Does “late style” exist?

Jonathan Reeve

Does “late style” exist?

Jonathan Reeve
Literary Modeling and Visualization Lab
Group for Experimental Methods in the Humanities
Columbia University

The Problems / Hypotheses

Edward Said: On Late Style

Cover of On Late Style
Cover of On Late Style


“The maturity of the late works of significant artists does not resemble the kind one finds in fruit. They are, for the most part, not round, but furrowed, even ravaged. Devoid of sweetness, bitter and spiny, they do not surrender themselves to mere delectation.” (Adorno, quoted in Said) - Beethoven’s Ninth Symphony

An Old Critical Concern

  • Shakespeare studies of the 18th C
  • Lipking, The Life of the Poet (1984)
  • Adorno on Beethoven
  • Recently: “stylochronometry” papers

Difficulties / Caveats

Ambiguities of “late”

  • Untimely? Ahead/behind one’s time?
  • Dead? “The late Mr. Darcy”
  • Running late

Ambiguities of “style”

  • Contradictory definitions of “late style”

Said: “[late style] has the power to render disenchantment and pleasure without resolving the contradiction between them. What hold them in tension, as equal forces straining in opposite directions, is the artist’s mature subjectivity, stripped of hubris and pomposity, unashamed either of its fallibility or of the modest assurance it has gained as a result of age and exile”

  • Style as content?

Potential causes

  • Proximity to death?
  • Foreknowledge of one’s death?
  • Old age? (Alterstil v. Spätstil)
  • Old critics? (Said)



  • Quantifies textual stylistic differences
  • Well-studied in authorship attribution

Document Embeddings

  • Averaged word embeddings
  • Encode semantic information about documents
  • Test “style as content” hypothesis

Corpus Creation

Corpus A

  • Hand-curated (sometimes with OCR)
  • Writers discussed by Said: Mann, Proust, Genet
  • Writers with well-known “late periods”: James, Dickens
  • All electronically-available texts of each (virtually all texts)

Corpus B

  • Machine-generated using
  • All Project Gutenberg works in English,
    • with the “PR” Library of Congress category (British Lit.)
    • by authors with more than 8 works
  • 141 writers
    • Pruned according to available publication metadata to 51 writers
    • ~900 total works



  • Documents vectorized to document-term matrices with 800 MFW
  • Pre-clustered according to:
    • Date
    • Years to Death
  • Documents randomly sampled to make same length

Dimensionality Reduction

  • TF matrices reduced to 5 dimensions
    • using Principal Component Analysis (PCA)
    • 5 Dimensions perform better than 2
      • (acc. to previous grid search study)
    • Plots are projections of first 2


  • Periods modeled using a Bayesian Gaussian Mixture model
    • 1-3 possible “periods” or clusters
    • Clustering happens in 5D
  • 20 trials, averaged


  • “Lateness”
    • Of a work: L2 norm of the 5D document vector
    • Of a writer:
      • Difference of Euclidean L2 norms of centroids of doc. vectors
        • PCA-reduced MFW vectors
        • or semantic document embedding vectors
      • as pre-clustered by date or years to death to 1-3 clusters
  • “Periodicity”
    • Adjusted Rand Index comparing BGM clusters with initial date-based 1D-clustering
      • Category-agnostic mutual information score


Corpus A

Jean Genet
Jean Genet

Marcel Proust

Marcel Proust
Marcel Proust

Thomas Mann

Thomas Mann
Thomas Mann

Mary Augusta Ward

Mary Augusta Ward (Mrs. Humpry Ward)
Mary Augusta Ward (Mrs. Humpry Ward)

Henry James

Henry James
Henry James

Charles Dickens

Charles Dickens
Charles Dickens

Willa Cather

Willa Cather
Willa Cather


Author Periodicity Score
James 0.472
Dickens 0.469
Genet 0.457
Mann 0.367
Conrad 0.177
Cather 0.177
Ward 0.166
Proust 0.023

Results: Corpus B

Mean Latenesses, Sorted
Mean Latenesses, Sorted

Mean Lateness, All

(Early style, not late style.)

Mean Periodicities

Mean Periodicies, Sorted
Mean Periodicies, Sorted

Mean Latenesses (TTD)

Mean Latenesses by Time to Death
Mean Latenesses by Time to Death

Doc Embeddings: Mean Latenesses

Mean Latenesses using Document Embeddings
Mean Latenesses using Document Embeddings

Doc Embeddings: Mean Periodicities

Mean Periodicities using Document Embeddings
Mean Periodicities using Document Embeddings

Doc Embeddings: Overall Mean

Mean: -0.0288
(Even stronger early style.)


  • Late Style Early Style
    • Even stronger when style=content
    • Late style is weakly present when calculated with TDD
      • Needs more study

Future Plans

  • Bigger corpora
  • Better stylometry (cosine delta?)
  • Better statistics (confidence intervals, hypothesis testing)
  • More precise engagement with criticism


  • Completely free and open-source
  • Help needed
  • Funding?