Skip to content

ADR-0003: Culinary terms carry a language tag and are never translated

Status

Accepted.

Date

2026-08-19

Context

The 1907 translator of Le Guide Culinaire left many terms in French. Espagnole, velouté, béchamel and despumation appear untranslated in an English text, because each names a specific preparation rather than describing one.

Normalising a corpus to one language is a common first step, and it is available here. Folding espagnole to "brown sauce" would make lookup simpler and the catalogue smaller.

It would also be wrong, and the cost grows as the corpus moves away from Europe. Mole is Nahuatl, from mōlli. The vocabulary beneath it is Spanish over a Nahuatl substrate, including nixtamal, comal, metate and epazote. Rendering mole as "sauce" discards the distinction the word exists to carry.

There is a second reason. This corpus is intended to test whether a taxonomy induced from one tradition generalises to another. Translating everything into English would test whether an English-shaped taxonomy fits English-translated descriptions. That test passes for the wrong reason.

Decision

A term is stored as its surface form, tagged with the language it was written in, and linked to a language-independent concept identifier. No code substitutes a translation for a term.

The model has three parts.

Surface form. The term exactly as the source writes it, including diacritics. Velouté is stored as Velouté.

Language tag. An ISO 639-1 code. Inference uses three signals: diacritics, French noun-adjective word order, and a lexicon of French terms that keep their spelling in English.

Concept identifier. A folded form that resolves orthographic variation only. Velouté and VELOUTE reach veloute. Folding decides nothing about meaning, so it never resolves equivalence across languages. Establishing that salsa española and espagnole denote one concept requires evidence.

An English rendering may be recorded beside a term as a gloss. A gloss is labelled as one and never replaces the term.

This anticipates a translation stage. When one arrives, it may translate prose, meaning method descriptions and commentary. It may not translate the term layer. An unguarded translation pass renders mole as "sauce" and destroys the ontology in one step.

Consequences

Positive

  • The catalogue can represent a concept with no equivalent in any other language, which is the correct result for much of the mole vocabulary.
  • Cross-language equivalence stays a claim requiring evidence rather than a side effect of string handling.
  • The generalisation test stays honest, because the corpus is not pre-shaped toward the taxonomy being tested.

Negative

  • Lookup is harder. A reader searching for "brown sauce" must be met by the ending-match fallback rather than by a normalised index.
  • Language inference is heuristic and will misclassify terms. The consequence is a wrong tag rather than lost information, because the surface form is preserved either way.
  • Every future stage must be checked for an implicit translation step.

References