stylometry-python

Authorship attribution and stylometric analysis in Python.

A lightweight, dependency-minimal library for measuring writing style, attributing authorship, and detecting stylistic shifts introduced by LLMs.

pip install stylometry-python

What is stylometry?

Stylometry is the statistical analysis of writing style. Every author has unconscious stylistic habits — frequency of function words, sentence length patterns, punctuation choices — that form a measurable fingerprint.

Mosteller & Wallace used it to resolve the Federalist Papers authorship debate in 1964. Patrick Juola used it to identify JK Rowling behind the pseudonym Robert Galbraith in 2013.

This library makes those techniques accessible in 5 lines of Python.

Quickstart

from stylometry import StyleAnalyzer

sa = StyleAnalyzer()

# Fit on known texts
sa.fit(zola_texts, label="Zola")
sa.fit(maupassant_texts, label="Maupassant")

# Attribute an unknown text
predicted, distances = sa.predict(unknown_text)
print(f"Predicted author: {predicted}")
# → Predicted author: Zola

# Measure stylistic shift (original vs LLM rewrite)
shift = sa.shift(original_text, gpt_rewrite)
print(f"Stylistic shift: {shift:.4f}")
# → Stylistic shift: 0.2409

Installation

pip install stylometry-python

Dependencies: numpy, matplotlib, scikit-learn — nothing else. Works 100% offline. No API keys. No GPU.

Development setup

On macOS (Homebrew Python), use a virtual environment to avoid externally-managed-environment errors:

python3 -m venv .venv
source .venv/bin/activate
python3 -m pip install --upgrade pip
python3 -m pip install -r requirements-dev.txt

Equivalent direct command:

python3 -m pip install -e ".[dev]"

Run tests:

python3 -m pytest

For a full local workflow (venv, tests, coverage, lint, format), see docs/DEVELOPMENT.md.

Continuous Integration

GitHub Actions runs lint + tests on each push and pull request:

• ruff check .
• black --check .
• pytest --cov=stylometry --cov-report=term-missing

Workflow file: .github/workflows/ci.yml

Publishing

Package publication is automated after a successful Release workflow run. Tags are generated from Conventional Commits by semantic-release. Semantic release runs only after CI passes on the target branch.

Release workflow: .github/workflows/release.yml Workflow file: .github/workflows/publish.yml

Core API

StyleAnalyzer(function_words=None, language='fr', min_words=50)

The main class. Handles vectorization, attribution, and visualization.

from stylometry import StyleAnalyzer

# French (default) — 41 function words
sa = StyleAnalyzer()

# Custom vocabulary
sa = StyleAnalyzer(function_words=['the', 'of', 'and', 'to', 'a', 'in'])

# English preset
sa = StyleAnalyzer(language='en')

vectorize(text) → np.ndarray

Convert a text to a style vector (L2-normalized function word frequencies).

v = sa.vectorize("Il pleuvait a verse. La nuit etait noire...")
print(v.shape)  # (41,)
print(v.sum())  # ≈ 1.0 after normalization

fit(texts, label) → self

Compute a centroid from a list of texts. Chainable.

sa.fit(zola_corpus, "Zola").fit(maupassant_corpus, "Maupassant")

predict(text) → (label, distances)

Attribute a text to the nearest centroid.

predicted, distances = sa.predict(unknown)

print(predicted)           # "Zola"
print(distances)           # {"Zola": 0.12, "Maupassant": 0.43}
print(sa.confidence(distances))  # "HIGH" / "MEDIUM" / "LOW"

shift(original, rewrite) → float

Measure the cosine distance between two texts in style space. Use this to quantify how much an LLM changed the style of a text.

shift = sa.shift(original, gpt4_rewrite)
# 0.00 = style unchanged
# 0.24 = significant shift (typical GPT-4)
# 1.00 = maximally different

cosine_distance(text_a, text_b) → float

Direct cosine distance between two texts.

d = sa.cosine_distance(text_a, text_b)

Visualization

plot_fingerprint(texts_dict, top_n=15)

Bar chart comparing function word frequencies across groups.

fig = sa.plot_fingerprint(
    texts_dict={
        "Zola": zola_corpus,
        "Maupassant": maupassant_corpus,
        "GPT-4": gpt4_corpus,
    },
    top_n=12,
    title="Writing fingerprints",
)
fig.savefig("fingerprints.png", dpi=150)

plot_clusters(texts_groups, labels)

PCA scatter plot — visualize stylistic distances between groups.

fig = sa.plot_clusters(
    texts_groups=[zola_corpus, maupassant_corpus, gpt4_corpus],
    labels=["Zola", "Maupassant", "GPT-4"],
    title="Do LLMs form a distinct stylistic cluster?",
)

plot_shift_distribution(originals, rewrites_dict)

Box plot of cosine shifts per model.

fig = sa.plot_shift_distribution(
    originals=original_texts,
    rewrites_dict={
        "GPT-4": gpt4_rewrites,
        "Claude 3": claude_rewrites,
    },
)

Code Stylometry

Apply stylometry to source code. Measure developer fingerprints.

from stylometry.code import CodeAnalyzer

ca = CodeAnalyzer()

# Fit on known code samples
ca.fit(alice_code_files, label="Alice")
ca.fit(bob_code_files, label="Bob")

# Attribute an unknown file
predicted, distances = ca.predict(unknown_file)
print(f"Predicted author: {predicted}")

# Detect Copilot patterns
copilot_score = ca.copilot_score(code_file)
print(f"Copilot likelihood: {copilot_score:.2f}")

Code features measured:

Examples

See the examples/ directory:

• examples/rowling.py — Reproduce the Rowling identification experiment
• examples/llm_shift.py — Measure GPT-4 stylistic shift on your own texts
• examples/code_attribution.py — Attribute code files to developers
• examples/visualizations.py — Generate fingerprint and cluster plots as PNG files

cd examples
python rowling.py
# → Most likely author: Rowling (distance: 0.18)
# → Second closest:     Rendell (distance: 0.31)

python visualizations.py
# → saves examples/results/fingerprints.png
# → saves examples/results/clusters.png

Visualization Preview

Fingerprint comparison:

PCA clusters:

Limitations

Stylometry provides probabilistic signals, not forensic proof.

• Minimum ~100 words per text for reliable results
• Function word analysis is language-dependent
• Cross-domain generalization degrades significantly
• LLM detection is prompt-dependent and model-dependent

See LIMITATIONS.md for a full discussion.

References

• Mosteller & Wallace (1964). Inference and Disputed Authorship: The Federalist.
• Juola (2015). The Rowling Case. DSH, Oxford.
• Stamatatos (2009). A Survey of Modern Authorship Attribution Methods. JASIST.
• Kestemont et al. (2020). PAN @ CLEF 2020 Authorship Verification.
• Caliskan et al. (2015). De-anonymizing Programmers via Code Stylometry. USENIX.

License

MIT