NRC VAD Lexicon: Emotional Analysis Resource

• NRC VAD Lexicon is an open-access resource assigning detailed valence, arousal, and dominance scores to over 55,000 English words and multiword expressions.
• It uses robust annotation protocols and psychometric validation, including high inter-annotator agreement and reliability metrics, to ensure precise emotion quantification.
• The lexicon supports diverse applications in computational linguistics, psychology, digital humanities, and public health through reproducible, quantitative emotional analysis.

The NRC Valence, Arousal, and Dominance (VAD) Lexicon is an open-access, large-scale resource that provides fine-grained, human-annotated scores for the emotional dimensions of over 55,000 English words and multiword expressions (MWEs). Its structure supports quantitative analyses of emotional tone in natural language across a variety of disciplines, including computational linguistics, psychology, digital humanities, and public health. Each entry assigns a numerical score to valence (positivity/negativity), arousal (activity/passivity), and dominance (power/submission), facilitating rigorous studies of emotional content in both individual words and compositional phraseology (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025, Qiu, 11 Dec 2025).

1. Resource Composition and Distribution Format

The NRC VAD Lexicon v2.1, published in 2025, extends the 2018 v1.0 release—originating with Saif Mohammad’s research group—to cover both unigrams and MWEs at scale. Version 2 comprises:

• Unigrams: 44,928
• Multiword Expressions (MWEs): 10,205
• Total entries: 55,133

Entries are distributed as plain-text UTF-8 CSV/TSV files, each with one lexical item per line, comprising:

term	valence ∈ [–1, 1]	arousal ∈ [–1, 1]	dominance ∈ [–1, 1]	num_annotations
"joy"	0.92	0.74	0.36	9
"funeral"	–0.78	–0.22	–0.50	8
"power surge"	0.54	0.63	0.55	8

Standard deviations and annotator counts are included, but the principal use case focuses on mean dimension scores. The lexicon is downloadable for non-commercial research from http://saifmohammad.com/WebPages/nrc-vad.html (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).

2. Emotional Dimensions: Valence, Arousal, and Dominance

The NRC VAD Lexicon operationalizes three theoretically distinct and largely orthogonal affective dimensions:

• Valence (V): Numeric mapping of pleasantness/unpleasantness, where high values signal pleasure (“joy”), and low values signal displeasure (“funeral”).
• Arousal (A): Degree of activation/energy, with high scores representing excitement or alertness (“explosion”), and low scores representing calmness or passivity (“breeze”).
• Dominance (D): Extent of power/control versus submission; high scores mean authority or agency (“command”), while low scores mean powerlessness or timidity (“timid”).
• In social cognition, dominance is also called “competence.”

Scores for each term are weighted means over ∼8 independent raters, mapped onto a continuous scale from –1.00 (most negative/inactive/submissive) to +1.00 (most positive/active/dominant). This scale is derived via a 7-point annotation protocol mapped from –3…+3 and linearly rescaled during post-processing (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).

3. Annotation Protocol and Psychometric Validation

Sampling and Annotation

• Unigram selection: Frequent English words from emotion/sentiment lexicons, word prevalence norms.
• MWE selection: Over 10,500 frequent MWEs, including noun compounds (“support system”), idioms (“kick the bucket”), and verb-particle constructions (“take off”), sourced from high-concreteness lists (Mohammad, 25 Nov 2025).
• Annotation platform: Amazon Mechanical Turk, targeting US/Canada native speakers.
• Quality control: Gold questions (~2%) with immediate feedback; annotators failing ≥20% are excluded.
• Annotator demographics: ~95% US, UK, Indian, Canadian English speakers; average age 34.

Reliability Metrics

• Split-half reliability (Spearman’s ρ/Pearson’s r):
  • Valence: ρ = 0.98, r = 0.99
  • Arousal: ρ = 0.97, r = 0.98
  • Dominance: ρ = 0.96, r = 0.96
• Cronbach’s alpha assesses internal consistency per dimension.
• Inter-annotator agreement: Mean pairwise correlations exceed r = 0.75 on held-out subsets. Standard error of the mean per term is ≈ 0.02 (Mohammad, 30 Mar 2025, Mohammad, 25 Nov 2025).

