📄 A Simple Text Analytics Model To Assist Literary Criticism: comparative approach and example on James Joyce against Shakespeare and the Bible

Summary

The paper proposes a deliberately minimal text-analytics workflow to support literary criticism. It is designed so that programmers and humanities scholars can both follow and extend it. The idea is to place a target author's books in a shared measurement space alongside reference literature. Here the references are Shakespeare's complete works (Project Gutenberg) and the King James Bible, divided into its 80 individual books. Each book is reduced to about 20 simple statistics:

• mean sizes and counts of sentences, tokens, stopwords, punctuation and known words;
• the depth of WordNet synsets for known words.

These statistics are z-scored, and all books are projected together with PCA. As a validation case the authors analyse five works by James Joyce. Ulysses and Finnegans Wake form a cluster far from everything else. Dubliners and A Portrait sit close to Bible books, and Stephen Hero falls between the two groups. The authors read this as quantitative support for the view that Joyce's style leaves meaning gaps for the reader to fill. The method is fast, fully scripted and public, and the reference set is easy to swap.

Key points

• Reference corpus:
  • Shakespeare: 36 plays plus 2 poetry batches.
  • King James Bible: 39 Old Testament, 14 Apocrypha and 27 New Testament books.
  • Joyce: Stephen Hero, Dubliners, Portrait, Ulysses and Finnegans Wake.
• Toolchain and code: Python, NLTK, Numpy and Matplotlib; scripts are in a public git repository. The whole run takes a few minutes.
• PCA components: the first three explain about 40.25%, 14.32% and 11.91% of the variance. Three components were kept instead of the usual two.
• POS tagging deliberately avoided: the authors argue that modernist authors bend vocabulary and syntax, which makes taggers unreliable on such texts.
• Surprising results:
  • Ulysses and Finnegans Wake have deeper (more specific) synsets, landing near Shakespeare. This contradicts the authors' abstraction hypothesis.
  • Dubliners has more unknown words than Ulysses.
  • Known-word fractions separate Joyce's works better than the principal components do.
  • Shakespeare's books occupy a narrower region than the Bible's, and the two collections are cleanly separable on PC1–PC2.

Interesting ideas

• I1. Judge a text by placing it in a fixed reference cloud of well-understood corpora rather than in isolation, so distances become interpretable.
• I2. Use the share of tokens found in a lexicon, computed both over all tokens and over content words only, as a cheap signal of how unconventional a text's vocabulary is.
• I3. Treat WordNet hypernym depth as a measure of how specific versus generic the vocabulary is, giving each text a concreteness profile.
• I4. Write down expectations before running the numbers, then openly report where results contradict them (such as the synset-depth surprise).
• I5. Choose transparent features and z-score them before PCA, so that no single measure dominates and every axis stays explainable to non-specialists.

For Mova

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

• I1 (learn): Keep a frozen set of reference texts and project new silver-data batches onto the same feature space to spot batches that drift away from known-good material.
• I2 (tokens): Compute this rate per sentence and flag low-coverage input as a warning that tagger and lemma output may be unreliable, before any parsing happens.
• I3 (world): Store synset depth on entities in the world graph so readers can tell vague mentions ("thing", "creature") from specific ones when resolving references or ranking answers.
• I4 (learn): Record predicted outcomes before each evaluation run, so unexpected results surface as explicit findings alongside the negative controls instead of being rationalised afterwards.
• I5 (answer): Use z-scored, named diagnostic features when comparing reader performance across test sets, so differences remain explainable rather than buried in opaque scores.