📄 Relative contributions of Shakespeare and Fletcher in Henry VIII: An Analysis Based on Most Frequent Words and Most Frequent Rhythmic Patterns

Summary

The play Henry VIII is widely believed to be co-written. This study tests which candidates contributed (Shakespeare, Fletcher, Massinger) and in what proportions. It joins lexical and metrical evidence: each verse line is encoded as a binary stress pattern ("rhythmic type") using the Prosodic library on EarlyPrint XML texts. A linear SVM (scikit-learn) is trained on the 500 most frequent rhythmic types plus the 500 most frequent words, drawn from scenes of plays by the three authors written around the same period.

The study works at two levels. First, whole scenes are classified. Second, Eder's rolling attribution slides an overlapping fixed-size window of lines across the text and turns classifier outputs into probabilities with Platt scaling. This lets the method detect authorship changes inside a scene.

Massinger turns out to be very unlikely as a contributor. The Shakespeare/Fletcher split largely confirms Spedding's 1850 division, with some boundary shifts that support Thomas Merriam's later revisions. The work shows that line-level prosody is a strong stylometric signal, and that window-based attribution can locate collaboration seams that do not coincide with scene boundaries.

Key points

• Training data: 4 Shakespeare plays (53 scenes), 4 Fletcher plays (90 scenes) and 3 Massinger plays (46 scenes).
• Validation protocol:
  • Each play is held out in turn, so no test scene shares a play with the training data.
  • Classes are balanced by random downsampling.
  • The whole procedure is repeated 30 times.
• Scene-level accuracy:
  • Rhythm alone scores 0.97–1 for Shakespeare and Fletcher, except Valentinian at 0.84. It scores 0.81–0.88 for Massinger.
  • Words alone score 0.95–1.
  • The combined 1000-dimensional model wins in three cases.
• Rolling attribution settings and validation: windows of k = 100 lines with step d = 5. On eight known plays, only 10 of 4412 five-line groups were misassigned, an estimated accuracy of 0.9977.
• Findings for Henry VIII:
  • Massinger receives only minority votes, in scenes 2.1 and 4.2.
  • The shift in scene 3.2 is placed around line 2200 rather than Spedding's 2081, matching Merriam.
  • Both scenes of act 4 come out ambiguous.

Interesting ideas

• I1. Holding out an entire play rather than random scenes prevents the model from learning work-specific quirks that inflate scores.
• I2. Classifying overlapping fixed-size windows and averaging calibrated probabilities per small chunk turns a document classifier into a change-point locator.
• I3. A stress pattern per line, written as a bit string, is a compact, interpretable prosodic signature built on top of morphology.
• I4. Repeating balanced random downsampling many times and reporting vote counts (e.g. 20 of 30) yields an honest, readable confidence measure.
• I5. Running each feature family alone and then combined shows where each signal disagrees, which is itself evidence (e.g. rhythm vs words in act 4).

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Evaluate parsers and taggers with whole documents or sources held out, so silver-data gains are not mere leakage from shared texts.
• I2 (frames): A similar sliding window over speaker or narrator style cues could flag where voice or speaker changes in dialogues and tales, helping attribute speech acts.
• I3 (tokens): The en crate could expose per-word lexical stress and per-sentence stress strings, giving explainable features for poetry questions in school-course tests.
• I4 (learn): Report witness-parser agreement as vote tallies over resampled runs, so silver labels carry a visible confidence rather than a single verdict.
• I5 (learn): Routinely ablate feature groups in the parser and treat divergence between them as a pointer to sentences needing repair rules or teacher review.