Evolution of the Poetic Canon in the Princeton Prosody Archive

Method

From name strings to poets

Named-entity recognition only says “Shelley is a person.” Entity linking says which Shelley — Mary or Percy Bysshe — and pulls “Mrs. Shelley,” “Mary Wollstonecraft Shelley,” and the bare surname onto one identity across two centuries.

• the linker ReFinED, zero-shot. A neural model that detects and disambiguates each mention jointly from context, mapping it to a Wikidata QID (Dante → Q1067) with a 0–1 confidence. No corpus-specific training; ~1 day on an A100 over 1.68M pages.
• the Wikidata trap Linking is only as clean as the knowledge base. Wikidata files Friedrich Fröbel (the kindergarten founder) and Elizabeth Barrett Browning (d. 1861) among “soccer players” and “screenwriters” — so the poet set must be filtered, not taken raw.
• how accurate? ~98% at the 0.90 threshold. On 50 hand-annotated pages (3 annotators), 56 of 57 high-confidence links were correct. The system errs toward silence: its misses are hard OCR forms (“Bobert Bums” → Burns, “THEOCRITUS”), not fabricated poets.
• the cascade From spans to a poet universe. 23.9M mention spans → 11.6M linked → 92,823 humans → poet-flagged (Q49757) → conf. ≥0.90 → seen in ≥2 works → born ≤1900. Noise lives in singletons — the stability gate drops 34% of names but only 0.4% of mentions.

Final poet universe: 3,164 poets · 737,194 mentions across 3,149 works, 1559–1929

Method

Finding quotations at scale

A quotation is an asymmetric match: we align the PPA, where quotations surface, against a reference body of poetry — the Chadwyck-Healey English Poetry Database, deduplicated to 246,724 reference poems, plus Shakespeare’s plays, which Chadwyck-Healey leaves out.

• the tool passim. Built for the Viral Texts project to find reprinted passages across noisy newspaper OCR. It needs no prior guess about which texts align, and is robust to OCR error and ragged passage boundaries.
• tuned for verse Short quotes, not long reprints. passim’s defaults assume long documentary reuse, so we lowered the bar — n-gram 25→15, min length 50→25, gap 600→300, max DF 100→10,000 — with final values chosen against a hand-annotated gold set.
• the unit Poem appearance, not raw record. OCR noise and paraphrase split one visible quotation into several alignment records, so we count a poem’s presence in a host work once — the level every question here operates at.
• then filter Matches aren’t yet a corpus. Host-side (Literary/Linguistic collections, collapse reprints, 1694–1920) and poem-side (hand-remove centos & adaptations, a biographical-plausibility cut) filtering follow — the counts are the analytic-corpus detour one slide up.

1,263,587 raw alignments → 509,030 appearances → 252,893 appearances of 37,406 poems across 3,962 host works

Method

Holding the archive constant

A raw count of distinct poems can’t separate a widening practice from a thickening archive — more host pages simply mean more chances to quote. So we model a rate, with the archive’s size netted out.

• a rate, not a count The log‑pages offset. Page volume enters with its coefficient fixed at 1 — double the host pages, double the poems, by construction. The model is then free to find the departures from that proportionality.
• a flexible trend A natural cubic spline s(y). Free to bend where the evidence calls for it; 7 degrees of freedom, chosen by AIC. “Natural” keeps the curve from flailing at the window’s edges.
• room to wobble Negative Binomial, not Poisson. Yearly counts vary far more than their mean — one dense anthology spikes a year — so we relax the equal‑variance assumption. α_disp = 0.5; post‑fit dispersion ≈ 1.01.

Modelled window 1694–1920  ·  1694 = first year with ≥5 PPA host works  ·  1920 = the PPA’s coverage horizon

Method

Reading each poem against the corpus

• the input Prevalence per decade. For each eligible poem, p̂ = k_it/n_t — the share of that decade’s PPA host works that quote it. Fitted with a beta‑binomial rather than a binomial, because the superstar poems are overdispersed.
• the headline number The slope b¹ᵢ. Gaining or losing ground relative to the shared trend — not in absolute terms. A positive slope outpaces the average poem; a negative one lags even as the whole practice widens around it.
• reading the slope Log‑odds per decade. +0.25 ≈ ×1.28 odds (~28% likelier to be quoted each decade); −0.20 ≈ ×0.82 (~18% less). “Credibly” rising/declining = the 95% credible interval on b¹ excludes zero.
• keeping it honest Partial pooling + author shift. Thin‑evidence poems are pulled toward the corpus trend; well‑attested ones pull it. u_a(i) separates a poem’s own movement from its author’s standing (Shakespeare: 37 works; Gray rests almost entirely on the Elegy).

Eligibility starts 18 yrs after the poet’s birth — a Tennyson zero in the 1700s is impossibility, not neglect  ·  population = ≥30 host works across ≥10 decades → 896 poems  ·  Bambi / PyMC (NUTS), R̂ ≤ 1.02

Method

From poems to periods

No new model — the same per‑poem slopes b¹ᵢ, re‑cut by the literary period each poem belongs to. Labels are Chadwyck‑Healey’s bibliographic categories (the database’s organizing logic).

• why a period can surprise A period is a different unit from poems or authors. It can rise even when most of its authors are flat (a few rise sharply), or fall even when most of its poems hold (the high‑volume ones decline). The ridge checks whether the stories aggregate.
• what’s included Only periods with ≥20 modelled poems. The catch‑all “Poems & Miscellanies, 1500–1900” — for poems that fit no named period — is excluded.
• inside the Restoration ridge It’s bimodal: the classical‑translation and didactic‑verse apparatus fell hard (≈ −0.15), while Dryden’s original verse declined only mildly (≈ −0.05, intervals straddling zero).
• inside Jacobean / Caroline Layered: short lyric and song rose, the Shakespeare plays and Milton’s longer poems held, the heroic‑narrative tradition gave way — lyricization in miniature.

Slopes in log‑odds per decade vs the shared trend  ·  Middle English sits furthest right — Chaucer and the founders the nineteenth century canonized