RESEARCH DOSSIER
Affective Dimensions and Emotion Taxonomy
Abstract. The organization of affective experience poses a central problem of psychological classification: can a few general dimensions describe emotional life, or must emotion taxonomy preserve many distinguishable kinds of experience? This critical narrative research monograph traces the history of pleasure–arousal models, positive and negative affect, the circumplex and core-affect accounts, basic-emotion categories, multidimensional semantic spaces, and emotional granularity. It critically evaluates foundational experiments, PANAS and SAM measurement, stimulus selection, factor and embedding methods, cross-cultural evidence, autonomic and neural data, and studies comparing videos, music, narratives, and everyday life. Its synthesis identifies reproducible broad valence and activation dimensions alongside finer, context- and method-dependent categorical distinctions; emphasizes that report, attribution, expression, and biological signals are separate observables; and proposes rigorous cross-domain tests of taxonomic generalizability. The work is a critical narrative research synthesis of published literature, not an original experiment, systematic review, or meta-analysis.
Keywords: affective dimensions; emotion taxonomy; valence; arousal; dominance; core affect; circumplex model; positive affect; negative affect; basic emotions; emotion categories; emotion semantic space; dimensionality reduction; emotion granularity; emotion differentiation; cross-cultural emotions; affective measurement; PANAS; Self-Assessment Manikin; human–artificial affective inference
Introduction: The Research Object of Affective Classification
Affective dimensions and emotion taxonomy concern the formal representation of how emotional states resemble, differ from, and relate to one another. A dimension locates observations along an interpretable continuum, such as pleasantness or activation; a taxonomy partitions or organizes phenomena into named kinds, families, or hierarchical relationships. An empirical emotion space is constructed from reports, comparisons, behavior, expressions, or physiological measures, with the selected observations determining what that space can establish. The central research question is which structure best represents the diversity of affective experience and under what conditions that representation generalizes. Affective space is therefore an empirical model with a specified measurement domain, rather than a ready-made list of emotions. Russell (1980) Lin et al. (2025)
The present monograph investigates valence, arousal, dominance, positive and negative activation, core affect, circumplex organization, category systems, semantic spaces, differences in emotional differentiation, and methods for estimating dimensionality. It addresses the relation between a person's reported feelings and the language used to describe them, while using converging expressive, cultural, and neural evidence to test the wider interpretation of taxonomies. Event appraisal, emotion regulation, neural generation, and affective development appear only where they establish the limits or validation of classification; their mechanisms belong to separate focused research objects. Barrett & Russell (1999) Fontaine et al. (2007)
The evidence is evaluated as a critical narrative research synthesis. Foundational experiments and scale-development studies establish what the original measures actually captured. Large modern stimulus samples and cross-cultural investigations test whether a few familiar dimensions preserve relevant structure. Methodological work is used to distinguish dimensions of ratings from dimensions of nature. There is no preregistered systematic search, pooled effect-size estimate, newly collected dataset, or experimental intervention in this monograph. Accordingly, conclusions concern the extent and limits of the evidence covered, rather than an exhaustive census of all published studies. Cowen & Keltner (2017) Han et al. (2024) Lin et al. (2025)
Three different tasks frequently travel under the name emotion classification: describing a current internal state, categorizing an external stimulus or another person's expression, and predicting what a person will do. Each may involve related affective concepts but produces a different observation matrix. An emotion word that separates self-reports in one task may become poorly distinguishable when mapped from facial motion, autonomic activity, or short speech clips. Scientific comparison must hold the target, stimuli, task, sampling, model, and success criterion explicit before concluding that two studies agree or contradict one another. Cowen & Keltner (2021) Siegel et al. (2018) Keltner et al. (2019)
From Emotion Words to Scientific Measurement Spaces
The history of affective classification includes philosophical inventories of passions, ordinary-language distinctions, Darwinian theories of expression, psychometric studies of adjectives, and multivariate measurement. The decisive modern methodological change was treating relations among observations as data. Instead of choosing one authoritative list of emotions, researchers asked participants to rate feelings, compare words, or describe responses to stimuli, then examined the resulting covariation and distances. These operations made models of affect falsifiable at the level of expected item relations and allowed competing structures to be tested on new samples. Russell (1980) Watson et al. (1988)
Mehrabian and Russell's environmental research in 1974 introduced pleasure, arousal, and dominance as responses through which varied settings could be characterized. Their semantic-differential approach linked evaluative judgments about situations to affective responses and helped establish the PAD vocabulary. Russell's 1980 article subsequently made a specific circular prediction for affective adjectives: pleasure, excitement, arousal, distress, displeasure, depression, sleepiness, and relaxation should occupy ordered positions around a plane. The historical importance lies in the relation among terms and the testable circular arrangement, beyond the particular words selected. Mehrabian & Russell (1974) Russell (1980)
An influential alternative lineage defined positive and negative affect as broad factors. Watson, Clark, and Tellegen developed the Positive and Negative Affect Schedule (PANAS) in 1988, providing brief measures whose scores were meant to track relatively independent dimensions of activated positive and negative experience. This factor-analytic structure does not directly contradict a two-axis pleasantness-by-activation circumplex: a rotation of axes in an idealized plane can yield different factor pairs, while questionnaire item selection and time instructions change empirical relations. Interpreting a scale therefore requires studying its content and response frame, not only the name attached to its factor. Watson et al. (1988) Barrett & Russell (1999)
Twentieth-century basic-emotion theories proposed recognizable classes linked to characteristic expressive and functional patterns, with Ekman's 1992 argument a major explicit formulation. Such categories address differentiation between kinds of emotional episodes, whereas adjective-space experiments often address variation in consciously reportable feeling. Later empirical programs drew on many more labels, video clips, sounds, cultures, and machine-learning methods. The resulting tension between a few broad axes and many distinct category patterns reflects both genuine explanatory disagreement and the expanding number of questions that can now be measured. Ekman (1992) Cowen & Keltner (2017) Lin et al. (2025)
Valence: Pleasantness, Evaluation, and Affective Significance
Valence is commonly operationalized as the position of a feeling or stimulus evaluation along an unpleasant-to-pleasant continuum. A report of pleasure is psychologically different from a researcher coding a stimulus as positive, a participant choosing to approach something, or an observer inferring that a facial expression is favorable. These indicators may correlate but measure distinct processes: experienced hedonic tone, evaluative meaning, behavioral preference, and social interpretation. Valence must therefore be defined as an observation-level construct before correlations among its measures are interpreted as one common latent property. Russell (1980) Bradley & Lang (1994)
A bipolar pleasure axis assigns opposite signs to pleasant and unpleasant feelings within the same coordinate. PANAS positive and negative affect scores use partly different activated adjective sets, so a person may score high on both within a chosen time window. The difference is one of instrument construction and timescale: two scores recording frequencies or intensities of selected emotions do not automatically form the endpoints of a single momentary judgment. Researchers seeking daily mood profiles should test whether their data support a bipolar continuum, two correlated factors, or a richer multidimensional solution. Watson et al. (1988) Barrett & Russell (1999)
Valence is often conflated with motivation. Yet anger may be experienced as unpleasant while organizing approach, confrontation, or efforts to remove obstruction. Interest may have relatively modest hedonic intensity while directing sustained engagement. The orientation toward or away from an object, the felt quality of that state, and the expected consequences of acting are separate variables. Valence remains useful because it summarizes a broad portion of experience, but taxonomy requires additional dimensions or categories when those distinctions determine psychological function. Fontaine et al. (2007) Keltner et al. (2019)
Neural reward research reinforces the value of operational distinctions. Berridge and Kringelbach describe pleasure-related systems with separable processes involved in hedonic impact, incentive motivation, and learning. A ratings-based pleasantness coordinate cannot by itself identify the causal neural circuitry or behavioral incentive value of an object. Stronger measurement designs combine self-report with choice behavior and carefully specified physiological or neural indicators, asking whether all sources support a common dimension or reveal meaningful divergence. Berridge & Kringelbach (2015) Lindquist et al. (2012)
Arousal: Activation, Intensity, and Temporal State
