Natural emotion vocabularies as windows on distress and well-being

The study examines whether the richness of actively used emotion vocabularies (EVs) in natural language corresponds with emotional functioning and well-being. Building on distinctions between active and passive vocabularies, the authors note that most prior work has assessed passive recognition or self-perceived emotional abilities rather than spontaneous emotion word use. Linguistic principles (e.g., Zipf’s law) and theories suggesting that language reflects mental habits, interest/expertise, and can shape experience motivate the hypothesis that active EVs align with lived emotional experiences. The authors predict that larger negative EVs will correspond to lower well-being, whereas larger positive EVs will correspond to higher well-being.

Prior research often assesses emotion-related abilities via passive tasks, such as multiple-choice tests of emotional intelligence, Likert ratings used to infer emotion differentiation, response latency measures, and labeling effects on neural regulation, all indicating benefits of recognizing and using emotion labels but not capturing spontaneous usage. Linguistics distinguishes active (produced) versus passive (recognized) vocabularies, which are not necessarily correlated and develop differently. Theoretical frameworks propose that word use reflects common mental operations (Zipf), interest/expertise in domains, and that language can scaffold and shape experience (Dewey; constructivist emotion theories; language-as-context). These perspectives suggest active EVs should broadly align with frequency and salience of emotional experiences and may subtly influence them. The study situates EV within validated text-derived markers (e.g., LIWC categories) known to relate to mental and physical health (I-words, we-words, health/illness, affiliation, achievement, leisure).

Two studies quantified active emotion vocabularies (EVs) and examined their associations with mood, personality, and well-being.

• EV definition and computation: EV is the diversity (not frequency) of emotion words spontaneously produced. Using custom software (Vocabulate) and LIWC-based dictionaries refined to 92 negative and 53 positive emotion terms, EV was calculated as: EV = (# unique emotion words / total words) × 100. Inflections of the same lemma were collapsed. Separate EV indices were computed for positive and negative emotions, and for emotion families: sadness, fear/anxiety, anger, and undifferentiated negative (e.g., bad, awful) to align with stressed mood.
• Study 1 (students): Participants were undergraduates in an online psychology class who completed 20-minute stream-of-consciousness essays at Time 1 (N=1,567 analyzable; ≥100 words and ≥70% LIWC-recognized tokens) and Time 2 (N=1,360) for test–retest. Mean age 18.8 (SD=2.0); 60.7% female. Essays averaged 665 words (SD=241) at Time 1 and 631 (SD=254) at Time 2. Measures included: general vocabulary size (type/token ratio of open-class words excluding emotion terms), LIWC cognitive processing, positive and negative emotional tone, and language markers of well-being (I-words, we-words, health/illness, affiliation, achievement, leisure). Self-reports: Big Five (BFI-44), single-item overall health (1–5), CESD-10 for depressive symptoms. State moods (sad, worried, angry, stressed; happy, enthusiastic, optimistic, calm) were rated pre- and post-writing (1–5) and averaged by valence.
• Study 2 (blogs): Public blogs from blogger.com (collected Aug 2004) were analyzed after applying the same inclusion criteria and de-duplication, yielding N=35,385 authors (text lengths 107–481,983 words; M=3,142; SD=6,572). Self-reported age and gender available for 27.4% (N=9,688; 52% female; age M=22.41, SD=8.06, range 13–88). Text-derived indices paralleled Study 1 (EV, general vocabulary TTR, cognitive processing, positive/negative tone, well-being categories).
• Reliability and statistics: Study 1 test–retest correlations assessed temporal stability. Study 2 split-half reliability computed by dividing each blog in half. Associations were estimated via Pearson and partial correlations (controlling for general vocabulary size and positive/negative emotional tone). Study 1 used 2,000 bootstrap replicates for 95% bias-corrected and accelerated CIs; Study 2 used 500 replicates. For mood change analyses (Study 1), partial correlations predicted post-writing target moods from corresponding emotion-family EVs, controlling for pre-writing levels of that mood, general vocabulary, and emotional tone.

Study 1 (students; N=1,567):

