📄 Some Lexical Collocational Patterns in Late Middle English Legal Texts

Summary

The paper asks why certain word combinations became fixed in fifteenth-century English legal writing, and whether they stayed technical or passed into everyday language. Iglesias-Rábade treats collocations as products of culture and domain, not just statistical accidents. He compiles two late Middle English corpora by drawing on the University of Michigan Humanities Text Initiative's online collection of Middle English prose and verse texts. The first is a large non-technical reference corpus (n-TC) of fiction, drama and religious texts. The second is a smaller Legal Corpus (LC) split into seven subgenres. Using WordSmith Tools 4, he lemmatizes spelling variants, ranks LC words by keyness against n-TC, and filters candidate pairs with Church and Hanks's Mutual Information. The analysis covers six lexical patterns, from noun–noun compounds to binomials such as "execute and perform". Candidates are judged on three criteria: institutionalization, grammatical fixedness and non-compositionality. The surviving text indicates that the patterns studied were largely confined to legal English. The work shows how a contrastive-corpus method can separate domain-bound formulae from general usage in a historical language.

Key points

• Corpus sizes:
  • n-TC: 1,261,015 tokens and 67,155 types. It draws on 18 works, including The Canterbury Tales, Confessio Amantis, Merlin and the York and Towneley Plays.
  • LC: 408,398 tokens and 26,032 types.
• LC subgenres: Depositions, Lincoln Documents, Gylds, Indentures, Petitions, Signet and Privy Seals, and Wills. Gylds dominate with about 51% of the words.
• Six patterns studied:
  • N1 + N2 (market day)
  • N and N (wil and testament)
  • Adj + N (gracious lord)
  • Adj and Adj + N (trusti and wellbeloued lord)
  • V + N (take profit)
  • V and V (execute and perform)
• Salience differs from frequency:
  • day is the most frequent LC noun (801 occurrences) but ranks only 10th in keyness, because it is also common in n-TC (2,082).
  • yere ranks first in keyness (649 in LC vs 257 in n-TC).
• N1 + N2 analysis:
  • It was limited to the top 24 salient nouns within the first 100 ranks, and only 7 of them actually formed N1 + N2 patterns.
  • For chirche (546 occurrences, 64 as N2), only four left-collocates appear 4 or more times: parische (28 in LC vs 2 in n-TC), cathedral, moder and cryst.

Interesting ideas

• I1. Ranking words by keyness against a general reference corpus, rather than by raw frequency, surfaces domain vocabulary that frequency alone hides (yere versus day).
• I2. Choosing collocation candidates in stages works as a cheap funnel: first keep only salient head words, then search their left-collocates, then filter by an association score.
• I3. Merging spelling and morphological variants under one lemma before counting is what makes association statistics meaningful for a non-standardized language.
• I4. Accepting a combination as a collocation only when it passes three tests (institutionalization, fixedness under grammatical restrictions, non-compositionality) turns a fuzzy notion into a checklist.
• I5. Binomials such as "goods and chattels" or "execute and perform" are treated as one conventional unit rather than two coordinated concepts.

For Mova

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

• I1 (world): Keyness against a general-English baseline could flag domain-specific nouns in a text or genre, marking which entities and concepts deserve dedicated ontology entries.
• I2 (tokens): The same funnel could mine candidate multi-word units, such as legal binomials or N+N compounds, for the tokenizer and lemmatizer to treat as single lexical items.
• I3 (learn): Silver-data statistics over noisy or historical text should be computed on lemmas, so that spelling variation does not split the counts and hide real regularities.
• I4 (frames): Multiword predicates like "take notice" could be tested with these criteria before being mapped to a single frame instead of being composed from their parts.
• I5 (tree): The parser or a post-pass could tag such fixed coordinations, so that later stages do not invent two separate entities or events from one legal formula.