Big Five Personality

Big Five (OCEAN) Personality Profiling

Overview

Infer Big Five personality trait positions from linguistic markers in a text corpus, then map each trait to concrete writing style constraints for voice replication. The core principle: personality influences word choice in measurable but modest ways -- the mean absolute correlation between Big Five traits and LIWC categories is ~0.14 (Yarkoni, 2010), with the strongest single correlation at ~0.23. This means text-based personality inference reveals tendencies, never diagnoses. Treat all outputs as relative positions on a continuum with wide confidence bands, not as fixed personality scores.

Research foundation: The 2022 meta-analysis by Stachl et al. in Psychological Bulletin (N = 85,724 across 26 studies) established that 20 linguistic categories constitute a "kernel of truth" correlating with both self- and observer-reported Big Five traits. Self-report correlations are small (|r| = .08-.14); observer-report correlations are small-to-medium (|r| = .18-.39). The 52 LIWC categories collectively explain ~5.1% of self-reported personality variance, but ~38.5% of observer-reported variance.

When to Use

• Characterizing an author's personality tendencies from a corpus of their writing
• Building a personality profile to inform writing style replication or voice matching
• Quantifying linguistic markers associated with each OCEAN dimension
• Comparing a writing sample's trait indicators against population baselines
• Feeding trait positions into archetype classification or persona construction

When NOT to use:

• Diagnosing personality disorders or clinical conditions (this is NOT a clinical tool)
• Making hiring, screening, or gatekeeping decisions based on inferred traits
• Corpus contains fewer than 1,000 words (see Insufficient Data Handling)
• Text is translated, heavily edited by others, or ghostwritten (author signal is diluted)
• Single-context writing only (e.g., all work emails) -- context constrains language and masks personality

digraph personality_decision {
    rankdir=TB;
    "Enough text?\n(1000+ words)" [shape=diamond];
    "Multiple contexts\nor topics?" [shape=diamond];
    "Clinical/hiring\npurpose?" [shape=diamond];
    "Proceed with\nfull analysis" [shape=box];
    "Report insufficient\ndata" [shape=box];
    "Proceed with\ncontext caveat" [shape=box];
    "STOP: Not appropriate\nfor this use case" [shape=box];

    "Enough text?\n(1000+ words)" -> "Multiple contexts\nor topics?" [label="yes"];
    "Enough text?\n(1000+ words)" -> "Report insufficient\ndata" [label="no"];
    "Multiple contexts\nor topics?" -> "Clinical/hiring\npurpose?" [label="yes"];
    "Multiple contexts\nor topics?" -> "Proceed with\ncontext caveat" [label="no, single context"];
    "Clinical/hiring\npurpose?" -> "STOP: Not appropriate\nfor this use case" [label="yes"];
    "Clinical/hiring\npurpose?" -> "Proceed with\nfull analysis" [label="no"];
}

Quick Reference

The Five Traits and Their Linguistic Markers

Key Metric Thresholds

Workflow

Copy this checklist and track progress:

Big Five Personality Profiling Progress:
- [ ] Step 1: Validate corpus suitability and size
- [ ] Step 2: Compute linguistic marker frequencies per trait
- [ ] Step 3: Score each trait dimension (raw indicators)
- [ ] Step 4: Normalize scores and compute confidence bands
- [ ] Step 5: Map trait positions to writing style constraints
- [ ] Step 6: Generate the personality-informed writing profile
- [ ] Step 7: Write findings to docs/analysis/15-big-five-personality.md

Step 1: Validate Corpus Suitability

Before analysis, verify the corpus can support personality inference.

Suitability checks:

If corpus fails suitability: Report the failure. State what was checked, what failed, and why personality inference is unreliable for this corpus. Do not force analysis on unsuitable data.

Step 2: Compute Linguistic Marker Frequencies

For each trait dimension, count occurrences of validated linguistic markers across the corpus.

Marker categories to measure (per LIWC-validated research):

# Validated marker categories derived from Pennebaker & King (1999),
# Yarkoni (2010), and Stachl et al. (2022) meta-analysis.
# These are CATEGORY-LEVEL indicators, not exhaustive word lists.

