🗣️ Good translation: the general manual

This chapter says what a good translation is, whatever the two languages are, and how to make and check one. The per-language and per-pair manuals in this folder apply it. The examples are mostly English, French and Ukrainian, the three languages of the first manuals; the principles hold for any pair.

1. What a translation is for

1.1 The old argument and what is left of it

Every age of translation has argued between two poles: follow the words, or follow the sense. The argument is older than any of the languages in this folder, and it is worth knowing in outline because its terms are still used.

• Sense for sense. Jerome, defending his Latin version of a Greek letter in 395, says that outside Scripture he renders "sense for sense and not word for word", and he calls on Cicero, who translated the Attic orators "not as a translator but as an orator", keeping the sense and fitting the figures and the words to Latin usage. Jerome's own example of the cost of literalness is to render Homer word for word into Latin prose: the order becomes ridiculous and the poet barely articulate (jerome-1893). Note his exception: in Scripture "even the order of the words is a mystery". The exception is not piety alone; it is a statement that some texts are read for their form, and the form must then be kept even at a cost in ease.
• Three ways. Dryden, in the preface to Ovid's Epistles (1680), reduces all translation to three heads: metaphrase, "turning an author word by word, and line by line"; paraphrase, "translation with latitude", where "his words are not so strictly followed as his sense"; and imitation, where the translator takes "only some general hints from the original" (dryden-1900). He chooses the middle. The three heads are still the clearest names for three kinds of product, and a client who asks for "a translation" of a slogan often wants the third.
• Three laws, in order. Tytler's Essay on the Principles of Translation (1791) states three laws: the translation should give a complete transcript of the ideas of the original; its style and manner should be of the same character as the original's; it should have all the ease of original composition. What makes the essay still useful is that the laws are ranked: where one must be sacrificed to another, it is "highly preposterous to depart, in any case, from the sense, for the sake of imitating the manner", and wrong to sacrifice sense or manner to "a fancied ease". His example is d'Ablancourt's French versions, admirable "so long as we forbear to compare them with the originals" (tytler-1791). Tytler also gives the test most translators still use: imagine how the author would have written had he written in the target language.
• Two ways, not a mixture. Schleiermacher (1813) says there are only two ways: either the translator leaves the author in peace as far as possible and moves the reader towards him, or leaves the reader in peace and moves the author towards the reader; and the two must not be mixed, or author and reader may miss each other entirely (schleiermacher-1813). The modern names are foreignizing and domesticating (Venuti 1995, card only).
• Retrospective and prospective. Postgate (1922) names the same split by the translator's gaze: a retrospective (receptive) translation looks back at the original and is content to give as full and exact an impression of it as possible; a prospective (adaptative) translation looks at the reader and makes sure "this reader's prepossessions shall not be shocked" (postgate-1922).

What is left of the argument is this: there is no single right distance between the source and the target; there is a right distance for a given text, reader and purpose. A safety notice, a contract, a novel, a hymn, a menu and a tweet each sit somewhere else.

1.2 The working definition

The definition this folder uses is the one behind the MQM framework: a quality translation shows the accuracy and fluency its audience and purpose require and meets the other specifications agreed, taking the end user's needs into account (lommel-2013). The definition does three things:

1. It separates accuracy (does the target say what the source says?) from fluency (is the target good text of its language, whether or not it is a translation?).
2. It makes quality relative to a specification. A subtitle that omits half the words may be excellent; a contract that omits one may be worthless.
3. It allows a translation to be better than its source: a translator who repairs a defective source is credited, not penalised (lommel-2013).

The practical consequence is the brief. Before translating, know at least:

When there is no brief, the default is a stand-alone text for an educated general reader of the target language, in the register of the source, with full accuracy. That default is what the rest of this chapter assumes.

2. Meaning before form

2.1 What "the meaning" is made of

A sentence carries several layers of meaning, and a translation can keep one and lose another. Check each:

A translation is judged layer by layer, and the brief says which layers matter most. In a technical manual the propositional layer is nearly everything. In a novel's dialogue, register and attitude carry as much as the facts. In verse, the form is part of the meaning.

2.2 Translation procedures

Vinay and Darbelnet's Stylistique comparée du français et de l'anglais (1958; known to us by its card) named seven procedures that are still the common vocabulary of translators. They run from closest to the source to furthest from it:

1. Borrowing: take the source word. le week-end; вікенд; a croissant.
2. Calque: translate the parts. gratte-ciel from skyscraper; хмарочос likewise; world view from Weltanschauung.
3. Literal translation: word for word, when the result is grammatical and idiomatic. The book is on the table → Le livre est sur la table → Книжка на столі.
4. Transposition: change the word class without changing the meaning. After he returned → Après son retour (verb → noun); upon arrival → після прибуття or, better in most Ukrainian prose, коли приїде (noun → verb).
5. Modulation: change the point of view. It is not difficult → C'est facile; Keep off the grass → Не ходіть по траві ("don't walk on the grass"); the English passive I was told → On m'a dit → Мені сказали.
6. Equivalence: a different expression for the same situation, typically fixed. Ouch! → Aïe ! → Ой!; It's raining cats and dogs → Il pleut des cordes → Ллє як з відра.
7. Adaptation: replace a situation unknown in the target culture by one that does the same work. A reference to Thanksgiving dinner in a children's story may become a Christmas dinner in a French version; in a news text it stays Thanksgiving, explained if needed.

The list is a tool for describing what a translator did and for choosing when the literal rendering fails. It is not a ranking: literal translation is right whenever it is idiomatic, and the further procedures are used when it is not. The usual mistake of a weak translator is to stop at procedure 3 when the result is grammatical but not idiomatic; the usual mistake of an ambitious one is to jump to 6 or 7 when 3 was fine.

2.3 Compensation, explicitation, implicitation

Three moves keep the balance of a text when single sentences cannot:

• Compensation: a loss in one place is made up in another. If a French tu between two characters cannot be shown in English at the moment it first appears, the intimacy can be shown by a first name or a contraction a little later. If a Ukrainian diminutive (хатинка, "little house") has no English diminutive that is not twee, little old house or cottage may do in one place and a warmer word elsewhere.
• Explicitation: the target says what the source leaves implicit. French and Ukrainian often need a verb where English has a preposition: the man in the car → l'homme qui était dans la voiture when the context makes "who was in" the point. Explicitation is needed when the target grammar forces a choice (a gender, an aspect, a formal or informal you); it is a fault when it adds information the reader did not need, and corpus studies find translated texts more explicit on average than originals (Baker 1993, outside the library; the explicitation hypothesis of Blum-Kulka 1986, outside the library).
• Implicitation: the target leaves implicit what the source spells out, because the target reader infers it or because the target language finds it heavy. English in order to is often just pour and щоб; the English habit of repeating a noun (the Commission ... the Commission) is often a pronoun or a zero in French and Ukrainian.

2.4 Shifts are normal

A good translation is full of shifts: changes of class, structure, unit or level between source and target (Catford 1965, card only). A noun becomes a verb, a passive becomes an active, two sentences become one. Shifts are not losses; they are what a target grammar demands. What must not shift is the meaning layer the brief protects. Studies of human translations show how much semantic difference even professional translations carry between source and target, most of it deliberate and harmless, some of it error (vandevoorde-2020). The checker's task is to tell those two apart.