Arousal in affective measurement often denotes subjective activation or energy, ranging from drowsy and quiet to alert and highly activated. Researchers also use arousal for autonomic mobilization, vigilance, stimulus intensity, or physiological activation. These meanings require distinct instruments. A person's report of feeling energized can change independently of heart rate, skin conductance, or motor readiness. A dimensional emotion model is clearest when its arousal coordinate refers explicitly to experienced activation and additional signals are treated as potential correlates to be tested. Russell & Barrett (1999) Bradley & Lang (1994)
The classic circumplex pairs arousal with pleasure as approximately orthogonal coordinates, allowing equally pleasant states to differ in energy: relief may feel calm, whereas excitement is active. Unpleasant states can likewise differ between fatigue-like sadness and highly activated alarm. This arrangement avoids reducing emotion classification to positivity versus negativity. It also helps researchers specify response instructions: a person can be asked about current activation, the average of a day, a remembered episode, or intensity elicited by a particular image, and these provide meaningfully different datasets. Russell (1980) Yik et al. (2011)
Kuppens and colleagues examined the relation between valence and arousal in subjective experience and highlighted that their statistical association is a substantive empirical question rather than a structural identity. Under different circumstances, very pleasant or very unpleasant states may differ in how intensely they are experienced; sampling routines and individual differences can affect the observed shape of the joint distribution. Thus the familiar square displaying four valence-arousal quadrants is a useful coordinate chart, while empirical states may cluster unevenly within it. Kuppens et al. (2013)
Arousal-based explanations face particular challenges when indicators carry different timescales. A skin-conductance response, a retrospective emotion word, and momentary self-reported activation do not necessarily refer to the same event boundary or recovery trajectory. Moreover, physiological activation can be caused by temperature, exertion, medication, posture, or threat appraisal. Affective taxonomy must avoid classifying bodily activation as a unique emotion without evidence connecting the measured response to the specific episode and context. Siegel et al. (2018) Barrett (2017)
Core Affect and the Geometry of the Circumplex
Core affect refers in Russell and Barrett's formulation to an elementary consciously accessible feeling that varies in pleasure and activation. An emotional episode, by comparison, is a more articulated event that can include an object, conceptual interpretation, attention, expression, and action. The distinction provides a scientific division of labor: a core-affect measure captures broadly felt condition, while a named emotion category adds representational and situational information. The coexistence of these levels explains why a two-dimensional model can successfully summarize a large amount of variability without uniquely identifying fear, gratitude, embarrassment, or jealousy. Russell & Barrett (1999) Barrett & Russell (1999)
The circumplex is a model of relations among variables projected into a plane. Its organizing claim is circular: items with neighboring angles should be more similar or correlated than those positioned at widely separated angles, with opposed poles showing negative relations. Russell's 1980 design used multiple approaches, including ordering of affective adjectives, similarity scaling, and participant reports. Importantly, a diagram with eight emotions evenly spaced around a ring is an illustrative coordinate system, not proof that the sampled emotional events populate the plane uniformly or that each word denotes one natural kind. Russell (1980)
Yik, Russell, and Steiger developed a 12-point circumplex structure of core affect in four correlational studies and evaluated different response formats. Their work shows that broader theoretical dimensions can support more fine-grained operational scales: the question is how many informative distinctions can be made around a plane, not whether the 12 points imply 12 biological emotion types. The ordering and angular relation of questionnaire items should be evaluated in the population and measurement format used, especially when an instrument is transported to a different language. Yik et al. (2011)
Two-dimensional reduction is most useful when scientific aims concern hedonic quality, activation, broad affective transitions, or parsimonious visualization. It is less sufficient when the outcome requires distinguishing states with similar pleasantness and arousal but differing social meanings or action tendencies. A taxonomy can therefore use the circumplex as a common background coordinate while organizing additional categorical distinctions. Whether the added complexity is justified must be decided by independent predictive performance, out-of-sample reliability, and the question under study. Fontaine et al. (2007) Cowen & Keltner (2021)
Dominance, Agency, and Additional Affective Axes
Dominance is often defined in PAD approaches as a sense of control, influence, or power within the affective response to a situation. It clarifies why two episodes with similar valence and activation may feel qualitatively different: alarm in an uncontrollable situation and energized confrontation in a situation experienced as manageable differ in perceived agency. Mehrabian and Russell proposed pleasure, arousal, and dominance within an environmental-affect framework, while Bradley and Lang compared a pictorial measure covering the same three dimensions with semantic differential ratings. Dominance is therefore a legitimate operational coordinate with a traceable measurement tradition. Mehrabian & Russell (1974) Bradley & Lang (1994)
Additional candidate dimensions may describe novelty, certainty, approach, social engagement, self-relevance, moral evaluation, or bodily activation. These axes may operate at different analytical levels. Certainty is sometimes an appraisal of a situation; approach may be an action tendency; dominance may be felt personal control; sociality may be the relational meaning of an experience. Their correlation with affective reports does not make them interchangeable fundamental components. A model gains scientific value by stating what each coordinate represents and what theoretically distinct observation it can explain. Fontaine et al. (2007) Lin et al. (2025)
Fontaine, Scherer, Roesch, and Ellsworth used a semantic approach to argue that the world of emotion terms cannot be fully represented in only two dimensions. Their work demonstrates the analytical consequences of including features beyond pleasure and energy, including power-related and unpredictability-related aspects. A four-dimensional result from semantic judgments should be interpreted as structure within that particular sampled meaning space; it provides a reason to compare more dimensions, not an automatic estimate of the number of independent neural mechanisms. Fontaine et al. (2007)
The choice of dimensionality is therefore inseparable from the purpose of classification. Affective state monitoring may require two reliable coordinates, interpersonal emotion interpretation may require perceived agency and social reference, and clinical assessment may require the distinctions that best forecast impairment or change. If several axes remain highly redundant after correcting for measurement error, their number may be compressed. If additional axes reproducibly predict independent outcomes, a richer representation is warranted. The research problem is model selection under explicit empirical and conceptual constraints. Lin et al. (2025) Han et al. (2024)
Basic-Emotion Families and Categorical Hypotheses
Categorical emotion theories propose that recognizable kinds of emotional episodes are scientifically useful units of analysis. Classical basic-emotion accounts concentrate on phenomena such as fear, anger, sadness, disgust, joy, and surprise, often emphasizing evolutionary functions and expressive regularities. In Ekman's 1992 formulation, basic emotions are proposed to differ by coordinated processes, antecedents, and characteristic responses. A category has scientific content when it predicts observable structures that survive comparison across people, tasks, and settings. The mere existence of an English label or a high agreement rate under forced-choice recognition provides a narrower result. Ekman (1992)
Categorical hypotheses can be arranged at different levels. Fear and disgust might be proposed as broad families, while alarm, horror, revulsion, and contamination concern can describe neighboring states with partly shared properties. Conversely, pride, awe, gratitude, amusement, and desire challenge a scheme restricted to a few negative emotions and one undifferentiated happiness category. The empirical question is whether category distinctions support reliable discrimination and prediction beyond their shared dimensional positions. Different modalities may preserve different distinctions, and no fixed count follows from the category concept alone. Tracy & Robins (2004) Keltner et al. (2019) Cowen & Keltner (2017)
Evidence from recognition experiments must be understood in its design. Participants may choose among supplied words, freely describe an expression, arrange stimuli by similarity, or predict the likely situation generating a behavior. Each method controls the semantic alternatives available to the respondent. Tracy and Robins' experiments on pride are important partly because recognition was tested with an open-ended response condition, complementing forced-choice tasks. Yet recognizing a posed expression remains different from demonstrating that all naturally occurring pride episodes possess the same signal. Tracy & Robins (2004) Gendron et al. (2014)
Modern accounts of basic emotion have expanded beyond a rigid set of six facial expressions. Keltner and colleagues discuss more numerous expressions, graded patterns, and contextual sources of meaning. This literature supports treating basic-emotion categories as hypotheses about coordinated functional and communicative regularities rather than exhaustive labels for all experienced affect. Rigorous comparisons therefore include novel populations, naturalistic settings, multiple modalities, and counterexamples in which familiar expressions occur without the corresponding inferred feeling. Keltner et al. (2019) Cordaro et al. (2018)
Categories as Prototypes, Families, and Context-Sensitive Regions