TRAIT_MARKERS = {
    'openness': {
        'positive_indicators': [
            'articles (a, an, the)',           # r ~ .10-.15
            'prepositions',                     # r ~ .08-.12
            'long words (6+ characters)',       # vocabulary complexity
            'insight words (think, know, consider, understand)',
            'tentative words (maybe, perhaps, guess)',
            'high type-token ratio',            # vocabulary diversity
            'abstract nouns',
            'cultural/intellectual references',
        ],
        'negative_indicators': [
            'present-tense verbs (overuse)',
            'concrete/simple vocabulary',
            'low type-token ratio',
            'home/family words (inversely correlated)',
        ],
    },
    'conscientiousness': {
        'positive_indicators': [
            'achievement words (earn, win, complete, success)',
            'exclusion words (but, except, without)',  # precision
            'negations used for precision',
            'longer average sentence length',
            'discourse markers (therefore, however, furthermore)',
            'future-tense verbs',
            'organized paragraph structure',
        ],
        'negative_indicators': [
            'filler words (um, like, basically)',
            'informal abbreviations',
            'very short sentences (fragmented)',
            'present-tense hedging',
        ],
    },
    'extraversion': {
        'positive_indicators': [
            'social words (friend, talk, share, together)',
            'positive emotion words (happy, great, love, wonderful)',
            '1st person plural (we, us, our)',
            'references to people and groups',
            'exclamation marks',
            'informal/conversational register',
            'shorter average word length',
        ],
        'negative_indicators': [
            'articles (task-focused writing)',   # inversely correlated
            'long words',
            '1st person singular (I-focused)',
            'formal register',
            'prepositions (complex structure)',
        ],
    },
    'agreeableness': {
        'positive_indicators': [
            'positive emotion words',
            'affirmation/agreement words (agree, wonderful, thank)',
            'family words (home, mother, friend)',
            'social process words',
            'hedging/softening (perhaps, might, I think)',
            'compliments and praise',
            'inclusive language (we, together)',
        ],
        'negative_indicators': [
            'swear words',
            'anger words',
            'death/violence references',
            'money/power words',
            'direct commands/imperatives',
            'competitive language',
        ],
    },
    'neuroticism': {
        'positive_indicators': [  # HIGH neuroticism markers
            '1st person singular (I, me, my, mine)',  # strongest marker
            'negative emotion words (awful, terrible, worried)',
            'anxiety words (nervous, afraid, tense)',
            'sadness words (cry, grief, sad)',
            'anger words (hate, annoyed, furious)',
            'past-tense verbs',
            'certainty-reducing hedges (I guess, maybe, not sure)',
        ],
        'negative_indicators': [  # LOW neuroticism (emotional stability)
            'positive emotion words',
            'present-tense verbs',
            'certainty words (always, definitely, clearly)',
            'fewer self-references',
        ],
    },
}

Measurement approach:

1. Tokenize the corpus (preserve case for emphasis detection, then lowercase for counting)
2. Count words falling into each marker category using a validated word list (LIWC-22 if available, or open alternatives like VADER lexicon for sentiment categories, NRC EmoLex for emotion categories)
3. Compute the frequency rate per marker category: count / total_words * 100
4. Track structural markers separately: average sentence length, type-token ratio, punctuation patterns, pronoun ratios

import re
from collections import Counter

def compute_basic_markers(text):
    """Compute foundational linguistic markers for personality inference.
    This provides the raw measurements; trait scoring is in Step 3."""
    words = text.lower().split()
    total_words = len(words)
    if total_words == 0:
        return None

    unique_words = set(words)
    sentences = re.split(r'[.!?]+', text)
    sentences = [s.strip() for s in sentences if s.strip()]

    # Pronoun categories
    i_pronouns = {'i', 'me', 'my', 'mine', 'myself'}
    we_pronouns = {'we', 'us', 'our', 'ours', 'ourselves'}

    # Word length distribution
    long_words = [w for w in words if len(w) >= 6]

    markers = {
        'total_words': total_words,
        'unique_words': len(unique_words),
        'type_token_ratio': len(unique_words) / total_words,
        'avg_word_length': sum(len(w) for w in words) / total_words,
        'avg_sentence_length': total_words / max(len(sentences), 1),
        'long_word_pct': len(long_words) / total_words * 100,
        'i_pronoun_pct': sum(1 for w in words if w in i_pronouns) / total_words * 100,
        'we_pronoun_pct': sum(1 for w in words if w in we_pronouns) / total_words * 100,
        'exclamation_rate': text.count('!') / max(len(sentences), 1),
        'question_rate': text.count('?') / max(len(sentences), 1),
    }
    return markers

Step 3: Score Each Trait Dimension

Convert raw marker frequencies into relative trait positions. Each trait is scored on a continuum, NOT as a binary classification.