3. Register and style

3.1 The register coordinate

Every text is placed by three coordinates, and a translation must keep all three or move them on purpose:

1. Language: English, French, Ukrainian.
2. Variety modifier: a national variety (British, American, Indian English; France, Québec, Belgian, Swiss French), a regional dialect (Scots; Normandy; Hutsul, Polissian), a sociolect (Cockney, verlan, surzhyk), a period (Early Modern English, the French of the classical age, the Ukrainian of the 1920s).
3. Style of speech: official and legal; neutral written (news, reference); academic; colloquial spoken; slang (and whose slang: a generation's, a trade's, a city's); a song; a house style (a newspaper's, an institution's, a publisher's).

The three coordinates combine freely: a Québec French tweet in slang, a British English EU regulation, a Ukrainian folk song in a Carpathian dialect, an American English academic paper. A translator first reads the source's coordinate (from vocabulary, grammar, spelling, punctuation and the medium), then chooses the target's coordinate: the brief usually fixes the language and variety; the style of speech is normally kept.

3.2 Keeping the style of speech

Each language marks style with different means, so the style is kept by the target's own means, not by copying the source's:

The commonest register error in translation is flattening: rendering slang as neutral speech and official prose as neutral prose, because the neutral middle is the easiest. The second commonest is register drift: a sentence of slang in a text that was neutral, because one source word had a slangy equivalent the translator liked.

Rule. Choose the target style of speech once per text (or per character, in fiction), write it down in the brief, and check the draft against it, paragraph by paragraph.

3.3 Address: you, tu, vous, ти, ви

English has one you. French and Ukrainian each have two, and a translation from English must choose at every address. The choice is part of the register coordinate:

• Between strangers, in service, in official and commercial texts: vous, ви (and in Ukrainian official letters the capital Ви addressing one person).
• Between children, close friends, family, in most of the internet, in prayer: tu, ти.
• The switch from one to the other is itself an event in a story and should be marked when it happens.

Tatoeba keeps several French translations of one English sentence, each by a named author, and they show the choice exactly: Do you speak French? is rendered Parlez-vous français ? (tatoeba:415441-14077), Parles-tu français ? (tatoeba:415441-1031677) and Tu parles français ? (tatoeba:415441-7281436): the formal inverted question, the informal inverted question and the colloquial intonation question. All three are correct; only the context says which one is right.

Translating into English, the distinction is lost and must sometimes be compensated (§2.3): with a first name, a title, a change in the tone of the reply.

3.4 House style

Institutions and publishers fix choices that the language leaves open: spelling variants, the serial comma, quotation marks, how numbers and dates are written, how foreign names are transliterated, which terms are used. The European Commission's English Style Guide is an open example: British spelling but -ise (organise), single curly quotation marks, numbers one to nine in words and 10 and above in figures, a hard space as the thousands separator (15 000), dates in full in running text (6 June 2016) (commission-1982). A house style is not a matter of taste for the translator; it is part of the brief, and violating it is an error of the "locale conventions" or "style" kind in §10.

3.5 Varieties, dialects and sociolects in the source

A source in a dialect or sociolect raises the hardest register question: the target has no Scots, no Québécois, no surzhyk with the same social meaning. The options are:

1. Normalise: translate into the standard target language and say nothing. Right for informational texts where the variety carries no meaning (a technical manual written by a Scot).
2. Signal: translate into the standard, and mark the variety by a few chosen features (lexical items, a contraction, a word order) and by narrative comment where the source has it. Right for most fiction.
3. Substitute: render the variety by a target variety with a similar social position (a rural dialect for a rural dialect). Risky: it brings the target variety's own associations and places a Glasgow character in Marseille. Sometimes right for comedy and stage.
4. Invent: a non-local, recognisably non-standard target speech (dropped endings, colloquial grammar, a few mixed words) that marks the speaker's position without placing them on a map. Often the best choice for a mixed variety such as surzhyk translated into English or French (see Translating from Ukrainian).

Whatever the choice, it is made per variety, once, and applied consistently.

3.6 Songs are a style of speech

Songs have their own grammar, in every language: English lyrics routinely use ain't, double negatives, dropped subjects and auxiliaries (Ø Got my mind set on you), me and him went, rhyme- and metre-driven word order (the road I travelled down), dialect spellings (goin', 'cause). French songs drop ne, elide freely, invert for rhyme. Ukrainian folk songs keep vocative forms, short adjectives (зелен гай), fixed epithets (червона калина, ясний місяць), diminutives, and some songs are sung in surzhyk without anyone hearing it as an error. A song is translated by §9, not by the norms of written prose.

3.7 Varieties and styles as deltas over the standard

The working model of this folder for every non-standard voice is a delta over the standard. A dialect, a sociolect such as surzhyk, colloquial speech, slang, rap, a song's grammar, an archaic or biblical style, a newspaper's house style: each is the standard language plus two deltas,

• a grammar delta: the constructions it adds, drops or changes against the standard (dropped ne in spoken French; ain't and double negatives in English song and rap; Russian infinitives in -ть in surzhyk; thou and -eth in the KJV; the vocative and short adjectives of Ukrainian folk song);
• a lexicon delta: the words it adds, replaces or re-means (bagnole, boulot; шас, канєшно; quid, knackered; unto, spake).

A translation reads the source as standard + deltas, translates the standard, and then applies the target's own deltas that serve the same purpose in the text: the same social signal, the same closeness, the same humour, the same rhythm. It does not look for word-by-word dialect equivalents (a Cockney innit for a surzhyk канєшно), because a target variety brings its own place and history. In practice:

1. Identify which deltas the source uses and what each is for: placing a character socially, making a song singable and casual, sounding old and sacred, sounding official.
2. Translate the content through the standard.
3. Choose target deltas with the same function, at a similar density, from the target's own repertoire: its colloquial grammar, its slang, its song grammar, its archaic style.
4. Keep the choice consistent through the text (one delta per voice).

This is also how Mova models style today: in tools/translate an author is a delta over the standard encoder (§18), learned only where the standard gets that author wrong.

A contrast: when the varieties are nearly languages. For English, French and Ukrainian the deltas are small against a shared standard: a dialect speaker reads the standard without training. For languages such as Arabic or Chinese the "dialects" are closer to semi-languages: Moroccan and Iraqi Arabic, or Cantonese and Mandarin, differ from each other and from the written standard (Modern Standard Arabic; written Standard Chinese) in sound, grammar and core vocabulary far more than Scots from standard English, and are often not mutually intelligible. There the model needs a higher level of abstraction: the dialect is its own system (its own standard with its own deltas), linked to the written standard as a sister rather than derived from it as a delta, and translating a dialect text means choosing first which system the target text is written in.

4. Terminology

4.1 Terms and words

A term is a word or phrase with a fixed meaning in a domain: liability, shall, pharmacovigilance, вексель. Ordinary words may be varied for style; terms may not. The rules:

1. One concept, one term, throughout the text (and throughout a project). If the source uses two terms for one concept, find out whether the author meant two concepts; if not, the translation may unify them and note it.
2. The target term is the one the target domain uses, not the one a dictionary lists first. A term is checked against the target domain's own texts: its laws, standards, textbooks, reference works.
3. Fixed names are not translated anew: the official name of a body, a law, a treaty, a programme, a product; the established title of a book or film.
4. New terms (the source coins one, or the domain has none in the target) are made deliberately: by borrowing, calque or a native coinage (§2.2), and recorded in the project's glossary with the reason.

