MovaTranslation: manuals for Mova
This folder is an open manual of good translation, written for Mova and for anyone who translates or checks translations. It has four layers:
- Good translation: the general manual: what a good translation is, whatever the languages. Meaning before form, register and style, terminology, culture-bound items, names, numbers and typography, songs and verse, quality checks with an error typology, and how Mova handles each point.
- Per language, two manuals, because the problems of reading a language and the problems of writing it differ:
- from the language: what a translator must notice in a source text in that language (ambiguities, implicit meaning, register markers, traps);
- into the language: what a text in that language must look like to read as native (grammar, word choice, punctuation, the pull of the source language).
- Per pair, a pair manual: the contrasts of the two languages side by side, both directions, with many examples.
- Per pair, two lists: the false friends and the context-dependent renderings, each a manual and a machine-readable list measured in the parallel corpora.
Contents
| file | what it covers |
|---|---|
| Good translation: the general manual | Principles; meaning before form; the register model (language + variety + style of speech); terminology; false friends; grammar transfer; idioms and culture; names, numbers, dates, typography; songs and verse; quality checks (an MQM-like typology); terminology by domain (IT, science, medicine, law, EU and Canadian bilingual practice); a catalogue of register deltas; measured usage in Wikipedia; machine translation and post-editing; subtitles, interfaces, plurals; a worked example; how Mova handles it |
| Translating from English | Reading English as a source: polysemy, phrasal verbs, nominal chains, the unmarked you, tense and aspect, hedging, the register signals of English, British and American sources |
| Translating into English | Writing English as a target: word order and end weight, articles, tense, idiom and collocation, plain English, house styles (the EU English Style Guide), translationese in English |
| Translating from French | Reading French as a source: on, the tenses of narrative, the subjunctive, the formal and the spoken registers, tu and vous, nominal style, French typography as a signal |
| Translating into French | Writing French as a target: the noun phrase, determiners, agreement, clitics, the tenses, register, anglicisms judged by usage, French typography |
| Translating from Ukrainian | Reading Ukrainian as a source: case and aspect, free word order and information structure, ти and ви, diminutives, the registers, surzhyk as a variety |
| Translating into Ukrainian | Writing Ukrainian as a target: verbs over nouns, passives, participles, case government, anglicisms and purism judged by usage, feminitives, the 2019 orthography, surzhyk when it is right |
| English and French: the pair manual | English and French side by side, both directions, with examples from open parallel corpora |
| English and Ukrainian: the pair manual | English and Ukrainian side by side, both directions |
| French and Ukrainian: the pair manual | French and Ukrainian side by side, both directions |
| false-friends/ | False friends of each pair, both directions, by domain, measured in the parallel corpora: en-fr, en-uk, fr-uk, with machine-readable lists (*.tsv) |
| context/ | Context-dependent renderings: for words and constructions with several valid renderings, which fits which sense, register, domain, collocation, text type and reader: en-fr, en-uk, fr-uk, with lists (*.tsv) |
| examples-queue.tsv | The entries of both kinds of list that still need an example sentence (planned: our own sentences translated by the Codex teachers and gated) |
| Sources | Every source cited, with licence and where it is held |
The stance of these manuals
- Meaning, purpose and reader decide. A translation is good when it says what the source says, in the way the target reader of that kind of text expects, for the purpose agreed. Neither the source's form nor a style dogma decides by itself.
- Neither borrowing nor purism is a rule. An anglicism in French or Ukrainian is sometimes the word people actually use and sometimes a lazy calque; a native coinage is sometimes the living word and sometimes a museum piece nobody says. The choice is made by measured usage in the target language: how often the borrowed and the native form occur, in which register, in which domain, in which period, and in the house style the text belongs to. The manuals say where these measurements come from and where Mova will make them (the Nightingales, below).
- Register is a coordinate, not a label. A text is placed by its language, a variety modifier (national or regional variety, dialect, sociolect such as surzhyk) and a style of speech (official, neutral written, colloquial, slang, a song, a publication's house style). A translation keeps the coordinate, or moves it on purpose and says so.
Conventions
- Examples from the open parallel corpora Mova holds carry their ids:
tatoeba:<en_id>-<fr_id>,tatoeba:<en_id>-<uk_id>,tatoeba:<fr_id>-<uk_id>(Tatoeba, CC BY 2.0 FR; each sentence has a named author, and rows with an unknown author are not quoted),europarl-v8:<line>(Europarl v8, European Parliament texts, quoted with attribution),dgt-tm-v2021:<line>(the DGT translation memory, reused under Commission Decision 2011/833/EU),bible-kjv-lsg:<BOOK>_<c>:<v>andbible-web-lsg:…(KJV or WEB with Louis Segond 1910, public domain; the KJV is quoted for study, see Sources),gut-<book>:<n>(Project Gutenberg book pairs with identified 19th-century translators, public domain),bible-web-kulish:…,bible-kjv-kulish:…,bible-lsg-kulish:…(the Kulish/Puluj Ukrainian Bible of 1905, public domain, against the WEB, the KJV and Louis Segond). The corpora and their licences are described in Sources. - The direction of a corpus pair is not always known. Europarl and DGT rows do not say which side is the original; Tatoeba sentences are written in either language; the Gutenberg pairs are French originals with English translations; the Bible pairs are two translations of a third text. Each example says what is known.
- Unmarked examples are constructed for these manuals.
- Glosses are in italics;
→marks a translation direction;✗marks a rendering we judge wrong,?one that is doubtful,✓one we recommend. - Counts carry their date and the file they were made on. Counts made for these manuals were made on 2026-11-06.
- Sources are cited in our own words, by the library slug, e.g. (lommel-2024), or (lommel-2013 I4) where the card numbers its ideas. Works under share-alike or non-commercial terms are paraphrased, never copied.
How Mova handles translation today, in one paragraph
Mova's own translator, tools/translate, is deterministic and learned from counts, not from a large model. It translates between English Bible editions through a language-free meta-world, from the Hebrew Bible into English, and between the parallel Tanakh's Hebrew and English site pages, and it measures every translation as the bits needed to repair it exactly. It does not translate French or Ukrainian. tools/parallel has built the English–French, English–Ukrainian and French–Ukrainian parallel corpora quoted here. French is read by Mova's parser after a treebank's tags are mapped (fr manual); Ukrainian has no Mova parser baseline yet. Usage counts in the manuals come from these corpora, from Europarl, and from the French and Ukrainian Wikipedias (dumps of 2026-10-01, streamed and counted, not quoted). Two parts are planned and being built: the Nightingales 🐦 (one per language, each sitting under the Eel and building a naturalness model of its language from that language's Wikipedia on the CPU, each with its own Sol consultant) and the Eel's translation rounds tr-<src>-<tgt>. Each manual's last section says exactly what exists and what is planned for its points.
The Elamur team
This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/translation/