Categories need not be mutually exclusive boxes with sharp boundaries. A prototype represents central or typical features, while less typical instances may vary in appraisal, bodily activation, expression, and social meaning. Family-resemblance organizations allow different instances of a category to share overlapping subsets of features without one feature appearing in every case. In a geometric representation, a category can occupy an extended region with fuzzy borders and gradated membership. Such representations accommodate meaningful classification while preserving internal heterogeneity. Russell & Barrett (1999) Cowen & Keltner (2017)
A prototype model predicts recognizable centers, yet it must specify how atypical instances are classified and what constitutes an error. If laughter appears in amusement, nervous appeasement, and social politeness, expressive similarity alone cannot decide among the categories. Context-sensitive classification instead combines features of the event, the person's aims, and the observed action. This motivates a hierarchy that keeps the dimension of pleasure distinct from the attribution of amusement and distinguishes a person's felt report from an observer's best social inference. Gendron et al. (2014) Keltner et al. (2019)
Hierarchical taxonomies can organize emotions into broad superordinate families and nested subcategories. A useful hierarchy should demonstrate greater within-group similarity than between-group similarity and must retain prediction when categories are tested on independent stimuli. The most visually attractive dendrogram is not necessarily uniquely determined: similarity metrics, sampling distributions, linkage criteria, and permitted labels all influence branching. Reporting alternative trees or uncertainty in membership is part of the evidential basis rather than a cosmetic technical choice. Lin et al. (2025)
Several theories therefore overlap in practice. A low-dimensional coordinate can locate a family of related experiences; a categorical model can specify regions or prototypes within that space; a semantic-space model can represent gradients of meaning between named experiences. Their difference lies in explanatory and predictive commitments. The key adjudicating test is whether the extra category structure predicts discriminable observations across methods after accounting for valence and arousal, and whether that improvement persists beyond the dataset used to discover the taxonomy. Cowen & Keltner (2021) Fontaine et al. (2007)
The Twenty-Seven-Category Study and Its Evidential Reach
Cowen and Keltner's 2017 investigation marks a substantial shift in the empirical scale of reported emotion taxonomy. Participants evaluated emotional experience elicited by 2,185 short video clips selected to cover richly varying evocative content. Statistical analyses of category judgments and additional reporting methods identified up to 27 reliably separable varieties of reported emotional experience. Category labels explained observed reports more effectively than only valence and arousal ratings for that stimulus domain. The finding concerns patterns of human reports induced by those videos, and its precision depends on the selected label vocabulary and analysis. Cowen & Keltner (2017)
The reported categories included familiar states and underrepresented experiences such as admiration, awe, amusement, aesthetic appreciation, nostalgia, romance, relief, and empathic pain. The investigators also observed smooth gradients connecting some adjacent categories, such as anxiety, fear, and horror. These results suggest that several emotional experiences are discriminable without every pair having a sharp natural boundary. The 27 dimensions are dimensions of reliable variation in categorical self-report, each interpretable by a distinct category, rather than 27 independent biological axes equivalent to the pleasure and activation coordinates of core affect. Cowen & Keltner (2017)
Study design shapes interpretability. A collection of evocative online clips was assembled with reference to candidate categories, and the response methods were optimized to detect those distinctions. This creates a serious question of selection-conditioned discovery: categories that were rarely evoked or were absent from the available vocabulary could not become equally visible. The split-half canonical-correlation procedure addresses reproducible patterns within the studied task, but reliable replication inside a stimulus inventory does not prove that the same number emerges from spontaneous events, childhood experience, other languages, or psychophysiological observation. Cowen & Keltner (2017) Lin et al. (2025)
The study is therefore strong evidence for a richer semantic organization of reported U.S. emotion experiences than a six-label or two-coordinate chart alone would express. It also provides an important constraint on claims that named emotions occupy isolated and discontinuous clusters. It leaves open whether the optimal latent space is stable across elicitors, languages, social contexts, and modalities. A scientific synthesis should preserve both its impressive stimulus coverage and the authors' explicit statement that the findings constitute an approximation of emotional experience structure, not a definitive inventory of nature. Cowen & Keltner (2017) Cowen & Keltner (2021)
Semantic Space Theory and High-Dimensional Organization
Semantic Space Theory treats the meaning of reported and attributed emotions as a geometry of relationships among categories and dimensions. Cowen and Keltner's 2021 synthesis describes computational methods for reconstructing the arrangement of states across subjective experience and expression. In this approach, categories become interpretable axes or regions within a multivariate space, while continuous paths connect related experiences. The framework makes a productive prediction: the structure can be compared across modalities and cultures, but the number of retained dimensions is likely to depend on what phenomena and judgments are sampled. Cowen & Keltner (2021)
One methodological advantage of semantic spaces is their capacity to preserve nuanced distinctions without committing to a small exhaustive dictionary. A participant may describe an episode as both relief and gratitude; the resulting report can be located by multiple categorical coordinates. Similarly, a visual or acoustic signal may convey mixed meaning rather than one exclusive emotion. Nevertheless, model flexibility increases the need for external tests. A high-dimensional space can fit an existing sample more closely while generalizing poorly if its additional coordinates capture noise, lexical conventions, or artifacts of stimulus selection. Cowen & Keltner (2017) Lin et al. (2025)
Another challenge is interpreting the word dimension. In a circumplex, a dimension is typically a broadly bipolar coordinate. In a large semantic embedding, a dimension may represent reliable separable variation tied to a category such as awe; its mathematical basis can be rotated without changing some predictions. Consequently, the statement that one experiment finds two dimensions and another finds 27 can conceal differing constructs, sample spaces, and reliability thresholds. Scientific comparison requires identifying the input variables, data transformation, criterion for retained dimensions, and the kind of generalization claimed. Lin et al. (2025)
Analytical synthesis points toward a layered picture. Valence and activation provide compact descriptive coordinates; categorical semantic structure preserves distinct experiences; relational and contextual features determine which category is meaningful in a particular episode. Each layer should be assessed for its incremental explanatory value. The most demanding outcome is not visually plausible arrangement but successful prediction of new reports, behavior, communication, or changes in context that competing models classify differently. Cowen & Keltner (2021) Han et al. (2024)
Taxonomies of Positive and Self-Conscious Emotions
Positive emotions reveal the inadequacy of treating every pleasant state as the same psychological phenomenon. Calm contentment, exuberant joy, interest, admiration, affection, gratitude, awe, and pride may share broadly pleasant tone while differing in attention, social orientation, perceived agency, and eliciting conditions. The study of affective taxonomy is enriched by distinguishing these functions through empirical categories. Cowen and Keltner's video-based findings include numerous positive categories, and the positive-emotion literature emphasizes the consequences of treating them as separable constructs. Cowen & Keltner (2017) Keltner et al. (2019)
Self-conscious emotions require a relation to a represented self or social identity. Pride can involve the interpretation of achievement or valued performance; shame and embarrassment involve other configurations of social evaluation, agency, exposure, and reputation. Tracy and Robins' 2004 experiments identified a recognizable posed expression of pride involving expanded posture and characteristic head and arm positions. Their evidence makes pride relevant to expressive taxonomy, while separating observable recognition from the claim that all pride episodes have identical facial and bodily signatures. Tracy & Robins (2004)
Categories such as gratitude and admiration may both be pleasant and socially directed, but their targets and implied relational structures differ. Awe can coexist with uncertainty or apprehension, and nostalgia can combine pleasant remembrance with loss. The significance of mixed or complex emotions is that single valence codes may collapse experiences whose temporal, social, and narrative implications are different. A sound taxonomy therefore permits category overlap or multiple labels and asks which differences are supported by participants' own reports rather than inferred from researchers' everyday use of terms. Cowen & Keltner (2017) Fontaine et al. (2007)
The scientific cost of expanding a category inventory is construct proliferation. New labels should identify separable patterns that improve reproducibility and theoretical explanation. If gratitude and appreciation cannot be distinguished in a particular dataset, combining them may be the parsimonious decision for that task. If contexts, reports, and behavioral consequences discriminate them reliably elsewhere, the category distinction becomes more useful. Taxonomy is a revisable empirical organization whose resolution should be matched to the measurable phenomenon. Lin et al. (2025) Cowen & Keltner (2021)