Scoring approach:

For each trait, compute a weighted composite from its marker categories. The weights reflect the meta-analytic correlation strengths from Stachl et al. (2022).

Scoring heuristic (when validated tools like LIWC-22 are unavailable):

For each trait dimension:

1. Count the frequency of positive indicators (markers associated with HIGH trait levels)
2. Count the frequency of negative indicators (markers associated with LOW trait levels)
3. Compute a directional score: (positive_rate - negative_rate) / (positive_rate + negative_rate)
4. This yields a raw score from -1.0 (strongly low-trait) to +1.0 (strongly high-trait)
5. Center of 0.0 = indeterminate / population average

Trait scoring table:

Step 4: Normalize Scores and Compute Confidence Bands

Normalization:

Convert raw directional scores (-1.0 to +1.0) to a 0-100 scale centered at 50:

normalized_score = 50 + (raw_score * 50)

Where 50 = population average (indeterminate), 0 = extreme low-trait, 100 = extreme high-trait.

Confidence bands:

The confidence band width depends on corpus size and context diversity:

Interpreting positions:

Midpoint scores (40-60) require special handling: If a trait scores in the 40-60 range, the confidence band likely spans from "low" to "high." Report this as "indeterminate" rather than "average." Average and indeterminate are different claims -- "average" implies knowledge of position; "indeterminate" correctly states the data cannot distinguish.

Step 5: Map Trait Positions to Writing Style Constraints

This is the critical translation step: converting trait scores into actionable writing constraints that enable voice replication.

Openness mapping:

Conscientiousness mapping:

Extraversion mapping:

Agreeableness mapping:

Neuroticism mapping (inverted as Emotional Stability for constraint clarity):

Step 6: Generate the Personality-Informed Writing Profile

Synthesize the five trait positions and their writing constraints into a unified writing profile. The profile should resolve constraint conflicts between traits (e.g., high Openness suggests complex sentences, but high Extraversion suggests shorter sentences -- resolve toward moderate complexity with varied rhythm).

Profile structure:

## Writing Voice Profile

### Trait Summary
| Trait | Score | Position | Confidence |
|-------|-------|----------|------------|
| Openness | [X] | [label] | +/- [N] |
| Conscientiousness | [X] | [label] | +/- [N] |
| Extraversion | [X] | [label] | +/- [N] |
| Agreeableness | [X] | [label] | +/- [N] |
| Neuroticism | [X] | [label] | +/- [N] |

### Composite Writing Constraints
- **Vocabulary level:** [simple / moderate / sophisticated / varied]
- **Sentence complexity:** [short-direct / moderate / long-complex / varied-rhythm]
- **Emotional tone:** [restrained / balanced / expressive / warm / intense]
- **Social orientation:** [task-focused / balanced / people-focused]
- **Formality:** [casual / moderate / formal / context-dependent]
- **Assertion style:** [hedged / balanced / direct / blunt]
- **Temporal orientation:** [past-reflective / present-focused / future-oriented]
- **Self-reference density:** [low / moderate / high]

Resolving trait conflicts:

When two traits push writing constraints in opposite directions, resolve by:

1. Prioritize the stronger-signal trait (the one further from 50 with the narrower confidence band)
2. Document the tension in the profile ("High Openness suggests complex vocabulary, but high Extraversion favors conversational brevity -- resolve toward varied vocabulary in casual structures")
3. Never silently drop a constraint -- if a conflict exists, name it

Step 7: Write the Report

Write all findings to docs/analysis/15-big-five-personality.md.

Report Output Template

The final report MUST be written to docs/analysis/15-big-five-personality.md with this structure:

# Big Five (OCEAN) Personality Profile

## Methodology
- **Approach:** Linguistic marker frequency analysis mapped to Big Five personality dimensions
- **Corpus:** [N words scored, N documents/texts, date range, source description]
- **Marker framework:** LIWC-validated categories (Pennebaker & King 1999; Yarkoni 2010; Stachl et al. 2022)
- **Scoring method:** Directional composite scoring from marker frequencies, normalized to 0-100 scale
- **Confidence calibration:** Based on corpus size ([N] words) and context diversity ([single/multiple])

