📄 Annotated English

Summary

English spelling predicts pronunciation poorly. Spelling reforms keep failing because they would make existing texts obsolete, hide etymology, create new homographs and break software that processes English. Hernández-Orallo proposes Annotated English, a set of diacritic-like marks placed above, below or around letters, while the spelling itself stays exactly as it is. The design centres on a body of default pronunciation rules. Only the parts of a word that break those rules get a mark, so regular spellings stay clean and a mark tells the reader "this is an exception". The system covers silent letters, stress, syllable separators, consonant values, semivowels, vowel qualities and diphthongs. Because the base text is untouched, an annotated book keeps the same length. A reader who does not know the marks can ignore them, and someone new to the system can pick up most marks after a few pages of examples. The target users are native children learning to read, learners of English as a foreign language, and anyone meeting unfamiliar technical words. The paper also gives a full LaTeX implementation of the symbols.

Key points

• No letter of the original text is changed. Pronunciation information is layered on top as marks, unlike respelling schemes such as Basic Roman Spelling and Interspel, or new alphabets such as Unifon.
• Default rules keep marks sparse. Earlier marked-letter systems lacked such defaults and would flag even regular words like "cat" and "cell". Examples include the 1836 edition of Webster's Elementary Spelling-book and dictionary diacritics before respelling and IPA took over.
• Stress is treated as essential. The author notes that many reforms neglect it, even though it separates noun/verb pairs such as "an imprint" vs. "to imprint", which differ only in stress.
• Printing an IPA line under every line is rejected because it would double book length and force readers to switch between two scripts.
• An appendix supplies LaTeX macros for all the marks, built on packages such as tipa, extraipa, mathabx, accents and yhmath. It also gives workarounds for package conflicts, for example with the LNCS class.

Interesting ideas

• I1. Mark only deviations from a default rule set, so that the absence of a mark itself carries information ("regular here") and rare marks stand out as exceptions.
• I2. Add information as a non-destructive overlay on the original text, so the base stays fully usable by any consumer that ignores the layer.
• I3. Design the annotation scheme so its rules can be inferred from a few annotated examples rather than from a manual.
• I4. Treat stress as a disambiguating feature for noun/verb homographs that look identical in spelling.
• I5. Judge a reform by the full cost of switching, not only by its end state, including obsolete texts, lost etymology and broken downstream software.

For Mova

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

• I1 (tokens): Lemma and morphology tables could store only the irregular forms against default inflection rules, so every non-default analysis is visibly flagged and easy to explain.
• I2 (tree): Dialect conversions and parser decisions could be kept as standoff layers over the untouched token stream, letting raw-text tools and explainability traces coexist.
• I3 (learn): Self-induced repair rules could be judged by whether they are this compact and learnable from examples, preferring few general defaults plus explicit exceptions over many narrow rules.
• I4 (back): When generating text toward speech or reading-aid output, the POS chosen in the meaning graph could drive a stress mark, so "imprint" is voiced correctly for its role.
• I5 (world): Changes to the ontology or verb-class inventory could be checked for the cost of migrating existing compiled knowledge and dependent readers before they are adopted.