---
title: 'NRC VAD Lexicon: Emotional Analysis Resource'
url: https://www.emergentmind.com/topics/nrc-vad-lexicon
type: topic
---

# NRC VAD Lexicon: Emotional Analysis Resource

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 [2503.23547][2511.19816][2512.10865].

## 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 [2503.23547][2511.19816].

## 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 [2503.23547][2511.19816].

## 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 [2511.19816].
- **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 [2503.23547][2511.19816].

## 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 [2512.10865].

**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 [2511.19816]:

```python
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 [2511.19816].

## 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 [2512.10865][2503.23547].

## 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 [2503.23547].
  
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.

Source: https://www.emergentmind.com/topics/nrc-vad-lexicon