Mova0.6 pre-release⌕
☰
home / docs / languages / translation / context / English and French: renderings by context

English and French: renderings by context

Many English and French words have several good translations, and the right one depends on the context: the domain, the register, the collocate next to the word, the syntactic frame, the reader. Raise is soulever with an issue, augmenter with prices, élever with children, lever with a hand and sensibiliser with awareness. French relever is note in a finding, meet with a challenge, raise with a threshold and fall within in relever de. This file lists such words and constructions in both directions, one row per word in one context, says which rendering is preferred there and why, and gives each row a corpus count and, where one exists, a corpus example. It goes beyond the lexicon sections of the pair manual (sections 3, 4.3, 7, 8, 17 and 18) and stays consistent with them. Where the word is a partial false friend, the row says so and points to the false-friend list.

Data and method

The list is our own, hand-made data. Each row names a source word or construction (src_match), optionally a context that must also be in the source sentence (ctx_match: a collocate, a frame word), the forms of the preferred renderings (good_match) and of the rendering to avoid (bad_match). tools/parallel (parallel choices) counted the rows on 2026-11-06 over the English–French corpora: Tatoeba, Europarl v8, the DGT translation memory, the Gutenberg book pairs and the WEB–Louis Segond Bible (counted and quoted), and EMEA, JRC-Acquis and the ECB corpus (counted only, never quoted). The counts are co-occurrence, not alignment: "a recommended rendering in N" means N sentence pairs where the source matches and the target contains one of the listed renderings somewhere, not necessarily as the translation of that word. DGT and JRC-Acquis repeat many sentences, which inflates their counts, and Europarl and DGT do not say which side is the original. Read the counts as orders of magnitude and as ratios between the preferred and the avoided form, not as precise frequencies.

Each example was picked by the tool (the shortest quotable sentence pair with the source pattern, the context and a preferred rendering, and without the avoided one) and then checked by eye. We read all 493 examples of the first full run; 36 were wrong in sense or showed an alternative rather than the preferred rendering. We tightened the patterns of 15 rows (for example as far as possible also matched the spatial le plus loin possible, and immediate release caught the release of a prisoner), set 21 hand-found corpus ids (verified by the tool), and left 5 rows without an example where the corpora have no clean case. After the final run we read the 10 examples the tool picked anew and replaced one more (an implement row had picked a legal protocol) by no example. In all, 487 of the 495 rows have a checked corpus example. An example shows one translator's choice in one sentence; it is evidence of usage, not a rule.

Each row also carries a difficulty: plain, or hard:<reason> (ambiguity, register, law, idiom, culture, aspect, case, poetry) when a good general translator would get it wrong without the entry. For rows without a corpus example we plan to write our own source sentences and have them translated by the Codex teachers through mova-teacher (Terra for plain rows, Sol for hard ones), gated by back-translation agreement, Mova's parse and pragmatics of both sides and a hand-checked sample; accepted sentences will get ids ours-tr:en-fr:<n>. None exist yet.

The literature on lexical choice agrees with the approach of one row per context: sentence-level translations go wrong exactly where deixis, formality and lexical cohesion depend on the wider text (voita-2019-when-good-translation-is-wrong); the formality of you cannot be read from English and has to be supplied as a constraint (sennrich-2016-controlling-politeness-neural-machine-translation); native speakers decide between near-synonyms by context, and those decisions can be written down as rules (barua-2024-language-models-disambiguate-lexical-choices); and French–English partial cognates are disambiguated by their context words (frunza-2007-tool-detecting-french-english-cognates), which is what ctx_match does here.

The list has 495 rows: 317 from English (209 plain, 108 hard) and 178 from French (96 plain, 82 hard). A few ratios show what the counts are good for: in DGT and its neighbours, shall enter into force co-occurs with the present entre en vigueur in 41 097 of 41 593 sentence pairs and with the future or devra in 245; fiscal policy (2 389) with budgétaire in 1 608 and with fiscal in 679, much of it in sentences that mention both budget and tax; make sense (1 780) with avoir du sens and its relatives in 976 and with faire sens in 8; realise that (2 919) with se rendre compte and its relatives in 1 532 and with réaliser in 514. The authored list is kept beside the counted one (columns up to difficulty); the tool adds n_src, n_good, n_bad, by_corpus, example_ok, frequency_rank and example_status.

Function words and constructions

Function words carry most of the context-dependent choices that a word list misses. English you has to become tu, vous or on; shall and should change meaning between a statute, a recital and a conversation; since, as, while and still are each two or three words in French. French on is we, people, someone or an English passive; en effet, aussi with inversion, or and dans la mesure où look transparent and are not.