4.2 Where terms are checked

• Termbases. IATE, the EU's interinstitutional terminology database, holds the EU's terms in its 24 languages and Ukrainian is not among them; its data are reusable with attribution (to be confirmed for the bulk export, Sources). TERMIUM Plus, the Canadian government's bank, covers English and French. Ukrainian terminology is spread over national standards (DSTU), the terminology of the laws and dictionaries; r2u.org.ua gathers dictionaries.
• Parallel corpora. A term's established equivalent can be read off a translation memory of the domain: the DGT translation memory for EU law (English–French here: dgt-tm-v2021), Europarl for parliamentary language. A concordance of the source term shows the target terms the institution's translators actually chose, and how often.
• Monolingual target corpora. A candidate term is checked in the target language's own texts of the domain: does it occur, with the meaning, in texts not translated from the source language?

4.3 Borrowing or native term: decided by usage

The question "should the target use the anglicism or the native term?" has no general answer, and both dogmas are wrong:

• Full borrowing fills the target with source words its readers do not use, or use only in another register: a Ukrainian text with імплементувати, челендж and скіл where впровадити, виклик and навичка are what the readers say is translationese, not modernity.
• Full purism fills the target with native coinages its readers do not use: a French text that writes mercatique where every French speaker of the domain says marketing is a museum, not a translation.

The decision is made by measured usage, in the right slice of the target language:

1. Frequency: how often does each candidate occur, in texts originally written in the target language?
2. Register: in which styles of speech? A borrowing that dominates in tech blogs may be absent from official texts, and the reverse.
3. Domain: the term of the domain (le marketing in business French) may differ from the general word.
4. Period: usage moves; a measurement is dated.
5. House style and law: official bodies fix terms (France's Journal officiel lists of recommended terms; Ukraine's 2019 orthography for spelling; a client's termbase); within that house style the fixed term wins, outside it usage decides.

A count made for this manual shows the shape of such data. In the French side of Europarl v8 (2 051 014 sentence pairs, 1996–2011 parliamentary debates; counted with case-insensitive whole-word matches, 2026-11-06):

The table answers four questions differently: défi has won so completely that challenge is an error in this register; week-end has won against fin de semaine (which is the normal word in Québec, so the variety modifier matters); marketing is the word and mercatique a curiosity; e-mail, courriel and courrier électronique compete, with the native forms ahead in this formal register. Europarl is one register (formal spoken, transcribed and translated) and one period, and part of its French is itself translated, so the counts describe that slice only. This is exactly the measurement the Nightingales are being built to make over each language's Wikipedia and other open corpora (§18).

5. False friends

False friends are words that look alike in two languages and mean different things. They are the most frequent single kind of mistranslation between related languages, and MQM lists false friend as a subtype of mistranslation (lommel-2024). Four kinds:

1. Total: the meanings do not overlap. English actual ("real") and French actuel ("current"); English magazine and Ukrainian магазин ("shop"); French librairie ("bookshop") and English library.
2. Partial: one meaning is shared, others are not. French contrôler is "check, inspect" more often than "control"; English control is often maîtrise or gestion. Ukrainian контролювати similarly leans to "check, supervise".
3. Collocational: the words mean the same, but combine differently. English make a decision, French prendre une décision, Ukrainian ухвалити рішення (in official style) or прийняти рішення (widespread, criticised by purists as a calque, and itself a case for §4.3).
4. Register: the meaning is the same but the register differs. English demand is stronger than French demander ("ask"); English pretend is "feign", French prétendre "claim". Ukrainian претендувати is "lay claim to".

Two habits catch most false friends:

• Distrust every look-alike. When a target word looks like the source word, check the dictionary of the target language, not the bilingual one.
• Read the target sentence alone. A false friend usually produces a grammatical sentence with a slightly odd meaning; read without the source, the oddness shows.

Co-occurrence is not translation. In Europarl v8, of the 12 492 English sentences with actually, 197 have actuellement on the French side (counted 2026-11-06), which might look like a common false-friend error; read, they are nearly all sentences where actually and an unrelated "currently" both occur, as in europarl-v8:227542, where actually applied is rendered l'entrée en vigueur effective and at present is actuellement. The French renderings of actually in that corpus are mostly réellement or vraiment (2 077 sentences), en fait (1 328) and effectivement (999). A false-friend check needs word alignment, not co-occurrence counts.

The pair lists. The false friends of the three pairs, both directions and by domain, are listed and measured in false-friends/: English–French, English–Ukrainian, French–Ukrainian, each with a machine-readable list (<pair>.tsv), the kinds above plus frequency, connotation, grammatical (gender, countability, aspect, prepositions), interlingual homographs, calques and false enemies, and for Ukrainian the traps that come through Russian, surzhyk and Polish. A partial false friend is a choice that depends on context; those choices, for words and constructions with several valid renderings, are in context/ (English–French, English–Ukrainian, French–Ukrainian).

6. Grammar transfer

Grammar transfer problems arise wherever one language forces a choice the other leaves open, or offers a construction the other lacks. They are predictable from the two grammars, which is why the pair manuals list them. The main families:

6.1 A category one side lacks

The rule is the same each time: the translator must find the information the target needs, from context, from the rest of the text, from the world, or from the client; and must not invent it when it is not there. The doctor called in a text that never says who the doctor is is a real problem for French and Ukrainian: the masculine is the traditional default; a feminine form is right when the doctor is a woman; for an unknown person the target's current usage and house style decide (French le médecin / la médecin; Ukrainian лікар / лікарка, see Translating into Ukrainian). Europarl shows the forced information appearing out of the speaker's own gender: Mr President, I am happy to take the floor… is Monsieur le Président, je suis heureuse de pouvoir m'exprimer… (europarl-v8:1484): the French speaker is a woman, and English never said so.

6.2 A construction one side lacks

• The English progressive (is reading) has no French or Ukrainian form; it becomes the present (lit, читає), être en train de when the ongoing action matters, or an aspect choice.
• English phrasal verbs (give up, put up with) are single verbs or fixed phrases in French (abandonner, supporter) and prefixed verbs in Ukrainian (кинути, терпіти).
• English nominal chains (road safety awareness campaign) need prepositions in French (campagne de sensibilisation à la sécurité routière) and cases in Ukrainian (кампанія з підвищення обізнаності щодо безпеки на дорогах), and often a restructuring.
• The French and English cleft (c'est … qui; it is … who) is rare in Ukrainian, which uses word order (Дзвонила Марія).
• Ukrainian impersonals (Мене нудить; Вікно розбито) need a subject in English and French (I feel sick; The window has been broken, On a cassé la fenêtre).

6.3 A construction both have, used differently

The hardest transfers are between constructions that exist in both languages with different frequency or register. The passive is the standard case: English uses it freely in neutral and official prose; French prefers on, a pronominal verb or an active; Ukrainian prefers an active, the -но/-то impersonal or a reflexive, and a passive with an agent in the instrumental is heard as bureaucratic (§ Translating into Ukrainian). Transferring the source's construction because the target has it is the core of translationese (§11): every sentence grammatical, the text as a whole foreign.