• Emotion word prevalence and EV size: 6.11% (SD=1.66) of essay words were affective (LIWC): positive M=3.62% (SD=1.28), negative M=2.40% (SD=1.15). EV means: negative EV M=0.55 (SD=0.36; ≈1 unique negative emotion word per 200 words), positive EV M=0.52 (SD=0.34; ≈1 unique positive emotion word per 200 words). Positive and negative EVs were modestly correlated (r≈0.16).
• Reliability: Test–retest stability was modest but significant (negative EV r=0.18, 95% CI 0.10–0.27; positive EV r=0.28, 95% CI 0.19–0.38; both p<0.001).
• Demographics: Negative EV associated with female gender; positive EV showed no bivariate gender relation; age unrelated to EVs.
• Construct validity: Both EVs correlated with LIWC cognitive processing, general vocabulary size (TTR), and corresponding emotional tone (e.g., negative EV with negative tone; positive EV with positive tone).
• Language markers of well-being: Negative EV associated with higher I-words and illness words and lower we-words and leisure words. Positive EV associated with higher achievement, affiliation, and leisure words; unexpectedly, also higher illness words.
• Self-reports: Negative EV correlated with higher neuroticism and depression, and lower overall health. Positive EV correlated with higher extraversion, agreeableness, and overall health, and lower neuroticism and depression. Many relationships remained significant after controlling for general vocabulary and emotional tone, indicating incremental validity.
• State mood convergence and specificity: Larger negative EV correlated with more negative mood both pre- (r=0.19, SE=0.02) and post-writing (r=0.21, SE=0.02). Larger positive EV correlated with more positive mood pre- (r=0.19, SE=0.02) and post-writing (r=0.22, SE=0.02) (all p<0.001). Emotion-family EVs predicted emotion-specific increases over writing when controlling covariates (partial rs): sadness EV→sadness change r=0.09***; fear EV→worry change r=0.09***; anger EV→anger change r=0.10***; undifferentiated negative EV→stress change r=0.09**; positive EV→positive mood change r=0.04 (p<0.10). Some cross-emotion negatives emerged (e.g., fear EV associated with less anger change; anger EV with less worry change).

Study 2 (blogs; N=35,385):

• EV size and reliability: Average unique emotion words per author: negative 6.55; positive 5.99. EV rates: negative M=0.29 (SD=0.21; range 0–2.66), positive M=0.33 (SD=0.21; range 0–2.49), ≈1 unique emotion word per ~300 words. Split-half reliabilities: negative EV r=0.27 (95% CI 0.26–0.29), positive EV r=0.28 (95% CI 0.27–0.29); p<0.001. Positive–negative EV correlation r=0.22 (p<0.001).
• Demographics and construct validity: Both EVs associated with female gender. EVs related to cognitive processing, general vocabulary, and corresponding emotional tone.
• Language markers of well-being: Negative EV correlated with more illness words and I-words, and fewer we-, leisure, and achievement words. Positive EV correlated with more achievement, leisure, and affiliation words; also positively with illness words and I-words (stronger than in Study 1). Most negative EV relationships, and several positive EV relationships, remained significant when controlling for general vocabulary and emotional tone, supporting incremental validity.

Overall pattern: Larger negative EVs align with markers of psychological distress and poorer physical health; larger positive EVs align with higher well-being and, in students, better psychosocial functioning. EVs show modest but acceptable stability and provide unique variance beyond general vocabulary size and emotional tone.

Findings support the hypothesis that active emotion vocabularies correspond with lived emotional experience and well-being. Negative EVs broadly tracked markers of distress (higher neuroticism and depression, poorer self-rated health, more self-focused and illness language, less social/leisure language), whereas positive EVs tracked markers of better adjustment (greater extraversion, agreeableness, achievement/affiliation/leisure themes, better self-rated health). The strong emotion-specific links between EV diversity and concurrent intensification of the corresponding mood suggest EVs index familiarity or expertise with those states and possibly a reflective tendency that can amplify present feelings. EVs demonstrated construct validity (associations with cognitive processing, vocabulary breadth, and tone) and incremental validity beyond general vocabulary and emotional tone, indicating EVs are not mere artifacts of verbal ability or text valence. However, causality cannot be inferred; EVs may reflect frequent experiences, interest/preoccupation, and/or language shaping experience. The blog analysis replicates and extends results in a large, heterogeneous sample, indicating generalizability beyond a student stream-of-consciousness task, while confirming that active EVs are typically small in natural language use.

The studies introduce and validate a computational method (Vocabulate) for quantifying active emotion vocabularies from natural text and demonstrate that EV diversity corresponds with well-being: larger negative EVs with greater distress and poorer health, and larger positive EVs with higher well-being. EVs show modest stability and unique predictive value beyond general vocabulary and emotional tone. The work provides a scalable tool and initial benchmarks for active EVs in everyday language and highlights their potential as semi-implicit markers of familiar emotional states. Future research should employ longitudinal and experimental designs to test causal mechanisms (e.g., whether generating multiple emotion labels intensifies states or aids regulation), refine measurement (education/intellectual functioning controls, richer health outcomes), integrate with related constructs (emotional intelligence, emotion differentiation, passive emotion knowledge), and examine clinical applications (e.g., whether positive EV may buffer maladaptive effects of rich negative EV).

Key limitations include: cross-sectional and observational designs precluding causal inference; student sample and introspective writing task in Study 1 limiting generalizability and potentially inflating EV; single-item self-reported health; lack of a direct education measure (only general vocabulary TTR as proxy); small effect sizes typical of word-count methods; ambiguity of word meanings (emotion terms not always used to describe internal states); EV best suited for population-level inferences rather than individuals; mixed and sometimes unexpected associations for positive EV (e.g., with illness and I-words); and partial availability of demographics in Study 2. Additional longitudinal and experimental work is needed to determine how EV changes relate to changes in well-being and to clarify adaptive versus maladaptive roles of expanding emotion vocabularies.