📄 Editing Middle English Medical Manuscripts : The Case of Glasgow University Library MS Hunter 509

Summary

The article walks through the editorial workflow for a single Middle English medical manuscript, Glasgow University Library MS Hunter 509 (H509). The manuscript holds a System of Physic, mostly the Middle English Gilbertus Anglicus, a prose adaptation (ca. 1400) of Gilbert the Englishman's thirteenth-century Compendium medicinae. The main text was copied around 1460.

The author first describes transcription, which turned out to be far from mechanical. Problems included:

• letter shapes that look alike;
• one graph serving for both þ and y;
• abbreviation signs that can stand for several letter groups;
• unseparated minims;
• dense punctuation that is hard to tell apart from stains.

For each problem she gives a practical way to resolve it. Lemmatisation, morphological tagging and concordance generation followed. The text was then published in two forms:

• a semi-diplomatic electronic edition with images side by side, part of a roughly 1.2-million-token corpus of medieval scientific prose;
• a printed edition that mixes diplomatic and critical methods.

A glossary keyed to MED headwords and a full list of lemmas complete the edition. The work matters because it makes an unedited scientific text usable for linguistic research.

Key points

• H509 has more than 170 folios (over 350 pages). Folios 14r–167v contain the Middle English Gilbertus Anglicus, arranged head-to-foot (de capite ad pedem).
• When a nasal abbreviation could be read as either m or n, the editor checked how the word is spelled in full elsewhere and chose the more frequent form. For example, 'emplastre' was preferred over 'enplastre'.
• Concordances show the lemma, five words of left and right context, the keyword and a folio/line reference. They can be sorted by lemma, by the −1 column or by keyword. For example, sorting by −1 showed that cold is mostly predicative, typically after the 3sg present of ben.
• The electronic edition is hosted at hunter.uma.es. It belongs to joint projects of Málaga, Murcia, Jaén, Oviedo and Glasgow, and its corpus is about 1,200,000 tokens.
• For illegible passages, the printed edition consulted MS Hunter 307 and Getz's edition of MS Wellcome 537. It uses two apparatus layers at the foot of the page: one for scribal errors, corrections, headings and catchwords, and one for marginalia.

Interesting ideas

• I1. When a reading is ambiguous, look up how the same text spells the word without abbreviation, and fall back to the majority variant if both forms occur; the corpus itself serves as the tie-breaker.
• I2. Where many letter combinations are possible (minims, the shared þ/y graph), the editor lists every candidate, checks each against a dictionary and lets context make the final choice.
• I3. The transcription is graphemic rather than graphetic: positional letter shapes are merged into one grapheme, while real distinctions such as thorn and yogh are kept.
• I4. A faithful diplomatic layer and a normalised critical layer are produced from one transcription, and every editorial change is recorded in an apparatus instead of silently overwriting the source.
• I5. Re-sorting a KWIC concordance by the word immediately to the left of the keyword quickly reveals the dominant construction, such as predicative cold after ben.

For Mova

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

• I1 (tokens): Our lemmatiser could resolve ambiguous spellings or tokens by counting unambiguous occurrences of each candidate in the same document before turning to global statistics.
• I2 (tokens): Out-of-vocabulary or OCR-damaged words could be expanded into a candidate set, filtered by the lexicon and then chosen by syntactic fit, with the rejected candidates logged for explainability.
• I3 (tokens): This suggests a normalisation layer that folds purely visual variants into one form but never merges characters that carry meaning.
• I4 (learn): Silver data and repair rules could keep the original parse beside each corrected version, along with a record of the edit, so that every fix stays auditable.
• I5 (tree): Sorting a treebank's concordances by head or left neighbour could serve as a cheap check that a parser's attachments for an adjective match its typical usage.