6.4 Word order and information structure

English word order is fixed by grammar; French is fixed with more room for dislocation; Ukrainian is free, and its order carries information structure: the given first, the new last. Translating into Ukrainian, the English order often has to be reversed to keep the emphasis: A man came into the room (new subject) → До кімнати зайшов чоловік; The man came into the room (given subject) → Чоловік зайшов до кімнати. Translating out of Ukrainian, the order's information becomes English articles, clefts, passives or intonation marks.

6.5 Three languages, one sentence

Tatoeba links one English sentence to its French and Ukrainian translations, so the same sentence can be read in all three (English id first; French and Ukrainian ids in the second and third columns are the second half of tatoeba:<en_id>-<id> in en-fr/tatoeba and en-uk/tatoeba; CC BY 2.0 FR, authors in each row). Each line shows a transfer problem of this section:

7. Idioms and culture-bound items

7.1 Idioms

An idiom's meaning is not the sum of its words. The options, in order of preference when the register allows:

1. An equivalent idiom of the same meaning and register: It's raining cats and dogs → Il pleut des cordes → Ллє як з відра.
2. An idiom of a different image with the same meaning: to have other fish to fry → avoir d'autres chats à fouetter ("other cats to whip") → мати інші клопоти (no image).
3. A plain paraphrase in the right register: when no idiom fits, or the image would mislead.
4. A literal rendering, only when the image is the point (a text about idioms, a character who mangles them) or when it is transparent and the target can absorb it.

The literal rendering is the most frequent machine error with idioms. In the SimIdioms benchmark of Ukrainian–English idiom translation, the commonest failure of the tested large models was a literal translation that kept the sense but did not find a target idiom, and the idiom's transparency predicted quality only weakly (petruniv-2026-simidioms-corpus-benchmark-ukrainian-idiom). In a WMT 2025 English–Ukrainian system, a calque such as вийшов зі свого шляху for went out of his way is listed among the typical errors (grozea-2025).

7.2 Culture-bound items (realia)

Realia are things, institutions and customs of one culture: le bac, the Ivy League, вишиванка, Thanksgiving, la Sécu, ЗНО/НМТ. Strategies:

Institutions and legal concepts need special care: a French préfet is not a British prefect, a Ukrainian обласна рада is a regional council, not a "regional parliament". A good default for official texts is to keep the national name in the original (transcribed) at first mention, with a descriptive translation beside it, and to use one form consistently after that.

7.3 Humour and wordplay

A pun cannot usually be kept; its effect can. The order of preference: a pun at the same place on other words; a pun nearby (compensation); a non-pun with the same tone; a footnote only in scholarly editions. A translation that kills the joke without trying is an accuracy error of the "style" kind; one that adds a joke where the source had none is an addition. The WMT case of Let's eat granny, whose comma joke disappears in a rendering through Їдьмо, shows the loss (grozea-2025).

8. Names, numbers, dates, typography

8.1 Personal and place names

• Personal names are transcribed, not translated: Olha, not Olga, for a Ukrainian Ольга; Volodymyr, not Vladimir. Each language has its own romanisation system; the EU recommends the official national systems and warns that a Russian Ольга is Olga and a Ukrainian one Olha (commission-1982). A person's own preferred spelling, when known, wins over the system (Yuliia, Yulia, Iuliia are all used by people named Юлія).
• Place names use the target language's established exonym where one exists and is current (Londres, Лондон; Kyiv, not the Russian-based Kiev, in current English and French: Kyiv, Kyïv or the older Kiev depending on house style), and transcription otherwise.
• Into Ukrainian, foreign names follow the 2019 orthography's rules for foreign words and names: English w as в (Вашингтон, Вілсон), th as т or з (Артур, Резерфорд), h as г (Гарвард), with traditional exceptions in х (хобі, хокей); the letter ґ for a foreign g in personal names is allowed besides г (Гете and Ґете) (ukrainskyi-pravopys-2019, §§ 122, 124).

8.2 Numbers, units, money, dates

English column after (commission-1982). DGT shows the French percentage space and decimal comma together: 3,00 % en masse for 3.00 % by mass (dgt-tm-v2021:1385). Conversions of units (miles to kilometres) and currencies are made only when the brief says so, and the original is kept in brackets when the figure matters.

8.3 Punctuation and typography

Each language has its own typographic conventions, and they are part of fluency:

• Quotation marks: English “ ” and ‘ ’ (the EU uses single curly quotes first); French « » with a no-break space inside; Ukrainian « » without spaces, and „ “ or “ ” inside (ukrainskyi-pravopys-2019, § 164).
• Spaces before marks: French has a (narrow) no-break space before ; : ! ? and inside guillemets; English and Ukrainian have none.
• Dialogue: English uses quotation marks; French uses guillemets and dashes (the 19th-century Gutenberg editions in this folder's corpora use -- for the dash); Ukrainian uses the dash for dialogue in prose and guillemets for quotations.
• Order of closing marks: in Ukrainian a full stop, comma, semicolon, colon or dash never goes before the closing guillemet (ukrainskyi-pravopys-2019, § 164); in American English, commas and full stops go inside quotation marks; in British and EU English they go where the logic puts them.
• Capitals: English capitalises more (days, months, nationality adjectives, title words); French and Ukrainian capitalise less (lundi, понеділок; français, французький).

9. Songs and verse

9.1 What is being translated

A song is words, metre, rhyme, sound and a tune that is already fixed. A translation keeps some of these and gives up others, and the kind of translation is chosen by what the result is for:

The client must say which; "translate this song" without a purpose usually means a gloss for understanding or a singable version for performance, and the two are different products.

9.2 Singable translation

The constraints, in the order they bind:

1. Syllables to notes: one syllable per note, except where the original sets a melisma. French counts syllables with the sung mute e (u-ne bel-le cho-se); English and Ukrainian do not have it, so a French line rendered into English often needs extra syllables or a different word.
2. Stresses on strong beats: English and Ukrainian have free lexical stress, and a stressed syllable must fall on a strong beat; French stress is on the last full syllable of the group, and French song tolerates stress mismatch that English song does not.
3. Vowels on long and high notes: open vowels sing; i and u on a sustained high note are hard.
4. Rhyme at the same places, or at fewer places with a strong rhyme rather than at all places with weak ones.
5. Meaning: the images that carry the song's sense stay; the rest may change.

The public-domain round Frère Jacques shows the trade-off. The French Frère Jacques, dormez-vous ? … Sonnez les matines asks the sleeping friar to ring the morning bells; the traditional English Are you sleeping, Brother John? … Morning bells are ringing keeps the tune, the question and the bells, and changes who rings them: the bells are already ringing. A gloss would say "ring the matins"; the singable version chose rhythm and image over the exact act. Both are good translations of different kinds.