4. Scoring, Aggregation, and Integration Workflows

Lexicon Lookup and Aggregation

Lexicon entries are matched to text tokens without lemmatization or stemming; only lowercasing and basic punctuation stripping are applied to input text. Each token $w_i$ in a given text is looked up to retrieve its (v_i, a_i, d_i) triple. Aggregation of VAD scores across a document or chapter $c$ is performed via simple unweighted averaging over $N_{(c)}$ matched tokens:

$$X_c = \frac{1}{N_c} \sum_{i=1}^{N_c} X(w_i), \quad X \in \{v, a, d\}$$

Where $V_c$ is chapter valence, $A_c$ arousal, and $D_c$ dominance. Tokens not found in the lexicon are skipped (Qiu, 11 Dec 2025).

Python integration:

The resource is loaded into a Pandas DataFrame, mapped to dictionaries for fast lookup, and processed with NumPy and wordcloud libraries for visualization. Bigrams (MWEs) are handled as space-delimited terms. Example code for loading and annotation is provided in (Mohammad, 25 Nov 2025):

import pandas as pd lex = pd.read_csv("NRC-VAD-Lexicon-v2.csv") def annotate_text(text, lexicon): tokens = text.lower().split() scores = {'valence':[], 'arousal':[], 'dominance':[]} for i in range(len(tokens)): t1 = tokens[i] row = lexicon[lexicon.term==t1] if not row.empty: for dim in scores: scores[dim].append(float(row[dim].iloc[0])) if i+1 < len(tokens): t2 = f"{t1} {tokens[i+1]}" row2 = lexicon[lexicon.term==t2] if not row2.empty: for dim in scores: scores[dim].append(float(row2[dim].iloc[0])) return {dim: sum(vals)/len(vals) if vals else 0.0 for dim,vals in scores.items()}

5. Multiword Expressions and Compositionality

The expanded coverage of MWEs addresses idioms, noun compounds, and verb-particle constructs that are frequent and emotionally salient. Emotional compositionality is examined via binned analysis:

• MWE score derivation:
  • For bigram $t = (w_1, w_2)$, actual VAD score: $s_t$; constituent scores: $s_{w_1}, s_{w_2}$.
  • Partitioning the [–1, 1] range into seven discrete bins; building $7\times7$ matrices for mean MWE scores.
• High/low emotionality thresholds:
  • High: $\theta_{\rm high} = +0.33$
  • Low: $\theta_{\rm low} = -0.33$
• Findings:
  • ~69% idioms, ~65% noun compounds, ~63% particle verbs are non-neutral (valence).
  • Negative MWEs outnumber positive, reflecting general “negativity bias.”
  • MWE valence is moderately compositional, but arousal and dominance show substantial non-compositional deviation—numerous highly emotional MWEs have constituent words with neutral scores (Mohammad, 25 Nov 2025).

6. Applications and Limitations

Domain	Example Application	Note
Psychology	Societal affect dynamics, competence in children	Developmental focus
NLP	Sentiment/emotion features, word embeddings	Supervised features
Digital Humanities	Automated emotion arcs, narratology	Literary trajectory
Public Health	Crisis language in social media, campaign analysis	Mental health monitoring
Social Sciences	Political framing, stereotype analysis	Warmth/competence/dominance

Limitations:

• Scores capture only the predominant sense—contextual shifts (e.g., “cold” as temperature vs. emotion) are ignored.
• Polysemy, speaker intention, and irony are unmodeled.
• Proper nouns and out-of-vocabulary terms remain unscored.
• Annotator pool demographics and the static nature of the lexicon may introduce subtle bias and lack of temporal drift sensitivity.
• Supplementary methods (close literary reading, context-aware models) are necessary for full interpretive coverage (Qiu, 11 Dec 2025, Mohammad, 30 Mar 2025).

7. Future Directions

Planned and proposed expansions include:

• Cross-lingual VAD lexicons for major world languages (Chinese, Spanish, Hindi, Arabic, German).
• Context-aware VAD annotation, including sentence-level and dynamic modeling to capture semantic/affective drift over time.
• Ongoing community updates to address changes in word usage and affective associations (Mohammad, 30 Mar 2025).

The NRC VAD Lexicon v2 forms an indispensable reference for computational affect analysis, enabling reproducible, scalable quantification of the positivity, activation, and power connotations of English lexicon items in both research and practical applications.