Elamur Research⌕
☰
home / papers / 2017 / Automatic Compositor Attribution in the First Folio of Shakespeare

Automatic Compositor Attribution in the First Folio of Shakespeare

Paper · 2017 · Maria Ryskina, Hannah Alpert-Abrams, Dan Garrette et al.

Summary

The paper presents the first automatic method for compositor attribution. This is the task of grouping the pages of an early printed book by the worker who set the type, without knowing in advance who those workers were. Bibliographers have long done this by hand, tallying spelling habits and checking spacing around punctuation. The authors build an unsupervised generative model with a latent compositor for each page. Given the aligned modern word, the compositor produces the historical (diplomatic) spelling through a log-linear distribution. That distribution uses word-specific features and Levenshtein edit-operation features. The compositor also produces the pixel width of spaces after medial commas, drawn from a multinomial. Parameters are fitted with EM. The model is tested on Shakespeare's First Folio using both a manual transcription and Ocular OCR output. Its page clusters match expert attributions about 87% of the time, and accuracy drops only slightly on OCR text. This suggests the approach could scale to books that have no careful transcription.

Key points

  • Data:
    • Bodleian diplomatic transcription of the First Folio (1623).
    • Ocular OCR run on the Bodleian facsimile images, with character error rate around 10–15% for most plays.
    • Modern text from the MIT Complete Works of Shakespeare, aligned to the diplomatic text by word-level edit distance.
  • Gold attributions:
    • Blayney (1996), with 8 compositors (C = 8).
    • An approximation of Hinman (1963), with 5 compositors (C = 5).
  • Metrics are one-to-one and many-to-one cluster accuracy. Random baselines score 22.5/49.6 against Hinman and 16.7/49.6 against Blayney.
  • The best variant, FEAT w/ EDIT + SPACE, scores:
    • about 87.6 (Hinman) and 87.3 (Blayney) on the manual transcription;
    • 86.7 and 85.9 on OCR.
    • Adding WORD features slightly lowers these results, more noticeably on OCR.
  • Training uses 75 EM iterations with 100 random restarts, keeping the run with the highest likelihood. The M-step for the feature weights uses gradient ascent.

Interesting ideas

  • I1. The model explains a variant spelling as a function of its canonical modern word, rather than modelling which words appear at all. This isolates the transcriber's habits from the author's word choice.
  • I2. Whole-word preference features and generic character-edit features (with optional neighbouring-character context) are included side by side. Training then decides which kind of evidence dominates.
  • I3. Learned parameters were checked against independent expert observations. For example, the learned u→DEL weight matched one compositor's known habit of spelling "yong".
  • I4. Unsupervised clusters were scored with both one-to-one and many-to-one mappings to gold labels, alongside a random baseline.
  • I5. A non-parametric multinomial over discretized pixel widths captures multi-peaked distributions while keeping EM updates trivial.

For Mova

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

  • I1 (tokens): Archaic or variant spellings in older fables and tales could be mapped to standard lemmas with edit-operation features conditioned on the modern form, making each normalization step traceable.
  • I2 (tokens): A lemmatizer or spelling normalizer could combine per-word exceptions with general suffix/edit rules and let data weigh them, instead of hand-picking one approach.
  • I3 (learn): Self-induced repair rules or learned weights could be sanity-checked against known linguistic facts, as an interpretability-based evaluation alongside accuracy.
  • I4 (learn): This evaluation scheme suits any induced grouping, such as automatically discovered verb classes or entity clusters, where the cluster labels have no fixed identity.
  • I5 (world): Discretized multinomials could model stored distributions of measures or values in compiled knowledge where a single Gaussian would hide several typical modes.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🏢GoogleMountain View · United States430 works · 424,363 citations🎓Carnegie Mellon UniversityPittsburgh · United States539 works · 179,533 citations🎓The University of Texas at AustinAustin · United States109 works · 56,942 citations

 Fields, topics and concepts

📚 Shelves 6The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksLiterature and Literary Theory · 519 works

Cite

Cite as Maria Ryskina, Hannah Alpert-Abrams, Dan Garrette and Taylor Berg-Kirkpatrick. 2017. Automatic Compositor Attribution in the First Folio of Shakespeare. Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/p17-2065

