📄 Early Modern English Medical Texts. 2010. Corpus [CD-ROM].

Summary

This is Peter J. Grund's review of Early Modern English Medical Texts (EMEMT), a CD-ROM corpus released in 2010 with a companion volume edited by Taavitsainen and Pahta. EMEMT follows the trend toward narrow but deep historical corpora that cover one domain over a limited period. It contains extracts of printed English medical writing from 1500–1700, grouped by broad category. It also includes rich metadata, images from the original books, search software by Raymond Hickey, and a spelling-modernized version produced with VARD.

Grund praises the contextual book and the presentation features, saying they raise the bar for how corpora are published. He also raises several criticisms:

• The categories rest on inconsistent principles.
• Manuscripts are excluded without much discussion.
• The corpus has no built-in periodization.
• "Medical" and "scientific" discourse are treated as if interchangeable.

His sharpest objection concerns modernization. Normalized spelling makes searching easier, but it does not automatically make results more accurate. Leftover variants and rule-induced errors still require careful checking.

Key points

• Size and coverage: EMEMT holds about two million words of extracts from 1500 to 1700. It continues Middle English Medical Texts (2005), and a Late Modern collection is planned.

• Categories: there are six main categories plus an Appendix:

  • General treatises and textbooks
  • Treatises on specific topics
  • Recipe collections and materia medica
  • Regimens and health guides
  • Surgical and anatomical treatises
  • Philosophical Transactions

  The last category is defined by publication format, while the others are defined by content.

• Software: texts are stored in extended ASCII and shown as a seven-branch tree. Users can switch to a list view to build their own groupings, which is also the only way to define periods. Each text links to a description with ODNB and EEBO references, along with page images.

• VARD modernization: this was done with Lancaster researchers. Only about 73% of the detected variants were modernized; the rest fell below a confidence threshold.

• Rule error: a blanket rule rewriting verbal –th as –s turned plural "sayth" into "says". This happened four times in Boorde's 1542 Dyetary of helth, and no post-hoc spot-check is reported.

Interesting ideas

• I1. Spelling normalization that applies only above a confidence threshold, leaving uncertain forms untouched, trades coverage for safety and makes the residue explicit.
• I2. The –th→–s failure shows that a morphological rewrite which ignores agreement breaks plural subjects, so context-blind rules corrupt grammar silently.
• I3. Grund notes that nobody spot-checked the automatic modernization after it ran, and that claims of greater precision must be measured rather than assumed.
• I4. Mixing classification axes (content for most categories, publication format for one) leaves categories uneven and hard to compare.
• I5. A focused domain corpus can refine conclusions drawn from general multi-genre data, as the intensifier study by Méndez-Naya and Pahta shows.

For Mova

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

• I1 (tokens): an archaic-form normalizer for old fables (hath, doth, -eth) could rewrite only high-confidence forms, keeping the original surface on every token so later steps can see both.
• I2 (tree): any spelling or inflection rewrite should consult the subject's number in the UD parse before changing a verb ending, rather than firing on the suffix alone.
• I3 (learn): every self-induced repair rule should be followed by a sampled audit and a negative-control check, so a rule's gain is demonstrated, not presumed.
• I4 (world): ontology and genre labels should keep separate facets (topic vs. format vs. audience) instead of one flat list mixing criteria.
• I5 (learn): evaluation should report results per domain slice, so averages over mixed corpora do not hide domain-specific parser or frame failures.