Emotion Vocabulary, Concepts, and Distinct Kinds of Similarity
Emotion language provides the most accessible observational material for taxonomy, but it also introduces its own structure. Words can resemble one another because of shared valence, arousal, social context, grammatical usage, translation history, or culturally familiar scripts. Two people can report very similar feelings using different terms; conversely, the same term can refer to substantially different experienced episodes. Research must specify whether the target is lexical semantic similarity, elicited feeling, or the relation between the two. A word-based map alone demonstrates organization of concepts under the task's linguistic constraints. Barrett (2004) Chaouch-Orozco et al. (2026)
Barrett's 2004 multimethod investigation addressed whether differences in emotion reports reflected the structure of emotion words or features of experience. The key construct, emotion granularity, concerns how precisely a person differentiates reported emotional states. This design illustrates why language is simultaneously a measuring instrument and a candidate mechanism influencing categorization. Individuals with different lexical repertoires may vary in naming capacity, attention to feelings, and the circumstances they recognize as distinct; these factors must be separated experimentally where possible. Barrett (2004)
Similarity judgments between words measure perceived proximity; category selection measures discrimination or label fit; free descriptions reveal participant-preferred vocabulary; physiological measures establish different observables. Combining modalities can show whether a linguistic structure transfers to elicited experience or expression. Independent agreement is more informative than interpreting a strong within-method correlation as cross-method validation. When one method includes linguistic category prompts and the other is genuinely open-ended, differences can reveal how much of the apparent taxonomy was supplied by the measurement instrument. Gendron et al. (2014) Lin et al. (2025)
Translated emotion names do not guarantee concept equivalence. Some languages distribute semantic distinctions differently or rely on context-specific expressions. Cross-language analysis should compare response functions, similarity relations, eliciting situations, and item interpretation before assuming that a shared English glossary names the same categories for every participant. A multilingual taxonomy is an empirical achievement requiring translation verification and measurement-invariance evidence; lexical overlap by itself is insufficient. Chaouch-Orozco et al. (2026) Cowen et al. (2024)
Operational Measurement: PANAS, SAM, and Dimensional Ratings
An affect measure is a rule for turning particular observations into interpretable scores. The PANAS uses two sets of affect adjectives to capture positive and negative affect under specified time instructions. It provides a standardized way to compare people or repeated assessments, but its factors should be understood in relation to the item content. The positively valenced items emphasize active engagement; the negatively valenced items emphasize distress-related states. A PANAS score therefore cannot be automatically substituted for pure hedonic valence or placed on a circumplex axis without evaluating the measurement model. Watson et al. (1988)
The Self-Assessment Manikin (SAM), developed and evaluated by Bradley and Lang, offers pictorial ratings of pleasure, arousal, and dominance. Its relative independence from elaborate emotion vocabulary makes it valuable when researchers wish to record broad affective reactions to pictures and other experimental materials. Yet the drawings still require instructions and interpretation; their dimensions describe self-reported experience, rather than directly recording bodily or neural activation. Comparing SAM with semantic differential instruments can examine whether different response formats preserve similar scores for the same stimuli. Bradley & Lang (1994)
Short repeated reports are valuable for reconstructing the affective landscape within one person's daily life. An assessment asking about the current moment has different memory demands than one asking for the past week. Repeated measures can distinguish a person's average affect from fluctuations around that average, and can examine whether category differentiation is stable across episodes. They also create reactivity, missingness, unequal event sampling, and dependence between successive observations. Statistical analyses must respect the resulting nested data structure rather than treating every prompt as an independent individual. Barrett (2004) Kuppens et al. (2013)
Scientific validity is specific to the target: a rating scale may reliably quantify pleasantness while weakly distinguishing anger from fear. Conversely, a categorical vocabulary may discriminate emotionally complex experiences yet offer an unstable measure of arousal. Validation should consider item representativeness, convergent evidence, test-retest or within-person reliability as appropriate, factor structure, discriminant predictions, and invariance across groups. Internal consistency is one property of a scale, not sufficient evidence that its score measures a unique natural kind of emotion. Watson et al. (1988) Lin et al. (2025)
Stimulus Selection and the Elicitation–Reporting Pipeline
The empirical taxonomy begins before any statistical analysis: the selection of events, videos, images, sounds, narratives, or social encounters determines the range of emotional reactions available for classification. A dataset rich in threat cues may produce differentiated fear, anxiety, horror, and disgust responses while offering limited evidence about gratitude or aesthetic appreciation. Researchers must therefore define the stimulus universe, describe the sampling mechanism, and explain whether stimuli were selected to represent predefined labels. Inference beyond that universe requires independent replication with qualitatively different elicitors. Cowen & Keltner (2017) Lin et al. (2025)
Laboratory elicitation provides control over presentation, timing, and response conditions but often samples brief, externally supplied events. Daily-life reports provide personal relevance and repeated observation while giving researchers less experimental control. Narratives permit contextual richness but add reading ability and inferential demands; expression tasks rely on observer attribution as well as the producer's internal state. Each design samples an aspect of affective organization. Their results become complementary only when construct labels and outcome definitions remain explicit. Han et al. (2024) Cowen et al. (2024)
Han and Adolphs' 2024 examination of emotions in narratives, videos, and everyday life contributes a direct comparison across elicitor formats. Their methodological relevance is that structures inferred from one type of stimulus can be evaluated against structures inferred from another, helping identify features that reflect shared organization rather than item-specific content. Such comparisons should use comparable judgment scales and explicitly model shifts in stimulus prevalence. A taxonomy that remains stable across task changes gains generality, but common broad axes alone do not establish complete category equivalence. Han et al. (2024) Lin et al. (2025)
An experimental pipeline should record recruitment, language, stimulus provenance, presentation order, response options, missingness, exclusion criteria, and methods used to select affective words. If a researcher selects images that experts already classified as joyful and then demonstrates strong recognition of joy within those images, a portion of the result reflects the design. Independent stimulus sampling, preregistered validation sets, and open reporting of ambiguous cases provide stronger tests than repeated application of one label set to one curated stimulus bank. Cowen & Keltner (2017) Gendron et al. (2014)
Dimensionality Reduction and the Problem of Counting Emotions
Principal component analysis, exploratory factor analysis, multidimensional scaling, canonical correlation, and machine-learning embeddings answer related but nonidentical questions. Principal components identify directions of maximal variance in an observed feature matrix; common-factor models estimate shared latent covariance under additional assumptions; multidimensional scaling reconstructs distances; canonical correlation identifies reproducible associations between variable sets. None of these procedures directly reveals the metaphysical number of emotions. The retained dimension count is a property of an explicit data-and-model pipeline. Lin et al. (2025)
Researchers should distinguish an exploratory representation of a dataset from a confirmed theory. Factor rotation changes the interpretability of estimated directions; different rotations may express the same underlying subspace with different named axes. Thresholds based on variance, eigenvalues, held-out prediction, reliability, or replication can select different dimensionalities. An interpretable 27-component solution and a strong two-component summary may therefore both be appropriate descriptions, provided the level of representation and predictive objective are stated. Cowen & Keltner (2017) Lin et al. (2025)
Lin and colleagues' 2025 methodological review analyzes the influence of stimuli, tasks, and algorithms on emotion spaces. A major lesson is that debates about the number of dimensions cannot be settled by comparing bare numbers extracted from different studies. A low-dimensional solution often captures broad recurring valence and arousal variation, while additional dimensions may recover more specific distinctions if the experiment samples them reliably. The scientific burden is to show incremental, stable information rather than simply selecting the model yielding the most impressive visual landscape. Lin et al. (2025)
Out-of-sample evaluation should hold out participants and, where practical, eliciting events or entire stimulus classes. A model learned from emotionally charged videos may perform differently when applied to narratives, ordinary social interactions, or music. Bootstrap uncertainty, model comparison, sensitivity to excluded labels, and tests of nonlinearity make dimensional interpretations more credible. The ultimate question is whether recovered axes or categories permit the intended scientific inference under a new sample, not whether the initial dataset can be displayed in a visually coherent map. Han et al. (2024) Cowen et al. (2020)
