Valence-Arousal-Dominance (VAD) Framework

Updated 2 December 2025

VAD is a dimensional framework that quantifies emotions along three continuous, orthogonal axes: valence, arousal, and dominance.
The model is operationalized through large-scale, crowdsourced lexicons and refined via deep learning and fuzzy logic for accurate affect prediction.
VAD bridges discrete emotion labels with continuous affect representations, enhancing applications in affective computing, speech synthesis, and social media analysis.

The Valence-Arousal-Dominance (VAD) dimensional framework constitutes a foundational model for representing, measuring, and predicting affective states in humans and artifacts such as text, speech, and images. VAD decomposes affect into three orthogonal, continuous axes—valence (pleasure-displeasure), arousal (activation-deactivation), and dominance (control-submissiveness)—with robust empirical and theoretical support from psychometrics, computational linguistics, affective computing, and neurocognitive science. This multidimensional space formalizes how emotions are encoded and decoded in language, perception, and artificial intelligence.

1. Formal Definitions and Theoretical Basis

The VAD framework emerged from classical factor-analytic studies (Osgood et al. 1957; Russell 1980), which demonstrated that human judgments of word meaning and affect reliably yield three principal components:

Valence (V): Measures the degree of positivity/negativity or pleasure/displeasure. Psychologically, it indexes the direction of behavioral activation—approach (appetitive) versus avoidance (aversive) (Mohammad, 30 Mar 2025).
Arousal (A): Encodes intensity, excitement, or energy versus calmness or passivity. This dimension is orthogonal to valence and modulates the magnitude of emotional responses (Mohammad, 30 Mar 2025).
Dominance (D): Captures the degree of perceived control, power, or influence over an emotion-inducing stimulus. High dominance implies feeling empowered, low dominance describes submission or being overpowered (Mohammad, 30 Mar 2025).

These axes underlie both semantic-evaluative word ratings and emotion terms. Multiple traditions interpret dominance with alternative labels (potency, competence), with cross-domain validation in psycholinguistics, social cognition, and NLP (Mohammad, 30 Mar 2025).

2. Lexicon Construction and Annotation Methodologies

VAD scoring for words and phrases has been realized via large-scale, crowdsourced lexicons. The NRC VAD Lexicon v2 defines standardized annotation protocols and reliability statistics for over 55,000 English entries (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025). Key aspects include:

Scale: Annotators rate each axis on a seven-point bipolar scale ( $-3$ to $+3$ ), with final scores rescaled to $[-1,1]$ (Mohammad, 30 Mar 2025).
Curation: Lexicon entries are drawn from high-frequency lemmas, psycholinguistic norms, and multiword expressions (MWEs), omitting proper names and rare words (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).
Aggregation: Per-term final score $s_i = \frac{\bar{r}_i}{3}$ , with $\bar{r}_i$ the mean raw rating for item $i$ (Mohammad, 25 Nov 2025).
Reliability: Split-half reliability (SHR) across 1,000 random splits exceeds $r=0.95$ for all axes (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).
Extension to MWEs: Valence, arousal, and dominance for MWEs are measured directly and compared to compositional predictions from their constituents, enabling systematic paper of emotional non-compositionality and idiomatics (Mohammad, 25 Nov 2025).

Best practices advise averaging across large token samples, using relative (not absolute) comparisons, and accounting for lexical ambiguity, particularly in domain-specific contexts (Mohammad, 30 Mar 2025).

3. Computational Methods for VAD Inference in Text and Multimodal Data

Contemporary computational pipelines leverage VAD lexicons and deep learning for affect prediction from text, images, speech, and multimodal sources (Mäntylä et al., 2016, Kervadec et al., 2018, Asif et al., 15 Jan 2024, Li et al., 24 Sep 2025):

Text Analysis: The range or mean of lexicon-derived VAD scores for tokens in a document forms the basis of affect mining. For example, bug reports with higher priority yield higher inferred arousal; bug vs. feature discussions yield lower valence (Mäntylä et al., 2016).
Fuzzy Modeling in VAD: To capture annotator and subjectivity uncertainty, interval type-2 fuzzy sets partition VAD into low/medium/high, generating probabilistic cuboid representations and supporting robust CNN-LSTM affect classification, as with EEG input (Asif et al., 15 Jan 2024).
Neural Embedding Approaches: Deep neural networks (ResNet, VAE) structure latent affect space along VAD axes, often improving performance over larger, entangled embeddings. Learned 3D representations (e.g., CAKE-3) can align with, or enhance upon, classical VAD geometry in facial-expression classification (Kervadec et al., 2018).
Unimodal and Multimodal Models: Dual-tower models independently predict VAD from speech and text, aligning predictions in a shared continuous VAD space and using uncertainty-aware fusion only when modalities are consistent (Li et al., 24 Sep 2025).
Emotion-to-VAD Mapping: Discrete emotion labels are mapped to continuous VAD space for conversion between categorical and dimensional annotation systems using fixed lexicon lookups or proxy-based human assessment (Wrobel, 16 Nov 2025, Jia et al., 12 Sep 2024).

