Explorer

The Honour Test

Same English word. Different worlds.

What this view is and isn’t. The project is not claiming Shakespeare owned the word honour: both corpora use it constantly. Founders use it about 18,600 times; Shakespeare uses it 726 times. The finding is about the collocates — the words that habitually appear next to honour in each corpus. Those collocational profiles diverge statistically across fourteen common abstract nouns; the G log-likelihood test detects the divergence cleanly even after Bonferroni correction. The Cramer’s V (phi) effect sizes that quantify the difference, though, register as “very weak” — or, at the very top, barely “weak” — under Stefanowitsch’s verbal scale (topping out at phi ≈ 0.10) on every target. The shift is real and statistically robust; the absolute magnitude is thin. Each collocate in the columns below is shown with its G value, its phi, and the verbal effect-size label so you can read both numbers at once.

In Shakespeare, honour’s closest collocates are pawn, mine, lord, Brutus: characters “pawn their honour” as security, swear “by mine honour,” address one another as “your honour, my lord.” In the Founders’ letters, the collocates are sir, esteem, letter, excellency: the language of letter-closing protocol — “I have the honour to be, Sir, your most obedient humble servant.” Same noun, two different conceptual worlds.

A legitimate caveat: the contrast is between two genres (Shakespearean drama and eighteenth-century correspondence) at least as much as between two thinkers’ concepts. A merchant’s letters of the same period would probably show the same letter-closing pattern. So the divergence is a genre signal at least as much as a conceptual-inheritance signal. The full caveat is in the Stylistic notes section.

The same word; two totally different things. And it isn’t just honour. The same kind of divergence shows up for thirteen other common words. Pick one below to see for yourself. The words that hang around the target in the Founders’ writing appear on the left; the words that hang around it in Shakespeare’s appear on the right.

Pick a word

Selected: honour · 40 statistically distinctive collocates across the two sides

The Founders' honour

Epistolary — what one has when corresponding with a Sir

  1. 1.sirG 218φ 0.026very weak
  2. 2.respectG 132φ 0.020very weak
  3. 3.greatG 109φ 0.019very weak
  4. 4.esteemG 90φ 0.015very weak
  5. 5.letterG 69φ 0.013very weak
  6. 6.excellencyG 67φ 0.014very weak
  7. 7.incloseG 41φ 0.010very weak
  8. 8.sentimentsG 38φ 0.010very weak
  9. 9.yrG 38φ 0.010very weak
  10. 10.greatestG 33φ 0.010very weak
  11. 11.writeG 31φ 0.009very weak
  12. 12.writingG 29φ 0.009very weak
  13. 13.presidentG 26φ 0.008very weak
  14. 14.mrG 23φ 0.008very weak
  15. 15.transmitG 22φ 0.008very weak
  16. 16.highestG 22φ 0.008very weak
  17. 17.respectfullyG 20φ 0.007very weak
  18. 18.gentlemenG 19φ 0.008very weak
  19. 19.informG 19φ 0.007very weak
  20. 20.congressG 19φ 0.007very weak

Bars are scaled to G log-likelihood; numbers to the right report G alongside phi (Cramer’s V effect size, 0–1 scale). All items pass Bonferroni correction, but the phi column shows how small the underlying effect actually is: across the project’s fourteen target nouns, the top collocates register as very weak — or, in a single case, barely weak — on Stefanowitsch’s verbal scale (topping out at φ ≈ 0.10). The G test confirms that there is a collocational difference; phi tells you it is a thin one.

Shakespeare's honour

Chivalric — pawnable, stainable, lord-bound

  1. 1.mineG 511φ 0.101weak
  2. 2.loveG 119φ 0.044very weak
  3. 3.lordG 114φ 0.041very weak
  4. 4.honourG 108φ 0.044very weak
  5. 5.pawnG 79φ 0.040very weak
  6. 6.comeG 66φ 0.032very weak
  7. 7.likeG 54φ 0.029very weak
  8. 8.brutusG 52φ 0.033very weak
  9. 9.tisG 49φ 0.030very weak
  10. 10.kingG 46φ 0.026very weak
  11. 11.beseechG 46φ 0.030very weak
  12. 12.heavenG 46φ 0.030very weak
  13. 13.exitG 46φ 0.030very weak
  14. 14.fatherG 42φ 0.025very weak
  15. 15.lifeG 41φ 0.023very weak
  16. 16.pleaseG 41φ 0.022very weak
  17. 17.coriolanusG 39φ 0.028very weak
  18. 18.renownG 39φ 0.028very weak
  19. 19.welcomeG 39φ 0.028very weak
  20. 20.uponG 38φ 0.018very weak

Bars are scaled to G log-likelihood; numbers to the right report G alongside phi (Cramer’s V effect size, 0–1 scale). All items pass Bonferroni correction, but the phi column shows how small the underlying effect actually is: across the project’s fourteen target nouns, the top collocates register as very weak — or, in a single case, barely weak — on Stefanowitsch’s verbal scale (topping out at φ ≈ 0.10). The G test confirms that there is a collocational difference; phi tells you it is a thin one.

Reading

What this contrast shows

The diagnostic case for the whole conceptual-inheritance argument. Shakespeare's honour can be pawned (Coriolanus 5.6: 'I pawn'd mine honour for his truth'); it lives in the company of mine, lord, love, come, kept. The Founders' honour lives in epistolary openings and closings — sir, respect, esteem, excellency, letter ('We have the honour to be, Sir, your most obedient humble Servants'). The word travelled across the Atlantic; the chivalric concept did not. The Founders inherited the lexis and rebuilt the meaning.

What to read next

The argument behind the pattern

The contrast you can see by clicking through the targets above is the project’s most vivid stylistic contrast — a supporting beat to the central Two Modes finding. The full argument (what it means that shared vocabulary carries divergent collocational worlds) is in the Honour Test case study. For the companion finding (Adams cites Shakespeare explicitly; Franklin sounds like Shakespeare without citing him), see Two Modes of Shakespearean Influence. Per-passage deep dives live in the case studies.

▸How we picked these words, for the methodologically curious

For each target noun, we count every word that occurs within five words of the target in either corpus. Each candidate is then scored with a standard word-distinctiveness test (G log-likelihood, the workhorse statistic for keyword and collocate comparison in corpus linguistics) against its frequency in the other corpus.

The lists above show the top words on each side after filtering out function words, proper names, and tokenisation fragments, and after applying a multiple-comparisons correction (Bonferroni) so that what survives is what survives at conventional significance levels. Bars are scaled to the maximum G-score within each side; the number beside each bar is the raw G-score.

The full per-target tables, including everything we filter out here, are in tables/cs3_*.csv in the research repository, or read the section on differential collocate analysis in the full paper.