Emotional Characteristics of MWEs

• The paper introduces a comprehensive VAD lexicon for over 10,000 MWEs, ensuring precise measurement of emotional characteristics via high inter-annotator reliability.
• It employs a robust methodology using Amazon Mechanical Turk with ratings scaled to [-1, 1], achieving Cronbach’s α above 0.95 for Valence, Arousal, and Dominance.
• Findings reveal MWEs exhibit stronger emotionality than unigrams, with both compositional and non-compositional patterns across idioms, noun compounds, and verb–particle constructions.

Multiword expressions (MWEs) such as idioms, noun compounds, and verb–particle constructions constitute a substantial portion of natural language, and their emotional characteristics play a critical role in both human communication and automated language understanding. Emotional properties of MWEs are well-captured by the primary dimensions of Valence (V), Arousal (A), and Dominance (D), as established by classical factor-analytic work. Comprehensive, highly reliable human-annotated VAD norms for over 10,000 English MWEs are now available, enabling rigorous analysis of both the intensity and the compositionality of affect encoded in multiword units and facilitating a broad range of research in linguistics, NLP, psychology, and related fields (Mohammad, 25 Nov 2025).

1. Theoretical Framework: Valence, Arousal, and Dominance

Valence, Arousal, and Dominance (collectively, VAD) are three largely independent latent dimensions underlying connotative and affective meaning.

• Valence measures the pleasantness–unpleasantness (pleasure–displeasure, positivity–negativity) continuum.
• Arousal quantifies degree of activation (calm/passive to excited/energized).
• Dominance reflects the sense of control or power versus submissiveness.

Let $X_1, \dots, X_p$ denote observed semantic ratings. A common-factor model posits

$$X_j = \lambda_{j1}F_1 + \lambda_{j2}F_2 + \lambda_{j3}F_3 + \varepsilon_j,\quad j=1,\dots,p,$$

where factors $F_1, F_2, F_3$ correspond to V, A, and D, respectively. Empirical studies find that these three factors suffice to explain most variance in semantic differential tasks, with antonym pairs such as “good–bad,” “excited–calm,” and “dominant–submissive” aligning with Valence, Arousal, and Dominance (Mohammad, 25 Nov 2025).

2. Annotation Methodology and Reliability

The NRC VAD Lexicon v2 extends the annotation of VAD norms to 10,073 English MWEs and 25,089 unigrams via Amazon Mechanical Turk. Each term receives ratings from approximately nine annotators, with each participant rating a term on a seven-point scale (−3 to +3) for each dimension. Ratings are averaged and rescaled to [−1, 1]:

$v_i, a_i, d_i = 2\bar{s}_i/6 - 1$

where $\bar{s}_i$ is the mean score for term $i$.

Rigorous quality control is employed using interspersed “gold” questions; annotators with less than 80% accuracy are excluded. Ratings show extremely high inter-annotator reliability:

• Valence: $r_{\mathrm{SHR}}=0.99$
• Arousal: $r_{\mathrm{SHR}}=0.98$
• Dominance: $r_{\mathrm{SHR}}=0.96$

Cronbach’s $\alpha$ for all three dimensions exceeds 0.95, indicating exceptional internal consistency (Mohammad, 25 Nov 2025).

3. Descriptive Statistics and Global Trends

The MWE-VAD lexicon provides VAD scores for 10,073 MWEs (including bigrams such as “kick the bucket,” “breath of fresh air”) and, combined with NRC VAD v1, encompasses over 55,000 entries. Score distributions across the MWEs are roughly bell-shaped and centered near zero but exhibit a slight positivity bias for Valence. The proportion of MWEs with non-neutral VAD scores ($|score|>0.33$) is substantial:

• Valence: ≈ 67%
• Arousal: ≈ 60%
• Dominance: ≈ 62%

A notable finding is that MWEs are more likely to exhibit strong emotionality than unigrams. Specifically, 67% of MWEs are non-neutral on Valence, compared to ≈55% for unigrams (Cohen’s $d\approx0.3$; $p<0.001$ by Welch’s t-test). Similar, but less pronounced, results hold for Arousal and Dominance, indicating a general amplification of emotionality in multiword units—most markedly for idioms (Mohammad, 25 Nov 2025).

4. Emotional Compositionality and Non-compositionality

Compositionality measures the degree to which the emotional connotation of an MWE can be predicted from its constituents. For bigram MWEs, the constituent mean for a dimension (e.g., Valence) is

$$\mu_i^{\mathrm{cons}} = \frac{1}{2}(v(\text{word}_1^i)+v(\text{word}_2^i))$$

Pearson’s $r$ quantifies compositionality:

• Valence: $r_v\approx0.72$ (strong)
• Arousal: $r_a\approx0.55$ (moderate)
• Dominance: $r_d\approx0.50$ (moderate)

However, analysis by constituent classes uncovers significant non-compositional behavior. For instance, 4.8% of neutral–neutral MWEs have negative Valence ($v_i\le-0.33$), and 1.7% are strongly positive ($v_i\ge+0.33$)—demonstrating emergent, idiomatic emotional meanings unattainable by constituent-based prediction. This suggests that MWEs frequently encode affective information beyond what is inferable from their parts (Mohammad, 25 Nov 2025).

5. Variability Across MWE Types

MWEs encompass heterogeneous types—including idioms/fixed expressions, noun compounds, and verb–particle constructions—each exhibiting distinct emotional profiles. Comparative analysis using type annotations reveals:

MWE Type	Non-neutral Valence (%)	High Arousal (%)	High Dominance (%)
Idioms	69	28	29
Noun Compounds	65	35	34
Particle Verbs	63	26	27

Idioms are most likely to show strong Valence; noun compounds lead in both high Arousal and Dominance. ANOVA analyses indicate that idioms differ significantly from noun compounds and particle verbs in the strength of Valence ($F(2,10070)\approx12.4$, $p<0.0001$). Similar patterns appear for Arousal and Dominance, substantiating the variability of emotional characteristics by MWE subclass (Mohammad, 25 Nov 2025).

6. Illustrative Examples and Representative Patterns

High and low compositionality, as well as emergent affect, are observable in concrete cases:

• High compositionality (Valence): “sunny disposition” (constituent V ≈ +0.7, +0.3; MWE V ≈ +0.65).
• Low compositionality (Valence): “cold shoulder” (constituents ≈ 0; MWE V ≈ –0.55).
• Emergent high Arousal: “burst into tears” (constituents neutral; MWE A ≈ +0.75).
• Emergent low Dominance: “at the mercy of” (constituents neutral; MWE D ≈ –0.60).

This pattern evidences the prevalence of both compositional and non-compositional affect in MWEs, with a marked subset displaying idiomaticity in emotional meaning (Mohammad, 25 Nov 2025).

7. Research and Applied Implications

VAD norms for MWEs support an array of research and downstream applications:

• Natural language processing: Enhanced sentiment analysis that captures non-compositional, phrase-level emotion and enables more robust affect-sensitive text classification.
• Affective science and psychology: Examination of idiom-based emotion regulation, metaphoric mapping, and appraisal structures within MWEs.
• Public health monitoring: Analyzing MWE usage in social media to track population mood shifts (e.g., during pandemics).
• Digital Humanities: Quantifying emotional arcs in narratives, detecting the affective intensity of fixed expressions.
• Social and political analysis: Dissecting the framing and rhetorical impact of emotionally loaded phrases (e.g., “war on drugs,” “support system”).

The MWE-VAD lexicon provides a reliable, fine-grained framework for investigating both expression and compositionality of emotional meaning in multiword language, with wide interdisciplinary utility (Mohammad, 25 Nov 2025).