English → French

<!-- entries domain=function src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=function src_lang=fr --> <!-- /entries -->

Grammar: frames that change the rendering

Here the syntactic frame, not the word, decides: make + adjective is rendre, make + verb is faire; for + duration with a state that still holds is depuis with the present; have just done is venir de; -ing after a preposition is an infinitive. In the other direction mettre une heure is take an hour and arrêter de is stop -ing.

English → French

<!-- entries domain=grammar src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=grammar src_lang=fr --> <!-- /entries -->

Everyday words

The high-frequency verbs of everyday English (get, take, keep, run, mean, save, miss) have no single French equivalent; the collocate decides. French passer, rendre, porter, tenir, compter are just as wide the other way. Tatoeba is the best source for these rows, with short sentences and named authors.

English → French

<!-- entries domain=everyday src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=everyday src_lang=fr --> <!-- /entries -->

Collocations

A collocation fixes the verb: an issue is soulevée, a promise tenue, a challenge relevé, a speech prononcé, a meeting a lieu. The calque is usually grammatical and always wrong in register (✗ rencontrer un défi, ✗ prendre place). Europarl gives large counts for these.

English → French

<!-- entries domain=collocation src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=collocation src_lang=fr --> <!-- /entries -->

Register: official, neutral, colloquial, slang, song

A translation keeps the register coordinate of its source (README, stance): a letter stays a letter (Madame, Monsieur, and Cordialement), slang stays slang (mec, trop bien, kiffer), a song keeps its address words, a euphemism stays a euphemism. The rows below give matching forms at each level; the T/V choice of French belongs here too, because salut commits the speaker to tu.

English → French

<!-- entries domain=register src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=register src_lang=fr --> <!-- /entries -->

Shared Latin and French vocabulary: partial false friends in context

These words share a form and part of their meaning, so the context decides whether the cognate is right. Actual is never actuel, but réaliser is right for a dream and doubtful for "become aware"; important is important for weight and large for size. The full false-friend list is False friends: English and French; the rows here show the choice inside a context.

English → French

<!-- entries domain=latin-french src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=latin-french src_lang=fr --> <!-- /entries -->

Politics and parliamentary language

Europarl shows how speeches and resolutions are rendered: I welcome is je me félicite de, calls on is invite, je tiens à is I would like to, intervention is speech. French administrative nouns (dossier, enjeu, volet, chantier, acteurs, instances, collectivités territoriales) need a different English noun in nearly every sentence; the Euro-English calques (actors, instances, put in place) are understood in Brussels and sound foreign elsewhere.

English → French

<!-- entries domain=politics src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=politics src_lang=fr --> <!-- /entries -->

EU law

EU legislative drafting has fixed equivalents, and the DGT memory records them at scale (pair manual section 17). English shall in enacting terms is the French present; should in recitals is il convient de or devrait; provide is prévoir; Member States shall ensure that is veillent à ce que; having regard to is vu; legal is juridique more often than légal. In the other direction arrêter is adopt or lay down, relever de is fall within, saisir a court is refer to (and a court is seised), intervenir is often be reached or occur.

English → French

<!-- entries domain=law-eu src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=law-eu src_lang=fr --> <!-- /entries -->

Economy and finance

Numbers and terms of account carry traps that cost money: billion is milliard, trillion is billion, fiscal policy is politique budgétaire, turnover is chiffre d'affaires or rotation depending on the collocate. The ECB corpus is counted but not quoted.

English → French

<!-- entries domain=economy src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=economy src_lang=fr --> <!-- /entries -->

Medicine

The EU package-leaflet template fixes many formulas (Tenir hors de la vue et de la portée des enfants, effets indésirables, à libération prolongée); EMEA is counted, not quoted, so most examples here come from Europarl, DGT and Tatoeba. Everyday medical words split by sense: medicine is médicament or médecine, drug is médicament or drogue, trial is essai or procès.

English → French

<!-- entries domain=medicine src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=medicine src_lang=fr --> <!-- /entries -->

IT and software

IT French borrows freely and the official coinages coexist with the borrowings (into-fr.md, pair manual section 18.1). The rows record which form the measured corpora use and which the official vocabulary recommends; neither is a rule by itself. The corpora are thin here: DGT has some IT procurement and standards text, Tatoeba some everyday computing.

English → French

<!-- entries domain=it src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=it src_lang=fr --> <!-- /entries -->

Science and academic prose

Academic English evidence, significant and matter each need a choice in French (données, preuves; significatif in statistics, important elsewhere; matière in physics, question in debate). French il ressort de, porter sur, expérience go the other way.

English → French

<!-- entries domain=academic src_lang=en --> <!-- /entries -->

French → English

<!-- entries domain=academic src_lang=fr --> <!-- /entries -->

Limits

Back to the translation manuals; the pair manual is English and French: the pair manual; the false friends of the pair are in False friends: English and French.

The Elamur team

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