Stylometry
Using measurable features of writing — function-word frequencies, sentence lengths, punctuation — to profile authorship.
Last updated
Stylometric analysis uses measurable features of writing — function- word frequencies, sentence lengths, punctuation patterns — to profile authorship. Because function words (the, of, but, if) are used unconsciously, they tend to fingerprint an author more reliably than content words.
Why it matters for your research. In corpora with anonymous, pseudonymous, or disputed authorship — political pamphlets, newspaper columns, epistolary collections — stylometry can argue for or against attributions. It rarely proves anything on its own, but it strengthens or weakens cases built on other evidence.
In Archēglyph. On the roadmap as an authorship- attribution plugin. Only useful when the corpus has enough per- candidate text to make the feature distributions stable.
Not to be confused with. Topic modelling, which surfaces what documents are about; stylometry surfaces how they’re written.
A worked example. Burrows’s Delta method compares the frequency of the ~100-150 most common words in a corpus — mostly function words like “the,” “of,” “which,” “not” — between a disputed text and a set of candidate authors’ known writing. The candidate whose frequency profile has the smallest weighted distance to the disputed text is the strongest match. The method works precisely because authors rarely think consciously about how often they write “of” versus “the” — the signal tends to survive even deliberate attempts to disguise content and vocabulary, which is why disputed-authorship studies of anonymous or pseudonymous periodical writing lean on function words rather than subject matter.
Where it breaks down. Delta and its relatives need enough text per candidate to make a frequency distribution stable — a few hundred words of disputed text against a one-paragraph writing sample from a candidate author will not produce a reliable result, no matter how distinctive the candidates’ styles seem on a close read. Co-authorship, heavy editorial revision, and translation all blur the function-word signal too, which is why stylometric findings are reported as evidence that strengthens or weakens a case, never as a standalone proof of authorship.
What counts as a “feature” beyond word frequency. Delta and its relatives are the best-known methods, but stylometry more broadly covers any measurable regularity: average sentence length, punctuation density, ratio of function words to content words, even letter or n-gram frequencies for corpora too fragmentary for reliable word counts. Different features suit different questions — sentence-length distributions can distinguish an author’s early and late style even when vocabulary hasn’t changed, which plain word-frequency comparison would miss entirely.
Relationship to other tools in this glossary. Stylometry usually runs on the same chunked, sentence-segmented text that feeds embedding and search, but it measures a different axis of the text — form rather than content — which is why it sits beside topic modelling in the roadmap rather than replacing it: a corpus can be profiled for what it discusses and who wrote it at the same time, from the same underlying chunks.
Reporting a result honestly. A responsible stylometric write-up states which feature set and distance measure were used, how many candidate authors were compared, and how much text each candidate contributed — a Delta score computed against five candidates with thousands of words each carries a different weight than the same score against two candidates with a few hundred words apiece. Treating the numeric output as a verdict rather than as one input to a larger argument is the most common way stylometric findings get overstated in secondary reporting.
Interaction with OCR and translation. Stylometric features assume the text reflects the author’s actual word and letter choices, so OCR noise (a misread letter changing an n-gram) and machine translation (which reflects the translator’s style, not the original author’s) both corrupt the signal before the analysis even starts. Running stylometry on a machine-translated text answers a different question — “does this translation’s phrasing resemble that one” — than running it on the original-language source.