## Corpus Suitability Assessment
- **Word count:** [N] words ([sufficient / insufficient / marginal])
- **Language:** [English / mixed]
- **Authorship:** [Single / multiple / unknown]
- **Context diversity:** [Single context / multiple contexts -- list them]
- **Editing level:** [Natural / lightly edited / heavily edited]
- **Temporal span:** [Single occasion / multiple occasions over N days/months/years]
- **Overall suitability:** [Suitable / suitable with caveats / unsuitable]

## Trait Scores

| Trait | Raw Score | Normalized (0-100) | Position | Confidence Band | Interpretation |
|-------|-----------|--------------------|-----------|--------------------|----------------|
| Openness | [+/-X.XX] | [0-100] | [label] | +/- [N] | [1-sentence interpretation] |
| Conscientiousness | [+/-X.XX] | [0-100] | [label] | +/- [N] | [1-sentence interpretation] |
| Extraversion | [+/-X.XX] | [0-100] | [label] | +/- [N] | [1-sentence interpretation] |
| Agreeableness | [+/-X.XX] | [0-100] | [label] | +/- [N] | [1-sentence interpretation] |
| Neuroticism | [+/-X.XX] | [0-100] | [label] | +/- [N] | [1-sentence interpretation] |

## Linguistic Evidence per Trait

### Openness
- **Marker frequencies:** [table of measured indicator rates vs. corpus baseline]
- **Key evidence:** [most prominent indicators and their rates]
- **Confidence notes:** [what strengthens or weakens confidence in this score]

### Conscientiousness
[Same structure]

### Extraversion
[Same structure]

### Agreeableness
[Same structure]

### Neuroticism
[Same structure]

## Writing Style Constraints (for Voice Replication)

### Composite Profile
| Dimension | Constraint | Derived From |
|-----------|-----------|--------------|
| Vocabulary level | [constraint] | [trait(s)] |
| Sentence complexity | [constraint] | [trait(s)] |
| Emotional tone | [constraint] | [trait(s)] |
| Social orientation | [constraint] | [trait(s)] |
| Formality | [constraint] | [trait(s)] |
| Assertion style | [constraint] | [trait(s)] |
| Temporal orientation | [constraint] | [trait(s)] |
| Self-reference density | [constraint] | [trait(s)] |

### Trait Conflict Resolutions
[Document any conflicts between trait-derived constraints and how they were resolved]

### Exemplar Sentences
[3-5 sentences that demonstrate the constrained writing style, drawn from or inspired by the corpus]

## Limitations and Caveats
- Text-based personality inference captures tendencies, not diagnoses. The strongest individual LIWC-trait correlation is ~0.23 (Yarkoni, 2010).
- LIWC categories collectively explain ~5.1% of self-reported personality variance (Stachl et al., 2022). Results are probabilistic indicators, not deterministic measures.
- Writing context strongly affects language: the same person writes differently in email vs. social media vs. academic papers. Scores reflect the writing context as much as the person.
- [Corpus-specific limitations from Step 1]
- [Confidence band acknowledgment: scores of X and Y overlap at their confidence intervals, meaning they may not be truly different]
- These results should NOT be used for clinical assessment, hiring decisions, or any evaluative purpose. They are intended solely for writing style characterization.

## Population Baseline Comparison
- All scores are normed against a theoretical population mean of 50.
- Without a matched comparison corpus, absolute positions should be interpreted cautiously.
- The profile is most useful as a RELATIVE description (which traits are strongest/weakest for this author) rather than an absolute placement.

## References
- Pennebaker, J.W. & King, L.A. (1999). Linguistic styles: Language use as an individual difference. *Journal of Personality and Social Psychology*, 77(6), 1296-1312.
- Yarkoni, T. (2010). Personality in 100,000 Words: A large-scale analysis of personality and word use among bloggers. *Journal of Research in Personality*, 44(3), 363-373.
- Stachl, C., et al. (2022). The kernel of truth in text-based personality assessment: A meta-analysis of the relations between the Big Five and LIWC. *Psychological Bulletin*.
- Hirsh, J.B. & Peterson, J.B. (2009). Personality and language use in self-narratives. *Journal of Research in Personality*, 43(3), 524-527.
- Boyd, R.L. & Pennebaker, J.W. (2017). Language-based personality: A new approach to personality in a digital world. *Current Opinion in Behavioral Sciences*, 18, 63-68.
- Tausczik, Y.R. & Pennebaker, J.W. (2010). The psychological meaning of words: LIWC and computerized text analysis methods. *Journal of Language and Social Psychology*, 29(1), 24-54.