9.3 The register of a song

Song lyrics have their own grammar (§3.6), and the translation keeps the register of the song, not the norm of prose:

• An English lyric with ain't and a double negative is not corrected into standard French or Ukrainian; it gets the target's colloquial song grammar: J'ai jamais rien eu (dropped ne); Нічого в мене нема (Ukrainian negative concord is standard, so the colloquial mark goes elsewhere: a particle, a short form, a slang word).
• A dropped subject in English (Ø Woke up this morning) is natural in Ukrainian (Прокинувся вранці) and needs a subject in French, often a colloquial one (J'me suis réveillé ce matin).
• Word order bent for rhyme is allowed in the target too, within what its songs allow; Ukrainian folk song tolerates inversions (червона калина and калина червона) that prose would not.
• Dialect spellings (goin', 'cause) are rendered by the target's spoken spellings where it has a tradition of them (French p'tit, j'suis), and by vocabulary where it does not.
• A song in a mixed variety (a Ukrainian song in surzhyk) is not "corrected" into the standard when the mixed voice is the song's point; see Translating from Ukrainian and Translating into Ukrainian.

9.4 Verse that is not sung

Verse translation follows the same table without the tune: the translator chooses whether to keep the metre, the rhyme, both or neither, and says so in the edition. Arnold's lectures On Translating Homer (1861) are the classic English argument that the translator's metre must give the reader the original's speed, plainness and nobility, whatever its form (arnold-1861); the argument applies to any verse.

10. Translationese and naturalness

Translationese is the set of statistical traits that distinguish translated text from text written in the target language. Individual sentences may be fluent and accurate; the text as a whole is not what a native writer would write. Known measures include a lower type–token ratio (less varied vocabulary), more explicit cohesive markers, and a different distribution of parts of speech (wein-2024-lost-translationese-reducing-translation-effect). The typical sources:

• Interference: the source's constructions, word order and collocations carried over (§6.3).
• Normalisation and levelling: translators avoid the unusual and fall back on the safe middle of the target language (§3.2's flattening).
• Explicitation: more connectives, more spelled-out relations (§2.3).
• Unique items: target-language words and constructions with no counterpart in the source are under-used in translations, because nothing in the source triggers them (Tirkkonen-Condit 2004, outside the library): Ukrainian particles, French voilà, English phrasal verbs.

Translationese matters for three reasons. Readers notice it, even when they cannot say what is wrong. It affects measurement: machine translation test sets whose source side was itself translated from the target language reward different systems than test sets with original sources, and the WMT campaigns stopped using such "reverse-created" data after this was shown (graham-2020-statistical-power-translationese-machine-translation). And translated text fed back into a language's corpora shifts that language's measured usage: a Nightingale that learns "natural Ukrainian" from a Wikipedia full of translated articles learns some translationese too (§18).

The cure is the second reading: after checking the translation against the source, read it again as a text of the target language only, and rewrite every sentence a native writer would not have written. The test question is Tytler's: how would a native author have written this?

11. Quality checks

11.1 An MQM-like error typology

The Multidimensional Quality Metrics framework is an open catalogue of translation issue types arranged in a hierarchy, from which a project selects the types it needs; when an error fits no child type it is filed under the parent (lommel-2013 I1, I2). Its current core has seven high-level dimensions with subtypes, and a scoring model of weights and severities (lommel-2024). The typology this folder uses follows it, with the subtypes that matter between our languages:

The MQM authors also proposed a verity branch for texts accurately translated but false for the target locale (a wiring colour that is right in one country and wrong in another) (lommel-2013 I3), and the use of the same fluency and accuracy checks on the source, to catch defects before translation (lommel-2013 I5). Both are worth keeping in a checklist.

11.2 Severity and score

Each error gets a severity:

• neutral: a preferential change, or a problem that is not the translator's fault; no penalty;
• minor: does not affect meaning or usability;
• major: affects meaning or usability, but the text remains usable;
• critical: makes a part of the text unfit for purpose (a safety instruction reversed, a wrong dose, a wrong legal party).

The original MQM scoring, taken from the LISA QA model, weighted minor, major and critical errors 1, 5 and 10 and divided the sum by the word count (lommel-2013); the authors noted that the weights were not validated (lommel-2013 I4). The current scoring model publishes the parameters openly: error-type weights, severity multipliers (an example scale is 0, 1, 5, 25 for neutral, minor, major, critical), a reference word count, an acceptable number of penalty points per reference word count, and a calibrated scale on which the pass threshold is set (lommel-2024). Two of its findings matter for any checker:

1. Thresholds depend on the text. A legal translation may pass only with five penalty points per thousand words; user-to-user technical help may accept twenty-eight on the same count (lommel-2024).
2. Sample size changes the arithmetic. Linear scoring works only near the sample size it was calibrated on; one major error fails a one-page sample, but seven pages should not need seven major errors to fail. Below about 500 words, scores are too uncertain for anything but statistical quality control (lommel-2024).

11.3 Who checks and how reliable it is

• Human evaluation remains the reference, but it is noisy. In a study of English–Ukrainian sentence pairs rated by eight professional translators, individual experts' test–retest agreement ranged from 0.21 to 0.83 (ICC), and student raters agreed at only 0.43–0.54 (chaplynskyi-2026-professional-translators-versus-quality-estimation). Professional status alone does not make ratings reliable; repeated items, calibration and several raters do.
• Post-editing (correcting a draft into an acceptable translation) gives targeted references, but a score built on edit distance to post-edits depends on how many references there are and of which kind, and monolingual direct assessment against a reference carries a bias towards that reference; the two methods see different things and are best used together (bentivogli-2018).
• Automatic metrics compare an output with a reference: character n-gram scores such as chrF give partial credit to near-miss word forms without a tokenizer (popovic-2015 I1); neural metrics and large-model judges correlate better with humans on average, but they are opaque and can be fooled. On the English–Ukrainian data above, two large-model judges reached r ≈ 0.81–0.82 with the experts, and no single automatic metric exceeded r = 0.747 (chaplynskyi-2026-professional-translators-versus-quality-estimation).
• Negative controls. A check that cannot fail is not a check. A metric should be shown to score a shuffled or deliberately corrupted translation lower; a human review should include a few seeded errors. Mova's own translator measures its controls beside every system (§18).

11.4 A working checklist

Before delivery, read the whole target twice, once against the source and once alone:

Against the source

1. Every sentence, figure, name, date and negation is there, once.
2. Terms match the glossary and are consistent.
3. No look-alike is a false friend.
4. The register coordinate is kept (or moved as the brief says).
5. Forced choices (gender, aspect, tu/vous, ти/ви, articles) are supported by the text or the brief.

Alone, as a target text 6. Every sentence reads as written in the target language, in the chosen register. 7. No construction is there only because the source had it. 8. Punctuation, spaces, quotation marks, numbers and dates follow the target's conventions and the house style. 9. Links, tags, placeholders and layout are intact. 10. A reader who knows nothing of the source understands what the brief says they should.

12. Terminology by domain

Terminology work (§4) differs by domain: who fixes the terms, where they are checked, how much freedom the translator has, and which errors are critical.