Physiology, Brain Activity, and the Limits of Cross-Modal Taxonomy
Biological measures offer important constraints on emotion classifications but cannot be treated as synonymous with reported categories. Changes in heart rate, respiration, electrodermal activity, and facial muscle activity may reflect overlapping bodily demands across affective conditions. Siegel and colleagues synthesized 202 laboratory studies of autonomic nervous system reactivity during emotion inductions, testing hypotheses about category-specific physiological fingerprints and context-sensitive populations of responses. Their multilevel meta-analysis documents why averaged physiology requires careful interpretation of between-study heterogeneity and within-category variation. Siegel et al. (2018)
Neuroimaging studies can examine whether particular emotional categories have reproducible neural patterns, whether affective dimensions explain activity, and how contextual features modulate those patterns. Lindquist and colleagues' 2012 meta-analytic review argued for interpreting emotion-related neural evidence through more distributed functional systems rather than assuming simple one-region-to-one-emotion correspondences. Brain-based classification depends on task, brain coverage, signal processing, and the number of trial examples; a classifier trained on a limited set of emotion labels may succeed without revealing a complete neural taxonomy. Lindquist et al. (2012)
Expression is another domain with its own evidential requirements. Keltner and colleagues review developments in emotional communication that encompass multiple modalities and numerous recognized affective states. Cordaro and colleagues examined free-response expression in five cultures and identified both shared expressive patterns and cultural differences. These findings establish that expressive meaning can exhibit regularities while also demanding culturally and contextually specific interpretation. An expressive category is not automatically equivalent to a participant's subjective feeling or to a neural activation pattern. Keltner et al. (2019) Cordaro et al. (2018)
A cross-modal taxonomy is strongest when categories discovered in one source predict distinct observations in another, beyond common valence and arousal. This could involve predicting conversational intentions from expressive meaning, evaluating naturalistic facial movement and self-reports within the same event, or testing whether category judgments generalize across unrelated stimulus corpora. Particular care is required when a machine classifier is supervised by human labels: high accuracy establishes recovery of the provided annotations, subject to sampling limits, and does not independently validate those labels as biological emotion essences. Cowen et al. (2024) Siegel et al. (2018)
Cross-Cultural and Cross-Linguistic Structure
Culture enters emotion taxonomy through linguistic concepts, eliciting conditions, normative display practices, attention, and the social meaning of responses. Cross-cultural similarity in a recognition task supports some common structure, but its degree depends on prompts and possible answers. Gendron and colleagues' 2014 study compared U.S. participants with Himba participants using emotion sorting without the usual conceptual cues and found important differences in organization. The result demonstrates the necessity of testing stimulus and response format when making claims of universal category mapping. Gendron et al. (2014)
Other comparative research identifies partial shared structure. Cordaro and colleagues studied free-response expressions of 22 emotions across five cultural settings, reporting both common patterns and systematic variations. Cowen and colleagues later examined emotional experience and expression across multiple countries and languages, reporting a substantial set of shared self-report distinctions along with variability across individuals and cultures. These programs imply that cross-cultural comparison should estimate the extent and domain of commonality rather than make a binary judgment that emotion categories are either universal or wholly culture-dependent. Cordaro et al. (2018) Cowen et al. (2024)
Cowen, Fang, Sauter, and Keltner examined emotional responses to music in the United States and China using 2,168 excerpts, reporting at least 13 distinguishable types of experience in their studied musical domain. The number is specific to this stimulus class and statistical approach; it should not be confused with a universal count of the emotions that humans can experience. Affective space may contain shared broad coordinates while music, faces, narratives, and social events furnish different distributions of specific feelings. Cowen et al. (2020)
A 2026 study by Chaouch-Orozco, Li, and Li asked native speakers of 15 languages to spatially arrange emotion concepts by perceived meaning. They documented similarities and systematic cross-language differences linked to cultural variables, including long-term orientation. Such semantic-arrangement evidence informs conceptual taxonomy, while leaving the lived structure of momentary emotion to be corroborated by experience-based measures. Cross-cultural validation therefore requires careful translation, matched procedures, diverse sampling, and tests of both conceptual and phenomenological comparability. Chaouch-Orozco et al. (2026)
Emotional Granularity and Individual Differences in Categorization
Emotional granularity, also called emotion differentiation, is the degree to which a person represents or reports distinct emotional experiences with fine resolution. Two individuals may assign similarly unpleasant valence to episodes but differ in whether they distinguish anger, disappointment, fear, guilt, and sadness. Barrett's 2004 work examined granularity as a problem at the intersection of feeling and language. The construct can describe within-person differentiation across real situations and should be kept distinct from vocabulary size, verbal fluency, or the number of categories offered by a questionnaire. Barrett (2004)
Repeated experience-sampling ratings often motivate differentiation indices based on the covariation of similarly valenced emotion terms across a person's sampled situations. When all unpleasant adjectives rise and fall together, the ratings offer little evidence that the individual systematically discriminates their meanings at the measured resolution. Yet the index also depends on the number of observations, actual emotional opportunities, scale use, and the valence and frequency of selected items. An apparently low-differentiation score may reflect restricted sampling or a narrow emotional range rather than a stable inability to distinguish experiences. Barrett (2004) Kashdan et al. (2015)
Tugade, Fredrickson, and Barrett examined resilience and positive emotional granularity in research incorporating psychophysiological and experience-sampling methods. Kashdan, Barrett, and McKnight later surveyed associations between differentiation and adjustment. These sources motivate the idea that more specific emotion representation can support more targeted responses to difficulty, but correlations alone cannot prove that increasing emotion vocabulary causes better mental health. Experimental manipulation, longitudinal mediation, and reliable measurement are required to establish such causal accounts. Tugade et al. (2004) Kashdan et al. (2015)
The taxonomic implication is substantial: emotional category structure may vary across people even within one language. A group-level map averaging experiences may conceal differences in semantic emphasis, preferred descriptions, and the actual distribution of events encountered. Multilevel representations can distinguish shared dimensions from individual deviations, while linguistic and situational assessments can test why those deviations arise. A classification system intended for individual monitoring should report uncertainty, avoid forcing rare states into fixed prototypes, and remain sensitive to person-specific patterns without sacrificing cross-person interpretability. Kuppens et al. (2013) Lin et al. (2025)
Mixed Emotions, Co-Occurrence, and Affective Ambivalence
Mixed emotions challenge a taxonomy that requires each episode to occupy exactly one positive or negative category. A person can experience gratitude for an ending relationship while grieving the loss it represents, or feel excitement about a new role alongside anxiety about uncertainty. The categories describe different appraisals and implications even when they are jointly reported at one moment. Broad valence may summarize an overall evaluation, while separate positive and negative ratings preserve co-occurrence. The measurement question is whether an instrument can represent simultaneous experiences and whether respondents understand the requested temporal window. Watson et al. (1988) Barrett & Russell (1999)
At the level of a two-axis chart, a mixed episode could be represented by an average or a trajectory, but each compression discards something different. Averaging strong pleasantness and unpleasantness can yield an apparently neutral score that incorrectly resembles indifference. A trajectory can show a switch between states yet conceal genuine simultaneity if the sampling interval is too coarse. Multi-label reports, component-level ratings, and dense temporal measurement can better distinguish mixtures from rapid alternation. These representational choices require validation against participants' immediate descriptions, rather than a decision made solely for a convenient visualization. Lin et al. (2025) Kuppens et al. (2013)
A useful taxonomy should therefore clarify its logic of category membership. A mutually exclusive model asks respondents to choose a dominant emotion. A graded model permits varying degrees of several labels. A hierarchical model assigns related categories to a common family without requiring identical emotional meaning. These alternatives make different predictions for mixed experience and for how people respond to ambiguous stimuli. Comparing their reliability and held-out prediction can reveal whether added flexibility captures stable experience or questionnaire noise. Cowen & Keltner (2021) Han et al. (2024)