Good Patterns

• Use validated marker lists from meta-analytic research (Stachl et al. 2022, Yarkoni 2010) rather than intuition-based word lists
• Report confidence intervals on every trait score -- never present a point estimate without a band
• Treat results as tendencies ("the writing suggests moderate openness") not diagnoses ("this person is open")
• Map traits to writing constraints -- the practical output is a replicable writing profile, not a personality label
• Compare against population baselines -- raw marker frequencies without normalization are uninterpretable
• Document corpus limitations prominently, not as a buried footnote
• Score all five traits even when some are indeterminate -- an indeterminate score is information ("the data does not distinguish this trait")
• Resolve trait conflicts explicitly when mapping to writing constraints -- never silently drop conflicting evidence

Anti-Patterns

Boundaries

This skill SHOULD produce:

• Trait position scores (0-100) for all five OCEAN dimensions with confidence bands
• Linguistic marker frequency tables showing evidence per trait
• Writing style constraints derived from trait positions
• Composite writing voice profile for style replication
• Explicit documentation of corpus limitations, confidence levels, and methodological caveats
• Written report at docs/analysis/15-big-five-personality.md

This skill should NOT:

• Diagnose personality traits, disorders, or clinical conditions
• Claim clinical validity or psychometric equivalence with validated personality instruments (NEO-PI-R, BFI-2)
• Be used for hiring, screening, admissions, or any evaluative decision about a person
• Present scores without confidence bounds
• Ignore that writing context affects language (professional email vs. casual social media vs. creative writing)
• Claim accuracy beyond what meta-analytic evidence supports (~5% of self-reported variance explained)
• Produce a profile from fewer than 1,000 words
• Treat a single writing context as representative of whole personality
• Conflate "presented writing identity" with "actual personality"

Insufficient Data Handling

digraph insufficient_data {
    rankdir=TB;
    node [shape=diamond, style=filled, fillcolor="#f0f0f0"];

    start [label="Total word count?", shape=diamond];
    tiny [label="< 1,000 words", shape=box, fillcolor="#ffcccc"];
    small [label="1,000-3,000 words", shape=box, fillcolor="#fff3cc"];
    moderate [label="3,000-5,000 words", shape=box, fillcolor="#ffffcc"];
    good [label="5,000+ words", shape=box, fillcolor="#ccffcc"];

    start -> tiny [label="< 1,000"];
    start -> small [label="1,000-3,000"];
    start -> moderate [label="3,000-5,000"];
    start -> good [label="5,000+"];

    tiny_action [label="STOP: Report insufficient data.\nDo NOT compute trait scores.\nReport only raw marker counts.", shape=box, fillcolor="#ffcccc"];
    small_action [label="Compute scores with\nLOW CONFIDENCE flag on every trait.\nConfidence band: +/- 20-25 points.\nSkip writing constraint mapping.", shape=box, fillcolor="#fff3cc"];
    moderate_action [label="Full analysis.\nModerate confidence.\nBand: +/- 15 points.\nMap writing constraints with caveats.", shape=box, fillcolor="#ffffcc"];
    good_action [label="Full analysis pipeline.\nBand: +/- 10-12 points.\nFull writing constraint mapping.", shape=box, fillcolor="#ccffcc"];

    tiny -> tiny_action;
    small -> small_action;
    moderate -> moderate_action;
    good -> good_action;
}

Common Mistakes

References

• Pennebaker, J.W. & King, L.A. (1999). Linguistic styles: Language use as an individual difference. Journal of Personality and Social Psychology, 77(6), 1296-1312.
• Yarkoni, T. (2010). Personality in 100,000 Words. Journal of Research in Personality, 44(3), 363-373.
• Stachl, C., et al. (2022). The kernel of truth in text-based personality assessment: A meta-analysis. Psychological Bulletin.
• Hirsh, J.B. & Peterson, J.B. (2009). Personality and language use in self-narratives. Journal of Research in Personality, 43(3), 524-527.
• Boyd, R.L. & Pennebaker, J.W. (2017). Language-based personality: A new approach. Current Opinion in Behavioral Sciences, 18, 63-68.
• Tausczik, Y.R. & Pennebaker, J.W. (2010). The psychological meaning of words. Journal of Language and Social Psychology, 29(1), 24-54.
• LIWC-22 Manual: Development and Psychometrics.
• IBM Watson Personality Insights: Validation and Methodology.
• Frontiers in Psychology: Linguistic measures of personality in group discussions (2022).