12.1 Information technology and software localisation

• Who fixes the terms. The platform vendors' style guides and glossaries (operating systems, office suites), open-source localisation teams (each with a glossary per language), and national bodies for official equivalents (France's FranceTerme, Québec's OQLF). Ukrainian open-source teams keep their own glossaries; many terms of the 2000s (файл, папка, вікно, меню, завантажити) are settled, newer ones are not.
• Interface strings are translated as terms: one string, one rendering across the product (Save → Enregistrer → Зберегти; Cancel → Annuler → Скасувати; Settings → Paramètres / Réglages → Налаштування). A user who sees two words for one button thinks there are two functions.
• Imperative or infinitive. English buttons and menu items are imperatives or bare verbs; French uses the infinitive (Enregistrer, Imprimer) for commands and the imperative (Cliquez sur …) in instructions addressed to the user; Ukrainian uses the infinitive or the imperative by house style (Зберегти / Збережіть), most often the infinitive on buttons.
• Address. English you in UI text needs vous or tu in French and ви or ти in Ukrainian, as a product decision (§3.3); mixing them inside one product is an error of locale conventions.
• Anglicisms by measured usage (§4.3): the IT domain is where borrowings are densest and where official equivalents compete hardest (e-mail / courriel; smartphone / téléphone intelligent / ordiphone; хештег / гештег; імейл / електронна пошта). Counts for these are in Translating into French and Translating into Ukrainian.
• Placeholders and markup ({name}, %d, <b>) are kept exactly, moved to where the target grammar needs them, and never translated. A target sentence often has to change its order around a placeholder because the placeholder's case or gender is unknown (§16).

12.2 Science and medicine

• Science uses international Greek and Latin terms that English, French and Ukrainian share with spelling adaptation (photosynthesis / photosynthèse / фотосинтез), plus national terms (chemical element names: oxygen / oxygène / кисень; hydrogen / hydrogène / водень: Ukrainian uses native names for several elements).
• Units and notation follow the International System; decimal commas in French and Ukrainian (§8.2); billion traps (§8.2).
• Medicine has the most consequential errors (MQM "critical", §11.2): a dose, a contraindication, a laterality (left / right), a negation in an instruction. Drug names: the international non-proprietary name (INN) is used in every language with spelling adaptation (paracetamol / paracétamol / парацетамол; but acetaminophen is the US name); brand names differ by country. Patient leaflets in the EU follow fixed templates and terms in every EU language; the EMA's parallel corpus (EMEA in the English–French holdings, terms not checked) is learning and test material, not shipped.
• Register in medicine has two levels: the clinician's terms (myocardial infarction, infarctus du myocarde, інфаркт міокарда) and the patient's words (heart attack, crise cardiaque, серцевий напад). A patient leaflet uses the patient's words with the term in brackets once.

12.3 Law

Legal translation is translation between legal systems, not only languages. Three rules:

1. A legal term names a concept of one system. English trust, consideration, estoppel (common law) have no French civil-law equivalent; French préfet, notaire, acte authentique have no common-law equivalent; Ukrainian нотаріус is close to the French notaire, not to an English notary public. Use the established equivalent where one exists, a descriptive rendering with the original in brackets where none does, and keep it consistent.
2. The form is part of the meaning. Article numbering, defined terms, the order of conditions and exceptions, shall/may distinctions are kept exactly; a "better style" that merges two conditions changes the law.
3. Which text is authentic. A translation of a law is either authentic (equally binding, as in the EU and in Canada) or for information (most translations of national law). The translator of an informative translation says so on the document.

12.4 Bilingual and multilingual institutions: the EU and Canada

• The European Union legislates in 24 official languages, and every language version of a regulation is equally authentic. Most drafts are written in English (and some in French), translated by the institutions' services, and revised by lawyer-linguists. Terms are kept in IATE, house styles in the Interinstitutional Style Guide and in each language's style guide (for English (commission-1982)). The DGT translation memory (dgt-tm-v2021) quoted in these manuals is the record of that practice for English and French. Ukrainian is not an EU official language; Ukraine translates the EU's law into Ukrainian for its own legal approximation, and those translations are not authentic EU texts.
• Canada has two official languages with equal status in federal law. Federal laws are co-drafted: two drafters, one anglophone and one francophone, write the two versions side by side, neither translating the other, and the courts read the two together (the "shared meaning" rule). Canada is also bijural: federal law must work both in the common law of most provinces and in the civil law of Québec, so federal statutes sometimes use doubled terms (real property and immovables; biens réels et immeubles). TERMIUM Plus, the government's terminology bank, holds the agreed equivalents. Canadian practice is the reference for English–French legal terminology outside the EU, and for the principle that a bilingual text is two texts, not a source and a copy.
• Ukraine has one state language (Ukrainian) by its Constitution and its 2019 law on the state language; official texts are written in Ukrainian, and translations into English are informative.

13. Register deltas: a catalogue

The principle of §3.7, applied: each non-standard voice is the standard plus a grammar delta and a lexicon delta, and a translation maps the source's deltas to target deltas of the same function. The tables list the main deltas of each language; the pair manuals pair them.

13.1 English

13.2 French

13.3 Ukrainian

Kulish's Bible gives the biblical delta of 19th-century Ukrainian: Не вбивати меш for Thou shalt not kill (bible-kjv-kulish:EXO_20:13) uses the synthetic future вбивати меш ("thou wilt kill") of south-western speech, and Господь пастирь мій, не мати му недостатку (bible-web-kulish:PSA_23:1) the future мати му ("I shall have"). A modern Ukrainian translation of an archaic English or French text may borrow such forms as its archaic delta, exactly because they have the function "old and solemn" for a Ukrainian reader.

13.4 Rap

Rap is the hardest song form to translate, because rhyme is dense (internal and multisyllabic), rhythm is speech-like and fast, and the vocabulary is a group's. Translations are of three kinds: subtitles and glosses (meaning only, the most common); adaptations in the target's rap tradition (a French or Ukrainian rapper re-making the track: rhyme density and flow kept, content changed); and rarely close singable translations. The deltas map by function: English ain't, zero copula and double negatives → French dropped ne, elision and verlan → Ukrainian colloquial grammar, slang and, where the persona fits, surzhyk or English borrowings. Brand names, places and quotations are kept; boasts and disses are rendered by the target genre's own formulas.

14. Measuring usage: method and first results

§4.3 says that borrowing against native, calque against norm is decided by measured usage. This section says how the measurements in these manuals were made, what they show, and how they mislead.

14.1 Method

1. Choose the slice that matches the target text: register (parliamentary speech, law, encyclopaedia, fiction, speech), variety and period. The slices used here are Europarl v8 (formal spoken, transcribed and partly translated, 1996–2011), DGT-TM (EU law), the French and Ukrainian Wikipedias (encyclopaedic, dumps of 2026-10-01), and the Ukrainian Wikisource (older literature and documents).
2. Count forms, not words: a lemma's inflected forms (впровадж-), and the multiword forms of a construction (брати участь in all its tenses).
3. Count the competitors in the same slice, so that the ratio, not the raw number, decides.
4. Read a sample of hits: counts include other senses (виклик is also "a call"), markup (navigation templates), names and quotations.
5. Date the count and keep the method with it.

The Wikipedia counts were made by streaming each compressed dump through one whole-word, case-insensitive pattern match (no disk copy), on 2026-11-06.

14.2 What the counts show

• Most purist battles are over in neutral written prose, in favour of the norm: протягом 130 149 against на протязі 268, брати участь 184 446 against приймати участь 559, наступн- 263 724 against слідуюч- 811 in the Ukrainian Wikipedia; défi 11 590 against challenge 14 in Europarl's French.

• Some borrowings have won outright: аеропорт 107 182 against летовище 1 765; marketing 270 against mercatique 1 (Europarl); менеджер, волонтер, смартфон without a native rival.