Mixed emotions also illustrate the distinction between a momentary feeling and a person's interpretation of a complex event. Narrative reflection may connect pride and regret across years, while brief experimental clips evoke rapid changes in several affective responses. A taxonomy needs a specified unit of analysis: instantaneous rating, bounded episode, remembered narrative, or long-term relationship. Without this temporal and situational unit, disagreement about whether mixed states are genuinely simultaneous may reflect mismatched measurement windows rather than competing evidence about affective organization. Russell & Barrett (1999) Han et al. (2024)
Emotion Spaces Across Music, Images, Narratives, and Daily Life
Emotion taxonomy has become increasingly sensitive to stimulus modality. Musical excerpts can evoke combinations of tension, nostalgia, awe, and pleasure without presenting the explicit social narratives commonly used to classify interpersonal feelings. Cowen and colleagues' 2020 study used 2,168 musical samples across U.S. and Chinese participant groups and reported at least 13 distinguishable categories of music-related emotional experience. The finding reveals differentiation in that domain; its count should not be treated as a universal classification of the entire human affective repertoire. Cowen et al. (2020)
Videos introduce changing visual, auditory, and social context. A single clip may carry cues to the emotion expressed by a performer, the event experienced by a protagonist, and the viewer's own reaction. These three targets can diverge. A participant who identifies a frightened person on screen might report curiosity rather than fear. Therefore, a dataset of attributed emotions and one of feelings evoked by the same materials can yield different taxonomies. Distinguishing the target avoids attributing every recognized label to the observer's experienced state. Cowen & Keltner (2017) Keltner et al. (2019)
Written narratives can simulate complex events but depend on readers' knowledge, reading proficiency, and prior experiences. Han and Adolphs compared emotional reports for narrative situations, videos, and more than ten thousand everyday experiences, finding recurrent low-dimensional structure across these stimulus families. Their study also identified continuous gradients rather than sharply separated clusters in sampled emotion reports. This combination supports broad cross-domain regularities while showing why a category boundary is often a useful research approximation rather than an empirically observed discontinuity. Han et al. (2024)
Daily-life sampling adds ecological variation and individual relevance. An individual's ordinary episodes arise from personal relationships, schedules, cultural expectations, and contexts missing from curated experimental sets. The resulting distribution of emotions can differ from that produced by evocative laboratory stimuli even when both datasets share interpretable valence and arousal factors. Generalizability tests therefore should include participants' own events and should report whether the comparison concerns shared axes, category counts, category positions, or population prevalence. Han et al. (2024) Lin et al. (2025)
Developmental and Lifespan Conditions of Emotion Classification
Taxonomy depends partly on how a person learns to distinguish experiences. Developing emotion vocabulary, attentional skills, and familiarity with social situations can alter which differences are reportable or recognizable. A child may use a broad negative label across several contexts that an adult describes with distinct terms, while the child's behavior may still vary meaningfully with the situation. The scientific task is to distinguish differences in emotional experience from differences in linguistic response capacity or the ability to use a particular rating scale. These are related but separable developmental questions. Barrett (2004) Lin et al. (2025)
Comparisons across age groups must be designed for equivalent tasks. Pictorial pleasantness ratings might reduce language demands but remain sensitive to how images are interpreted. Open-ended verbal descriptions provide nuanced category evidence yet depend heavily on vocabulary. Recognition with predefined alternatives can mask different semantic understandings because a familiar word may be chosen by elimination. A developmental taxonomy should combine multiple methods, adjust for reading and cognitive demands, and test whether identical labels reflect similar experiences across ages. Bradley & Lang (1994) Gendron et al. (2014)
Lifespan changes may influence the events people encounter as well as their emotional representations. A sample of adult workplaces and romantic relationships affords situations that differ from those experienced in adolescence or old age. Cross-sectional differences in category use may reflect circumstance, cohort vocabulary, health, or response preferences. Longitudinal within-person designs offer stronger evidence of change in differentiation, although repeated-measures designs remain vulnerable to practice and changing stimulus distributions. The object of classification remains the measured experience in its context. Kuppens et al. (2013) Barrett (2004)
Research on lifespan emotional development belongs to a more specialized scientific dossier; its relevance here lies in a methodological boundary. A category discovered in a broad adult internet sample should not be presented as automatically developmentally invariant. Evidence for transport to a new age group requires matched elicitors, comprehensible response options, and predictive validation. Where such evidence is missing, a taxonomy can be treated as a useful adult reference while the precise developmental generality of its dimensions and categories remains open. Cowen & Keltner (2017) Lin et al. (2025)
Distinguishing Dimensional, Categorical, and Hybrid Models
Dimensional theories and category theories make overlapping but testably different claims. A compact dimensional model predicts that broad aspects of reports can be reconstructed from a small set of coordinates. A discrete-category model predicts structured differences associated with identified kinds of experience. A hybrid model posits both common affective background dimensions and category-specific information. Comparing them requires a defined outcome: explained self-report covariance, identification of eliciting circumstances, prediction of expressive signals, or prediction of action. A model's apparent superiority in one outcome cannot settle its adequacy for all of them. Russell (1980) Cowen & Keltner (2021)
Suppose two stimuli receive similar unpleasantness and activation ratings, yet one yields predominantly anger reports and the other fear reports. A two-axis rating system correctly summarizes broad feeling quality but cannot distinguish these labels from those coordinates alone. The taxonomic model might improve classification through perceived cause, power, threat, or agency. Whether that gain is meaningful depends on its stability across participant samples and on whether the added variables predict observable consequences. Mere assignment of labels by a researcher cannot establish such validity. Fontaine et al. (2007) Lin et al. (2025)
Conversely, a vocabulary containing dozens of partially redundant labels may predict a dataset accurately while obscuring shared psychological organization. Lower-dimensional models can reveal commonalities among labels and support parsimonious longitudinal tracking. Complexity must be evaluated against held-out performance and against the interpretability and availability of inputs. When the practical objective is broadly tracking mood or monitoring shifts in activation, two coordinates may be preferable to an unstable inventory whose categories cannot be reliably sampled. Watson et al. (1988) Yik et al. (2011)
Hybrid approaches become especially promising when their levels are explicitly linked. A probabilistic category model can allow several named states with similar coordinates but different contextual profiles. A semantic map can locate these states along graded boundaries. A physiological model can then test whether additional categories predict bodily differences beyond common affective activation. Such layered models should specify which components are measured, which are inferred, and which are proposed mechanisms. Only tests of convergent and incremental validity can determine whether the resulting structure represents more than overlapping descriptive vocabularies. Siegel et al. (2018) Cowen & Keltner (2021)
Replication, Measurement Invariance, and Generalizability
Reproducible affective taxonomies require more than a visually stable factor plot. Replication can ask whether a model yields similar spatial structure on new participants, new stimuli, new regions, and new response formats. A circular arrangement might be stable under adjective ratings yet change when words are replaced by free descriptions. A 27-category model might classify a curated set of short videos but perform differently in ordinary interaction. Each replication claim should identify exactly which feature replicated: number of components, item angles, distances, category membership, or prediction. Cowen & Keltner (2017) Lin et al. (2025)
Measurement invariance is especially important for cross-language and cross-age comparison. The same translated item may evoke different typical situations, levels of activation, or social implications. If the scale's loadings or response thresholds differ, comparisons of observed means may reflect altered measurement rather than actual emotional experience. Factor models can formally test aspects of invariance, while interviews, qualitative elicitation, and semantic similarity tasks can reveal why apparent differences emerge. Failure of strict invariance need not eliminate useful comparisons, but it narrows what their numbers can support. Fontaine et al. (2007) Chaouch-Orozco et al. (2026)
Reliability must be distinguished from external validity. Participants can consistently agree on categories for posed expressions even if everyday expressions are more context-sensitive. A dimensional measure may be internally consistent across items while insufficiently covering emotions central to the target population. Likewise, a statistically stable embedding can reproduce semantic conventions shared by an internet sample without measuring the organization of experience in another community. Robust evidence requires comparing the taxonomy against independently observed criteria and disclosing the limitations of the sampled domain. Gendron et al. (2014) Han et al. (2024)
