📄 Early middle English manuscript as an object of linguistic analysis

Summary

This is a historical-linguistics study, not an NLP paper. It asks how the structure of English changed between late Old English and early Middle English, using the Proverbs of Alfred as evidence. The text survives in four 13th-century copies of a lost original: Trinity College (T), Jesus College (J), Maidstone (M) and Cotton Galba (C). Arngart edited all four. M and C are fragmentary or unreliable, so the author sets J and T side by side and compares them line by line. J keeps many Old English inflections: gendered and case-marked pronouns, dative endings, adjective–noun agreement and plural markers. T, copied in roughly the same period, drops endings, levels forms by analogy and moves toward analytic constructions. The author also uses alliteration rules to choose between proposed identifications of the opening place-name. The main conclusion is that the gap between the two copies is structural rather than chronological. On that view, the question of when Middle English began should be answered by system-level change, not by dates.

Key points

• The four witnesses differ in date, size and reliability:
  • T: second half of the 13th century, 37 sections, 668 lines. The scribe was probably Norman, used the continental letter w and confused dental spirants with stops.
  • J: copied shortly after 1276, 23 sections, 456 lines, with almost no errors.
  • M: first half of the 13th century, 17 sections, 266 lines, full of mistakes.
  • C: mostly burned in the 1731 Cottonian fire. About 100 lines survive, including sections 19, 20, 28 and 29.
• Opening place-name: Seuorde (J) is preferred over Siforde (T), linking the place to Seaford rather than Shefford. In Germanic alliterative verse, sc/sk alliterates only with sc/sk and never with plain s, and Seuorde alliterates on s.
• Case and agreement loss in T:
  • Line 8: J's feminine dative singular Þare becomes the unmarked Þe in T.
  • Line 43: J's accusative hine becomes the dative-derived him in T.
  • Line 28: the verb lusten requires a dative object, which J marks on louerde and T levels.
• Line 24: J places the preposition on after its noun. The T scribe apparently misread this and added an extra in before the noun.
• Line 35: T's sawen gives sagu a weak-declension plural that it did not have in Old English. The author reads this as analogical leveling.

Interesting ideas

• I1. Two copies made at about the same time can still preserve different grammatical stages, so aligning them line by line exposes exactly where structure changed.
• I2. An independent formal constraint, here the alliteration pattern, can settle which of two readings is genuine.
• I3. A copyist who did not understand a postposed preposition "repaired" the line by inserting a redundant one, so a reader's mistakes reveal which constructions it cannot handle.
• I4. Grammatical features can collapse (Þare→Þe, hine→him) or be filled in by analogy (sawen), so the surface form alone may stop carrying gender, number or case.
• I5. Differing ink shades suggest the J scribe left hard passages blank and returned to fill them in later.

For Mova

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

• I1 (learn): when witness parsers disagree, align their outputs token by token and classify each disagreement by construction type, instead of only counting agreement rates.
• I2 (learn): when silver-data candidates conflict, an orthogonal check such as agreement or valency can act as a tie-breaker that does not depend on the parsers themselves.
• I3 (tree): flag doubled or stranded adpositions in parses as a diagnostic for constructions the parser misanalyses, and turn these cases into repair rules.
• I4 (tokens): when the morphology layer must infer a feature from context rather than read it off the form, it should record that the value was inferred, which keeps explanations honest.
• I5 (learn): defer low-confidence items to a later pass, for example a higher LLM teacher tier, instead of forcing an early guess.