• Some criticised calques are alive at scale: в / у якості 33 141, являється 25 977, відноситься до 8 208 in the Ukrainian Wikipedia; réaliser que 360 against se rendre compte que 352 in Europarl. Here usage alone does not acquit them: careful editing replaces them, and the replacement is always correct. A translator follows the careful register in neutral and official prose.

• Some pairs are genuinely shared: вертоліт 15 381 / гелікоптер 9 854; фотографія 46 842 / світлина 8 503; прийняти / ухвалити рішення 8 851 / 11 117. Register and house style decide.

• Spelling reforms win slowly: проєкт 177 478 against проект 104 922 seven years after the 2019 orthography; хештег 443 against the rule's гештег 42.

• Feminitives are established in encyclopaedic Ukrainian (авторка 10 413, директорка 3 415, лікарка 2 742, міністерка 1 247) and in institutional French (Madame la Présidente 20 135 against Madame le Président 6 483 in Europarl).

• The French Wikipedia agrees on most points and warns on others: courriel 13 696 against e-mail 4 449; se rendre compte que 9 827 against réaliser que 4 535; chargé de 140 564 against en charge de 4 744; but week-end 37 117 against fin de semaine 1 953, start-up 18 556 against jeune pousse 1 650, smartphone 14 472 against ordiphone 70. Its raw counts for challenge (119 752 against défi 49 112) and deadline (54 423) are inflated by names of competitions and an English-language trade publication cited in references: the slice's markup matters (§14.3).

French Wikipedia counts are in Translating into French §11; Ukrainian ones in full in Translating into Ukrainian §11.

14.3 How counts mislead

• Co-occurrence is not translation: English actually with French actuellement in the same Europarl sentence is almost never the false friend (§5).
• One slice is one register: Wikipedia is not speech; Europarl is not a novel. Вікенд (1 150 in the Ukrainian Wikipedia) is far more frequent in speech; на протязі is more frequent in older texts (897 in the Wikisource against 4 859 протягом).
• Translated text is inside the slice: part of every Wikipedia and part of Europarl's French is translated, so the counts include translationese.
• Editors' norms are inside the slice: Wikipedia editors correct to the norm, so it leans normative; social media would lean the other way.
• Senses and markup inflate counts: кейс includes phone cases; мапа includes map templates; наступн- in the Wikisource includes navigation links.

The Nightingales (§18) are being built to make these measurements by domain and period with the slice stated, and to score a whole translated sentence rather than one word.

15. Machine translation, post-editing and large models

Most translation today starts from a machine draft. The manual's principles apply to the draft as to any source; what changes is where the errors hide.

15.1 What machine drafts get wrong

Machine output is fluent, and its errors are the ones fluency hides:

• Literal idioms and calques: the sense survives and the idiom does not (petruniv-2026-simidioms-corpus-benchmark-ukrainian-idiom); went out of his way → вийшов зі свого шляху (grozea-2025).
• Invented words and wrong-language words: a WMT 2025 English–Ukrainian system produced non-words and russianisms (зірветься for вибухне) and transliterated slang (брух) (grozea-2025).
• Defaults instead of evidence: the masculine for an unknown person, the stereotype for a role noun (the nurse → feminine), vous everywhere or tu everywhere.
• Inconsistency: a term rendered three ways in one document, because sentences are translated separately.
• Omissions and additions that read smoothly: a dropped negation, an added clause; large models also hallucinate content not in the source.
• Lexical impoverishment: machine output tends to fewer distinct words and more frequent constructions than human translation (the translationese of §10), and different systems borrow anglicisms at different rates: in German, some commercial systems used fewer anglicisms than the human references and others more (shaitarova-2023-machine-vs-human-exploring-syntax).
• Formatting: lost or moved placeholders, broken markup, extra line breaks (in the WMT system above, 50 of 1 251 texts had to be re-translated because the model inserted empty lines) (grozea-2025).

15.2 Post-editing

• Light post-editing: make the draft correct and understandable; accept unidiomatic but correct wording. For internal or short-lived texts.
• Full post-editing: make the draft equal to a good human translation: register, idiom, terminology, house style. For published texts.
• Method: read the source first, not the draft; then the draft against the source sentence by sentence; then the draft alone as a target text (§11.4). A post-editor who reads the draft first is primed by it and keeps its structures, and post-edits are known to carry that bias (bentivogli-2018).
• When to throw the draft away: when more than about half of a sentence changes, translating afresh is faster and better.

15.3 Large models as translators and judges

Large language models translate well on average and are now used as quality judges; on English–Ukrainian sentence pairs two such judges agreed with professional raters better (r ≈ 0.81–0.82) than any single automatic metric (≤ 0.747) (chaplynskyi-2026-professional-translators-versus-quality-estimation). Their failures are those of §15.1, and they are opaque: a judge's score does not say which layer of meaning (§2.1) was lost. Use them as drafters and as a second opinion, with the human or measured check deciding. Mova's own rule is stricter: a large model never answers Mova's own assessment questions; it may teach, and the score is Mova's own (§18).

16. Constrained formats: subtitles, interfaces, plurals

16.1 Subtitles

Subtitles are condensed translation: the viewer reads while watching, so the text is limited by reading speed and screen space. Common professional practice is at most two lines, about 37–42 characters per line, a reading speed in the region of 15–20 characters per second, a minimum display time of about one second, and line breaks at syntactic boundaries. Omission is expected and is not an error unless it loses something the viewer needs (MQM's example: omissions are marked in a full translation and tolerated in subtitles, §1.2) (lommel-2013). Register is kept by word choice more than by length: a slang word costs no more characters than a neutral one. French subtitles keep French typography (spaces before ? !), Ukrainian subtitles theirs.

16.2 Interface strings and placeholders

• Placeholders change the grammar around them. Delete {count} files? → French Supprimer {count} fichiers ? works for the plural; Ukrainian needs the right plural form of файл for the number (§16.3); {name} liked your photo needs the past verb's gender in Ukrainian (вподобав / вподобала) and French agreement if a participle follows ({name} est connecté(e)). Solutions: restructure to avoid agreement (Користувач {name} вподобав… still masculine; {name}: подобається ваше фото), or get the gender from the data.
• Length: German and French strings are often a third longer than English ones; Ukrainian similar. Buttons truncate; the translator checks in context.
• Concatenation: strings built from pieces ("You have " + n + " new " + item) cannot be translated correctly in languages with agreement; a translator must report them as an internationalisation error (MQM's design and markup dimension).

16.3 Plural categories

Languages count differently. The Unicode CLDR plural rules, which software uses, give:

A translation into Ukrainian of an English string with two plural forms needs four.

17. A worked example

One short English text, translated into French and Ukrainian, with each decision named. The text is a constructed notice from a city library's website.

Our library will be closed from 24 to 26 December. Books borrowed before the 20th can be returned at the main desk or in the drop box by the entrance. If you've lost your card, don't worry: just ask a member of staff and they'll issue a new one. Happy holidays!

Brief (assumed). Public notice, informational and friendly; general readers; France for French, Ukraine for Ukrainian; full accuracy.

Decisions.

Result.