Publication bias and analytic flexibility create additional problems. Studies yielding memorable maps or high classification accuracy are often easier to communicate than null or unstable solutions. Researchers may choose among many transformations, label subsets, distance metrics, and factor-retention thresholds. Preregistration where appropriate, public analysis code, clear reporting of all relevant analytic decisions, and multiverse or sensitivity analysis can reveal whether a proposed structure depends on a narrow modeling choice. The methodological recommendations of Lin and colleagues provide a useful foundation for such transparent comparisons. Lin et al. (2025)
Interpretation of Classifiers and Automated Emotion Recognition
Automated emotion recognition is frequently presented as a test of affective taxonomy, but the meaning of its output depends on its training labels. A classifier trained to distinguish anger, fear, and joy from annotated images learns to reproduce an annotation scheme under a particular data distribution. It does not establish that every labeled individual experienced the predicted emotion, nor that the same categories organize internal experience across modalities. Evaluation must specify annotation provenance, cultural sampling, contextual information, uncertainty, and performance on new populations. Keltner et al. (2019) Gendron et al. (2014)
Continuous prediction of valence and arousal offers an alternative to exclusive categories. It can be appropriate for music recommendation, exploratory human-computer interaction, or mapping subjective stimulus ratings, provided the system is evaluated against ratings actually collected from target participants. However, two numeric outputs do not confer transparency about mechanisms. If input video features correlate with the typical mood of an edited dataset, a model may achieve performance without identifying emotion from a person. External validity and constructive clarity remain necessary regardless of output format. Cowen et al. (2020) Lin et al. (2025)
Multi-label recognition can represent uncertainty and mixtures more faithfully than a single forced-choice label, yet it introduces calibration and evaluation challenges. A system should be assessed for whether it can say that evidence is ambiguous, whether its reported confidence corresponds to actual accuracy, and whether mistakes are systematically associated with demographic or cultural conditions. Models can also infer what an observer would call an expression rather than what its producer feels. Those are different scientifically useful tasks, and reporting should distinguish them. Cowen et al. (2024) Cordaro et al. (2018)
Within the Artificial Era, computational affective coordinates can be used to describe signals, outputs, reward variables, or user-inferred states of artificial systems. This is a legitimate modeling approach when the coordinates have operational definitions and predictive tests. The existence of an artificial valence score, a labeled embedding, or a system's emotion-related language is not sufficient evidence of felt pleasure or distress. Attribution of artificial subjective experience would require a separate evidential framework relating proposed mechanisms to independently testable indicators. Lin et al. (2025) Barrett (2017)
Practical Uses and the Ethics of Affective Classification
Emotion taxonomies contribute to psychological research by organizing stimuli, defining self-report variables, comparing populations, and specifying outcomes for studies of emotion dynamics. The choice between broad dimensions and finer categories should follow the intended inferential task. A study of global pleasantness may need one carefully validated coordinate, whereas a study of interpersonal communication needs to distinguish emotions associated with divergent social meanings. Clear construct definitions permit findings from separate laboratories to be compared without assuming that identical category names indicate identical measures. Russell (1980) Lin et al. (2025)
Measurement systems can also assist with repeated personal reports, educational research, and context-sensitive emotion understanding. A compact dimension scale reduces burden when participants respond frequently, but it can lose information about which situations elicit particular feelings. Fine-grained category ratings provide richer data while placing greater demands on attention and language. A hybrid assessment can combine a brief valence-arousal core with optional category descriptions, allowing researchers to examine whether finer distinctions add reliable predictive information. Such choices should be evaluated empirically rather than standardized across all populations by assumption. Bradley & Lang (1994) Barrett (2004)
Ethical concern arises when affective classification is treated as a direct reading of hidden mental states. An observer's inference about someone's fear, anger, or pleasure may be mistaken, particularly when context is absent or the person belongs to a population poorly represented in the training data. Automated systems magnify the consequences when tentative classification is used for surveillance, educational selection, employment, or consequential judgments. The scientific response is to constrain claims to validated outcomes, document uncertainty and generalization gaps, and distinguish voluntary first-person report from externally attributed emotion. Gendron et al. (2014) Siegel et al. (2018)
A taxonomy does not determine whether an emotion is appropriate, pathological, sincere, or socially desirable. Those judgments depend on additional normative, clinical, contextual, and interpersonal information. Even a reliable category score cannot alone establish diagnostic criteria or explain why a person acted. Maintaining this boundary strengthens the scientific use of classification by preventing the substitution of a descriptive coordinate for explanations, decisions, or moral evaluations requiring independent evidence. Keltner et al. (2019) Lin et al. (2025)
Outstanding Scientific Problems and Testable Research Questions
The first unresolved problem is whether an affective space has a stable intrinsic dimensionality across tasks and observational modalities. Existing work repeatedly finds broad valence and activation structure, but estimates of additional dimensions vary with the stimulus universe, response task, and algorithm. A decisive comparative program would sample natural episodes, videos, narratives, and other elicitors; collect both categorical and dimensional reports; and preregister model comparisons on held-out participants and held-out stimulus families. Such a design could clarify which dimensions generalize and which are conditioned on method. Han et al. (2024) Lin et al. (2025)
A second problem is the relationship between conceptual meaning and subjective affect. Cross-linguistic arrangement tasks demonstrate culturally patterned relations among emotion words; experience reports and physiological data address other levels of the emotional process. A research program could test the same participants across semantic arrangement, daily experience sampling, and stimulus evaluation, estimating which representational structures predict which outcomes. It would be especially important to assess translation quality, individual semantic expertise, task demands, and within-person stability before interpreting language differences as differences in emotional feeling. Chaouch-Orozco et al. (2026) Barrett (2004)
A third problem concerns whether emotion categories are better modeled as discrete clusters, fuzzy regions, or continuously varying gradients. Model evaluation should avoid using the same features both to define categories and to claim independent discovery of those categories. Multi-method evidence offers a stronger route: clusters derived from first-person reports might be examined for incremental prediction of expressive, behavioral, or contextual variables. Counterexamples and ambiguous episodes should be preserved rather than removed solely because they complicate the taxonomy. The competing models can then be compared by calibration, generalization, and explanatory reach. Cowen & Keltner (2017) Han et al. (2024)
A fourth problem is person-specific and developmental variation. Emotion differentiation appears to vary across individuals, yet the scientific meaning of a differentiation index can be influenced by item selection and the available distribution of situations. Better longitudinal designs would jointly model category reports, context, linguistic knowledge, intensity, and time. This would help distinguish genuine changes in representational resolution from shifts in daily experience, and would establish which aspects of shared taxonomy are stable enough to support comparison across people. Barrett (2004) Kashdan et al. (2015)
A final problem is the interpretation of computational emotion representations in artificial agents. A machine may encode a numerical valence estimate of user feedback or generate emotion descriptions that are useful for prediction. The correct research questions concern the observable target, the data used to construct that estimate, and its transfer to new settings. Inferences about intrinsic artificial feeling require additional operational criteria and independent theoretical and empirical work. Separating useful predictive classification from assertions about experience makes comparative research in the Artificial Era methodologically stronger. Lin et al. (2025) Barrett (2017)
Critical Scientific Synthesis: What the Evidence Supports
The evidence establishes that affective experience can be meaningfully summarized using broad dimensions, particularly pleasantness and activation. Russell's circumplex and the later core-affect literature provide a coherent account of relations among affective ratings, while PANAS and SAM demonstrate multiple ways of operationalizing important portions of affective variability. These instruments describe overlapping but distinguishable psychological constructs. The existence of a common low-dimensional structure does not imply that every episode is exhausted by those coordinates or that different rating procedures are automatically equivalent. Russell (1980) Russell & Barrett (1999) Watson et al. (1988)
The evidence also establishes informative distinctions among numerous reported emotional experiences. Cowen and Keltner's video study identified 27 category-related dimensions of variation, and their music research found at least 13 distinguishable varieties within that domain. These figures are conditioned on stimulus modality, vocabulary, sampling, and the analytical method. They challenge taxonomies restricted to a handful of exclusive labels, but they cannot establish one universal final count of all emotions. Cross-domain research favors graded relations and broad shared dimensions rather than simple categorical discontinuities. Cowen & Keltner (2017) Cowen et al. (2020) Han et al. (2024)