@inproceedings{ryskina2017automatic,
  title = {{Automatic Compositor Attribution in the First Folio of Shakespeare}},
  author = {Maria Ryskina, Hannah Alpert-Abrams, Dan Garrette and Taylor Berg-Kirkpatrick},
  booktitle = {Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers)},
  year = {2017},
  doi = {10.18653/v1/p17-2065},
  url = {https://doi.org/10.18653/v1/p17-2065}
}

References

22 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Moshe Koppel, Jonathan Schler, and Shlomo Argamon. 2009. Computational methods in authorship attribution. J. Am. Soc. Inf. Sci. Technol. 60(1):9– 26. https://doi.org/10.1002/asi.v60:1. ↗
  2. Taylor Berg-Kirkpatrick, Alexandre Bouchard-Côté, John DeNero, and Dan Klein. 2010. Painless unsupervised learning with features. In Proceedings of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies.
  3. D. F. McKenzie. 1969. Printers of the mind: Some notes on bibliographical theories and printing-house practices. Studies in Bibliography 22:1–75. find
  4. Taylor Berg-Kirkpatrick, Greg Durrett, and Dan Klein. 2013. Unsupervised transcription of historical documents. In Proceedings of the Association for Computational Linguistics. find
  5. D. F. McKenzie. 1984. Stretching a point: Or, the case of the spaced-out comps. Studies in Bibliography 37:106–121. find
  6. John O’Connor. 1975. Compositors D and F of the Shakespeare First Folio. Studies in Bibliography 28:81–117. find
  7. Peter W. M. Blayney. 1991. The First Folio of Shakespeare: In Conjunction with the Exhibition at the Folger Shakespeare Library, Washington, DC, April 1, 1991-September 21, 1991. Folger Shakespeare Library. find
  8. Pervez Rizvi. 2016. Use of spellings for compositor attribution in the First Folio. The Papers of the Bibliographical Society of America 110:1–53. find
  9. Peter W. M. Blayney, editor. 1996. The First Folio of Shakespeare: The Norton Facsimile. Norton. find
  10. Gary Taylor. 1981. The shrinking compositor A of the Shakespeare First Folio. Studies in Bibliography 34:96–117. find
  11. Ian R. Burrows. 2013. ”The peryod of my blisse”: Commas, ends and utterance in Solyman and Perseda. Textual Cultures: Texts, Contexts, Interpretation 8(2):95–120. find
  12. Paul Werstine. 1982. Cases and compositors in the Shakespeare First Folio Comedies. Studies in Bibliography 35:206–234. find
  13. Christos Christodoulopoulos, Sharon Goldwater, and Mark Steedman. 2010. Two decades of unsupervised POS induction: How far have we come? In Proceedings of the Conference on Empirical Methods in Natural Language Processing.
  14. Philip Gaskell. 2007. A New Introduction to Bibliography. Oak Knoll Press. find
  15. Charlton Hinman. 1963. The printing and proofreading of the First Folio of Shakespeare, volume 1. Oxford: Clarendon Press. find
  16. David I. Holmes. 1994. Authorship attribution. Computers and the Humanities 28(2):87–106. find
  17. Jonathan Hope. 1994. The authorship of Shakespeare’s plays: A socio-linguistic study. Cambridge University Press. ↗
  18. Trevor H. Howard-Hill. 1973. The compositors of Shakespeare’s Folio Comedies. Studies in bibliography 26:61–106. find
  19. Trevor H. Howard-Hill. 1976. Compositors B and E in the Shakespeare First Folio and Some Recent Studies. Self-published. find
  20. Trevor H. Howard-Hill. 1980. New light on compositor E of the Shakespeare First Folio. The Library 6(2):156–178. ↗
  21. Matthew L. Jockers and Daniela M. Witten. 2010. A comparative study of machine learning methods for authorship attribution. Literary and Linguistic Computing 25(2):215–223. ↗
  22. Patrick Juola. 2006. Authorship attribution. Foundations and Trends in Information Retrieval 1(3):233– 334. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/ryskina-2017-automatic-compositor-attribution-first-folio/