FR: La bibliothèque sera fermée du 24 au 26 décembre. Les livres empruntés avant le 20 peuvent être rendus à l'accueil ou dans la boîte de retour, près de l'entrée. Vous avez perdu votre carte ? Pas d'inquiétude : adressez-vous à un membre du personnel, qui vous en délivrera une nouvelle. Joyeuses fêtes !

UK: З 24 до 26 грудня бібліотека не працюватиме. Книжки, які ви взяли до 20 грудня, можна повернути на абонементі або в скриньку для повернення біля входу. Загубили читацький квиток? Не хвилюйтеся: зверніться до працівників бібліотеки, і вам видадуть новий. Зі святами!

Notes. The French turns the conditional into a question (Vous avez perdu votre carte ?), a modulation that keeps the friendly tone; the space before ? and : follows French typography. The Ukrainian puts the dates first (the new information, the closure, last), drops the possessive our, uses the participle-free relative які ви взяли rather than взяті вами (§6.3, Translating into Ukrainian), repeats грудня, and uses a question too. Neither target has a word for word drop box; both describe it.

18. How Mova handles it

18.1 What exists today

Mova's own translator is tools/translate. It is deterministic, learned from counts, and has no large model translating or teaching inside it. What it does:

• It reads a text down the standard stack of its language (tokens, Universal Dependencies tree, frames, world) and from there into a meta-world: a graph of concepts with coarse grammar and roles, which belongs to no language and no author. It writes the meta-world out with a standard encoder of the target language and the target author's delta.
• The languages it covers are English (between Bible editions: JPS 1917, KJV, WEB, ASV, and, with translate many, every English edition eBible holds) and Biblical Hebrew (the Leningrad Codex as MACULA trees, into English; and the parallel Tanakh's Hebrew and English site pages in both directions). It does not translate French or Ukrainian.
• An author is a delta over the standard: a sparse set of overrides learned only where the standard gets that author wrong. «Author» can be any style: besides the Bibles, the school's plain book lines, its conversational test questions (the half-slang voice), its formal exam questions and the tales are styles with their own deltas. This is the register coordinate of §3.1, implemented for English: the language is the standard, the style of speech is the delta. It has no variety modifier separate from the author yet.
• Typography per author is learned (--typo on): each edition's quotation marks and possessive (Moses’ or Moses'), as in §8.3.
• Its quality measure is exact: a translation is scored by the residual, the bits needed to repair it into the target byte for byte, split by stage (reader, encoder, stack) and by class: lexical, construction, order, function words, punctuation and case, morphology. These classes are an error typology in the sense of §11.1, measured, not judged; they are closest to MQM's accuracy (lexical, construction), linguistic conventions (morphology, punctuation and case, function words) and style (order).
• Logical closeness (logic.rs) compares two readings' events: a matched event whose negation differs, whose modals differ or whose role is filled by two different names is a contradiction; an entity that is agent on one side and patient on the other is a reversal. These are the critical accuracy errors of §11.2, counted per 1 000 words. A sensitivity control puts one change of meaning into each target verse (negation flipped, roles swapped, a number changed, events reordered) and checks that the distance moves, the negative control of §11.3.
• Normal mode scores output against the target by exact match and chrF.
• Every run measures negative controls beside the systems: verses shuffled inside blocks, the wrong author's delta, lexicons and tables learned on shuffled pairs.

Parallel corpora. tools/parallel converts the open English–French corpora quoted in these manuals into one aligned format and aligns Project Gutenberg book pairs with the Gale–Church length model. The same converters built the English–Ukrainian and French–Ukrainian corpora (Tatoeba EN–UK 217 646 and FR–UK 34 310 sentence pairs; the Kulish/Puluj Bible of 1905 against the WEB, the KJV and Louis Segond), 2026-11-06.

Reading the source languages. English is read by Mova's own parser and frames. French is read by the same parser after a treebank's tags are mapped (UD French GSD, train → dev, LAS 87.00 with own tags, 2026-10-22; fr manual). Ukrainian has no Mova parser baseline.

18.2 What is planned and being built

• The Nightingales 🐦 (tools/nightingale, being built). One Nightingale per language sits under the Eel and builds a naturalness model of its language from that language's Wikipedia, on the CPU, with its own Sol consultant. In this manual's terms, a Nightingale is the measuring instrument for §4.3 and §10: it would give the frequency of a borrowed against a native term, per domain (from the article's category) and per period (from revision dates), and score how far a candidate sentence lies from text written in the language. Its limits must be stated with each figure: Wikipedia is one register (neutral written, encyclopaedic); part of every Wikipedia is translated, often from English, so the model contains translationese of its own; and Wikipedia is CC BY-SA, so its models are for measurement inside Mova, not for shipped data. Mova holds the English Wikipedia today; the French and Ukrainian dumps are to be fetched for their Nightingales.
• The Eel's translation rounds tr-<src>-<tgt> (being built): rounds of the Eel, Mova's learner, that would train and measure a translation direction (for example tr-en-fr) on the parallel corpora, with the same gates and held-out discipline as its parser and autoencoder rounds.
• What neither does: no large model translates for Mova or answers its quality questions. The Sol consultants advise the Nightingales; the measurements are counts.

18.3 Where Mova stops

• No French or Ukrainian translation; no articles, gender or aspect transfer between different languages is implemented (the English editions share them).
• The register model is per author, for English; no variety modifier, no sociolect, no song grammar.
• No terminology control, no termbase, no false-friend list.
• No MQM annotation: the residual classes are measured, not mapped to MQM severities.
• The parallel corpora are aligned by length only; their direction of translation is often unknown.

Sources

Classics of translation theory (public domain): Jerome, Letter 57 to Pammachius (trans. 1893) (jerome-1893); Dryden, Essays, ed. Ker 1900 (dryden-1900); Tytler, Essay on the Principles of Translation (1791; 1907 reprint) (tytler-1791); Schleiermacher, Ueber die verschiedenen Methoden des Uebersezens (1813) (schleiermacher-1813); Arnold, On Translating Homer (1861) (arnold-1861); Postgate, Translation and Translations (1922) (postgate-1922). Known only by their cards: Vinay and Darbelnet 1958; Catford 1965; Venuti 1995. Outside the library: Blum-Kulka 1986; Baker 1993; Tirkkonen-Condit 2004 (logged in docs/usage.tsv).

Quality and translationese (open access): (lommel-2013) (CC BY-NC-SA, paraphrased); (lommel-2024); (bentivogli-2018); (graham-2020-statistical-power-translationese-machine-translation); (wein-2024-lost-translationese-reducing-translation-effect); (vandevoorde-2020); popovic-2015; (chaplynskyi-2026-professional-translators-versus-quality-estimation); (grozea-2025); (petruniv-2026-simidioms-corpus-benchmark-ukrainian-idiom).

Style and orthography: European Commission, English Style Guide (8th ed., updated November 2025) (commission-1982); Український правопис (2019) (ukrainskyi-pravopys-2019).

Corpora and counts: Tatoeba EN–FR, EN–UK and FR–UK, Europarl v8, DGT-TM 2021, the Kulish/Puluj Bible (1905), the French and Ukrainian Wikipedias (dumps of 2026-10-01, counted only) (see Sources). Further cited in §§12–16: (shaitarova-2023-machine-vs-human-exploring-syntax), (fellerer-2022-pro-drop-sw).

The Elamur team

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/languages/translation/general/