The physiology and expression literatures constrain stronger claims about category-specific biological signatures. Emotion words and labels may predict meaningful average differences, yet measurement heterogeneity, context, task design, and individual variation complicate one-to-one mapping from body or face to felt emotion. Cross-cultural evidence includes both shared expressive organization and differences in conceptual mapping. The most defensible taxonomy recognizes the distinct evidence supplied by verbal report, expression, autonomic physiology, and neural response rather than treating them as interchangeable indicators. Siegel et al. (2018) Lindquist et al. (2012) Cordaro et al. (2018) Gendron et al. (2014)
Three conclusions organize a practical scientific framework. First, every representation must identify its target: current feeling, semantic judgment, stimulus attribution, expression, or behavior. Second, its complexity should be selected by independent predictive value after accounting for common factors, not by an abstract preference for fewer or more dimensions. Third, generalization across contexts, languages, ages, individuals, and modalities is an outcome to be demonstrated. These principles permit dimensional, categorical, and hybrid models to be compared in a common evidential language while preserving meaningful differences in their scientific commitments. Fontaine et al. (2007) Lin et al. (2025)
Conclusion
Affective Dimensions and Emotion Taxonomy identifies a central scientific problem: how the diversity of emotional experience is structured, measured, compared, and represented. The history of the field shows a continuing movement between compact models of pleasantness and activation, inventories of differentiated emotion kinds, and newer multivariate spaces capable of describing gradual transitions among them. The principal empirical finding is that affective organization exhibits both broad regularities and distinctions that are meaningful at finer resolution. The kind of data being classified determines how far any particular representation can be generalized. Russell (1980) Cowen & Keltner (2021) Lin et al. (2025)
The strongest research designs connect clearly defined constructs to transparent measurement and independent evaluation. Self-reports, language judgments, psychophysiological data, and expressive observations can converge, but they also sample different features of emotional life. Distinguishing these measures clarifies why disagreements about emotion number and identity sometimes reflect different scientific tasks. Affective classification is therefore best approached as an empirical family of representational models whose predictive scope must be demonstrated rather than assumed. Han et al. (2024) Siegel et al. (2018)
Scientific progress depends on reliable cross-context replication, representative stimulus sampling, appropriate dimensionality-reduction methods, and direct tests of category information beyond broad valence and activation. Cross-cultural work and studies of emotional differentiation further show that semantic concepts and individual reporting patterns are integral to the evidential landscape. Future taxonomies should preserve uncertainty, account for mixed states, and explain what specific observations each proposed dimension or category improves. Chaouch-Orozco et al. (2026) Barrett (2004)
For Psychology by Bogdanova, the resulting research object is a stable reference point between the Science of Emotion and Motivation and more specialized studies of appraisal, emotion construction, temporal dynamics, and emotional communication. Within the Artificial Era, the same methodological standards enable careful comparison of human reports and artificial representations without equating a computed affective score with lived experience. The concluding position is a layered, explicitly operational science of affective organization, grounded in evidence and open to further discriminating experiments. Lin et al. (2025) Barrett (2017)
References
- Barrett, L. F. (2004). Feelings or words? Understanding the content in self-report ratings of experienced emotion. Journal of Personality and Social Psychology, 87(2), 266–281. https://doi.org/10.1037/0022-3514.87.2.266
- Barrett, L. F. (2017). The theory of constructed emotion: An active inference account of interoception and categorization. Social Cognitive and Affective Neuroscience, 12(1), 1–23. https://doi.org/10.1093/scan/nsw154
- Barrett, L. F., & Russell, J. A. (1999). The structure of current affect: Controversies and emerging consensus. Current Directions in Psychological Science, 8(1), 10–14. https://doi.org/10.1111/1467-8721.00003
- Berridge, K. C., & Kringelbach, M. L. (2015). Pleasure systems in the brain. Neuron, 86(3), 646–664. https://doi.org/10.1016/j.neuron.2015.02.018
- Bradley, M. M., & Lang, P. J. (1994). Measuring emotion: The Self-Assessment Manikin and the Semantic Differential. Journal of Behavior Therapy and Experimental Psychiatry, 25(1), 49–59. https://doi.org/10.1016/0005-7916(94)90063-9
- Chaouch-Orozco, A., Li, X., & Li, P. (2026). Cultural attitudes towards the future shape the semantic structure of emotion concepts. Humanities and Social Sciences Communications, 13, Article 387. https://doi.org/10.1057/s41599-026-06623-3
- Cordaro, D. T., Sun, R., Keltner, D., Kamble, S., Huddar, N., & McNeil, G. (2018). Universals and cultural variations in 22 emotional expressions across five cultures. Emotion, 18(1), 75–93. https://doi.org/10.1037/emo0000302
- Cowen, A. S., Brooks, J. A., Prasad, G., Tanaka, M., Kamitani, Y., Kirilyuk, V., Somandepalli, K., Jou, B., Schroff, F., Adam, H., Sauter, D., Fang, X., Manokara, K., Tzirakis, P., Oh, M., & Keltner, D. (2024). How emotion is experienced and expressed in multiple cultures: A large-scale experiment across North America, Europe, and Japan. Frontiers in Psychology, 15, Article 1350631. https://doi.org/10.3389/fpsyg.2024.1350631
- Cowen, A. S., Fang, X., Sauter, D., & Keltner, D. (2020). What music makes us feel: At least 13 dimensions organize subjective experiences associated with music across different cultures. Proceedings of the National Academy of Sciences, 117(4), 1924–1934. https://doi.org/10.1073/pnas.1910704117
- Cowen, A. S., & Keltner, D. (2017). Self-report captures 27 distinct categories of emotion bridged by continuous gradients. Proceedings of the National Academy of Sciences, 114(38), E7900–E7909. https://doi.org/10.1073/pnas.1702247114
- Cowen, A. S., & Keltner, D. (2021). Semantic space theory: A computational approach to emotion. Trends in Cognitive Sciences, 25(2), 124–136. https://doi.org/10.1016/j.tics.2020.11.004
- Ekman, P. (1992). An argument for basic emotions. Cognition & Emotion, 6(3–4), 169–200. https://doi.org/10.1080/02699939208411068
- Fontaine, J. R. J., Scherer, K. R., Roesch, E. B., & Ellsworth, P. C. (2007). The world of emotions is not two-dimensional. Psychological Science, 18(12), 1050–1057. https://doi.org/10.1111/j.1467-9280.2007.02024.x
- Gendron, M., Roberson, D., van der Vyver, J. M., & Barrett, L. F. (2014). Perceptions of emotion from facial expressions are not culturally universal: Evidence from a remote culture. Emotion, 14(2), 251–262. https://doi.org/10.1037/a0036052
- Han, Y., COVID-Dynamic Team, & Adolphs, R. (2024). A shared structure for emotion experiences from narratives, videos, and everyday life. iScience, 27(7), Article 110378. https://doi.org/10.1016/j.isci.2024.110378
- Kashdan, T. B., Barrett, L. F., & McKnight, P. E. (2015). Unpacking emotion differentiation: Transforming unpleasant experience by perceiving distinctions in negativity. Current Directions in Psychological Science, 24(1), 10–16. https://doi.org/10.1177/0963721414550708
- Keltner, D., Sauter, D., Tracy, J. L., & Cowen, A. (2019). Emotional expression: Advances in basic emotion theory. Journal of Nonverbal Behavior, 43(2), 133–160. https://doi.org/10.1007/s10919-019-00293-3
- Kuppens, P., Tuerlinckx, F., Russell, J. A., & Barrett, L. F. (2013). The relation between valence and arousal in subjective experience. Psychological Bulletin, 139(4), 917–940. https://doi.org/10.1037/a0030811
- Lin, C., Han, Y., Keles, U., Xu, Y., Lu, J., Zheng, R., & Adolphs, R. (2025). Investigating the structure of emotion: Tools, pitfalls and recommendations. Affective Science, 6(4), 726–744. https://doi.org/10.1007/s42761-025-00337-6
- Lindquist, K. A., Wager, T. D., Kober, H., Bliss-Moreau, E., & Barrett, L. F. (2012). The brain basis of emotion: A meta-analytic review. Behavioral and Brain Sciences, 35(3), 121–143. https://doi.org/10.1017/S0140525X11000446
- Mehrabian, A., & Russell, J. A. (1974). The basic emotional impact of environments. Perceptual and Motor Skills, 38(1), 283–301. https://doi.org/10.2466/pms.1974.38.1.283
- Russell, J. A. (1980). A circumplex model of affect. Journal of Personality and Social Psychology, 39(6), 1161–1178. https://doi.org/10.1037/h0077714
- Russell, J. A., & Barrett, L. F. (1999). Core affect, prototypical emotional episodes, and other things called emotion: Dissecting the elephant. Journal of Personality and Social Psychology, 76(5), 805–819. https://doi.org/10.1037/0022-3514.76.5.805
- Siegel, E. H., Sands, M. K., Van den Noortgate, W., Condon, P., Chang, Y., Dy, J., Quigley, K. S., & Barrett, L. F. (2018). Emotion fingerprints or emotion populations? A meta-analytic investigation of autonomic features of emotion categories. Psychological Bulletin, 144(4), 343–393. https://doi.org/10.1037/bul0000128
- Tracy, J. L., & Robins, R. W. (2004). Show your pride: Evidence for a discrete emotion expression. Psychological Science, 15(3), 194–197. https://doi.org/10.1111/j.0956-7976.2004.01503008.x
- Tugade, M. M., Fredrickson, B. L., & Barrett, L. F. (2004). Psychological resilience and positive emotional granularity: Examining the benefits of positive emotions on coping and health. Journal of Personality, 72(6), 1161–1190. https://doi.org/10.1111/j.1467-6494.2004.00294.x
- Watson, D., Clark, L. A., & Tellegen, A. (1988). Development and validation of brief measures of positive and negative affect: The PANAS scales. Journal of Personality and Social Psychology, 54(6), 1063–1070. https://doi.org/10.1037/0022-3514.54.6.1063
- Yik, M., Russell, J. A., & Steiger, J. H. (2011). A 12-point circumplex structure of core affect. Emotion, 11(4), 705–731. https://doi.org/10.1037/a0023980