4. Bridging Discrete and Dimensional Frameworks

VAD serves as a bridge between categorical emotion taxonomies and continuous affect representations:

Lexicon-based Mapping: Discrete emotion labels (e.g., “anger”, “joy”) are associated with specific points in VAD space, facilitating conversion between annotation regimes (Park et al., 2019, Wrobel, 16 Nov 2025, Jia et al., 12 Sep 2024). For instance, “joy” may map to high valence, high arousal, moderate-high dominance (Wrobel, 16 Nov 2025).
Computational Transformation: Earth Mover’s Distance or similar loss functions align model outputs with VAD-distributed categorical anchors, permitting zero-shot VAD regression from discrete-labeled corpora (Park et al., 2019).
Proxy-Based Human Mapping: Participants create animations as proxies for emotion labels and rate these on VAD scales to empirically anchor discrete classes in continuous space. This approach yields robust, reproducible mappings and supports dataset harmonization (Wrobel, 16 Nov 2025).
Cluster-based Bridging: K-means clustering in VAD space regroups open-vocabulary emotion terms into discrete categories (e.g., happy/sad/worried), enabling direct regression from rich multimodal inputs and subsequent mapping back to emotion labels (Jia et al., 12 Sep 2024).

These strategies support both interpretability and analytical flexibility, allowing cross-paper comparison, inter-corpus integration, and open-vocabulary emotion generation.

The VAD framework underpins a wide array of research and engineering applications:

Workplace and Productivity Analytics: Mining issue-tracker and communication data for VAD signatures reveals correlates of productivity, resolution time, and burnout risk, supporting unobtrusive well-being monitoring (Mäntylä et al., 2016).
Speech and Conversational Generation: Emotional text-to-speech (TTS) systems employ VAD (ADV) for continuous, interpretable emotion control, enabling both discrete (label-driven) and continuous (axis-driven) prosody modulation (Liu et al., 15 May 2025).
Political Stance and Social Media Analysis: Disentangling VAD in deep variational frameworks enables robust, interpretable stance detection, outperforming standard LLMs and transfer models, and enhancing generalization across topics (Xu et al., 26 Feb 2025).
Emotion Recognition in Multimodal HCI: Pipelines integrating facial, vocal, and textual cues perform dimensional regression in VAD space, supporting mapping to both basic and open-vocabulary emotion sets, and yielding improved performance and interpretability (Jia et al., 12 Sep 2024, Li et al., 24 Sep 2025).
Psycholinguistic and Digital Humanities Research: Massive VAD lexicons empower empirical studies of affect across historical texts, domain-specific content, and cross-linguistic corpora, providing quantitative measures of language-linked emotion and stereotype (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).
Biomedical Signal Processing: Fuzzy VAD modeling combined with EEG features for emotion recognition outperforms crisp or no-VAD approaches. It also supports generalization across individuals, increasing robustness of mental health monitoring systems (Asif et al., 15 Jan 2024).

6. Validation, Reliability, and Methodological Considerations

Reliability and validity of VAD-based analyses are ensured through rigorous annotation and computation:

Split-Half Reliability: High SHR ( $r > 0.95$ ) at the word and phrase level among annotators (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).
Perspective Sensitivity: Annotating from reader versus writer perspectives yields distinguishable VAD profiles and inter-annotator agreement; reader perspective ratings are richer in intensity and correlate more with mapping to categorical emotions (Buechel et al., 2022).
Statistical Best Practices: Use of Bonferroni correction for multiple comparisons, effect size reporting (Cohen's $d$ ), and k-fold cross-validation is standard in applied VAD modeling (Mäntylä et al., 2016).
Uncertainty Modeling: Explicit representation of confidence in VAD prediction (via variances or type-2 fuzzy sets) improves both interpretability and accuracy, especially in multimodal and cross-domain settings (Asif et al., 15 Jan 2024, Li et al., 24 Sep 2025).

7. Future Directions and Research Horizons

The VAD dimensional approach has catalyzed new inquiry and methodological expansions:

Multi-Word and Cross-Linguistic Expansion: NRC VAD v2 and related lexicons aim to cover thousands more MWEs and extend to non-English languages, supporting global and multilingual research (Mohammad, 25 Nov 2025, Mohammad, 30 Mar 2025).
Unified Emotional Generation: Neural architectures for speech and conversational AI increasingly integrate VAD for controllable, nuanced affective synthesis, as seen in state-of-the-art TTS (Liu et al., 15 May 2025).
Fuzzy and Probabilistic Modeling: Deeper integration of fuzzy logic and probabilistic partitioning of VAD space addresses inter-subject variability and annotation ambiguity (Asif et al., 15 Jan 2024).
Open-Set and Unsupervised Emotion Discovery: VAD-structured embedding spaces support detection and analysis of emotions beyond fixed taxonomies, paving the way for open-vocabulary emotional modeling and personalized affective computing (Jia et al., 12 Sep 2024, Kervadec et al., 2018).
Bridging Discrete-Dimensional Annotation Bottlenecks: Proxy-based and computational mapping frameworks increase interoperability of datasets, supporting large-scale multimodal emotion learning (Wrobel, 16 Nov 2025, Park et al., 2019).

The cumulative evidence demonstrates that the VAD framework offers a principled, quantifiable, and extensible scaffold for multidimensional emotion analysis, supporting both fine-grained research and scalable deployment in affective, social, and computational applications.