled as | Distinguishing feature |
|---|---|---|
| Emotion | category: emotion | Episodic, object-directed, multi-component (appraisal + physiology + expression + feeling) |
| Feeling | category: feeling / abstract.feeling | The subjective, conscious component of affect; may occur without a full emotion episode |
| Mood | category: mood | Diffuse (no object), lower intensity, longer duration (hours–days) |
| Affect | affect, abstract.core_affect | The broadest term; the continuous valence/arousal substrate |
| Drive | category: drive | Homeostatic pressures (hunger, thirst, sleep, lust, fatigue) |
| Motivation / Desire | category: motivation, category: desire | Goal-directed wanting and behavioral tendencies |
| Need | category: need | Requirements for well-being whose lack motivates |
| Appraisal | category: appraisal | Cognitive evaluation; input to, not the same as, emotion |
| Physiological state | category: physiological_state | Somatic components (sympathetic arousal, crying, goosebumps) |
| Behavioral tendency | category: behavioral_tendency | Action readiness (escape, attack, withdrawal, reparation) |
| Trait | category: trait | Enduring disposition (neuroticism, trait anxiety, BIS/BAS) |
| Sentiment | category: sentiment | Enduring object-directed attitude (hatred, loyalty) |
| Mental state | category: mental_state | Attentional/cognitive states tied to affect (rumination, absorption) |
These are linked, not siloed: e.g., emotion.fear is linked to appraisal.threat (appraised_as), physio.fight_flight_response (physiological_association), behavior.escape (motivates, behavioral_expression_of), drive.fatigue → emotion.irritation (triggered_by), and trait.trait_anxiety as a dispositional correlate.
The ontology does not assume emotions are mutually exclusive or binary. A state at a moment is represented as a distribution over nodes — e.g., {Fear: 0.7, Curiosity: 0.6, Excitement: 0.4, Sadness: 0.2} — where the weights are best understood as degrees of activation/prototypicality, not literal neural measurements (see 02-ontology-design.md §5 and 05-confidence-model.md). This is justified by:
(Larsen et al. 2001); the evaluative space model makes positivity and negativity separable (Cacioppo & Berntson 1994); bittersweet experiences (nostalgia, meaningful endings) are commonplace and experimentally reproducible (Sedikides et al. 2015; Larsen et al. 2021).
joy+sadness, schadenfreude as pleasure+resentment) are encoded as blend_of edges. Where the "blend" interpretation is contested, the edge confidence is low and the dispute is noted.
gradients (anxiety→fear→horror), supporting fuzzy membership rather than all-or-nothing states.
states that can persist beneath foreground emotions; the graph links them with similar_to and co_occurs_with edges.
parents, synonyms, evidence, confidence). data/relationships.json — one object per edge (source, target, type, directionality, strength, confidence, evidence).
edge types. Query the graph by edge type, confidence, and strength.
integrity checker.
graph; the Valence × Arousal view shows the dimensional structure.