🗣️ False friends: English, French, Ukrainian

This folder holds Mova's lists and manuals of false friends for the three pairs of the translation manuals, in both directions: False friends: English and French, English and Ukrainian: false friends and French and Ukrainian: false friends, with the machine-readable lists en-fr.tsv, en-uk.tsv and fr-uk.tsv. The general rules are in general.md §5; the companion lists of context-dependent renderings say which of several valid renderings fits which context, and a partial false friend usually has an entry there too.

1. What a false friend is

A false friend is a word (or a construction) that looks or sounds like a word of the other language and leads a translator to the wrong meaning, the wrong register or the wrong grammar. The look-alike is usually a cognate that drifted, a borrowing that narrowed or widened, or a chance coincidence. The kinds used in the lists (kind column):

Kinds combine with + (partial+register). Two channels matter for Ukrainian in particular:

• Through Russian. Many Ukrainian translators and readers also know Russian, and Russian words with a Ukrainian look-alike pull the meaning (ru неделя "week" against uk неділя "Sunday"), and Russian constructions pull the phrasing (мати місце, приймати участь, слідуючий). The lists mark these with the domain russian and judge them descriptively: what the norm and the editors' practice choose in neutral and official prose, and where a criticised form is right, in a character's speech or a song (into-uk.md §3.6 and §4). Surzhyk items (domain surzhyk) are a variety with regularities, not a list of errors; Polish (domain polish) matters for western Ukrainian and diaspora usage, where Polish-channel words differ from the standard.
• Through Latin and French. English and Ukrainian both took Latin and French words, Ukrainian often through Polish, German or Russian, and the meanings drifted on both sides (en ambulance / uk амбулаторія, fr cabinet / uk кабінет, en billion / fr billion). These have the domain latin-french in the lists with Ukrainian.

The research literature treats these words as a lexicon to build and to score. Cognates and false friends between French and English can be separated by orthographic similarity measures combined in a classifier, and partial cognates disambiguated in their sentence (frunza-2007-tool-detecting-french-english-cognates); false-friend pairs can be extracted from parallel bi-texts (nakov-2009-unsupervised-extraction-false-friends-parallel) or from monolingual corpora with a small bilingual dictionary, with a graded measure of how "false" a pair is (uban-2020-automatically-building-multilingual-lexicon-false); meaning shifts between related languages can be found as outliers among true cognates (uban-2025-friend-or-foe-computational-investigation); and parallel sentences tell interlingual homographs from cognates (han-2022-automating-interlingual-homograph-recognition-parallel). Our lists are not built by those methods: they are written by hand and measured in the corpora, as below; the methods are where an automatic extension of the lists would start.

2. Checking a false friend by measured usage

A list says what a translator should do; the corpora show what translators did. Each list row carries three patterns: the source word (src_match), the recommended renderings (good_match) and the look-alike (bad_match). The tool parallel choices reads every sentence pair of the pair's corpora once and counts, per row:

• n_src: sentence pairs whose source side contains the word;
• n_good: of those, the pairs whose target side contains a recommended rendering;
• n_bad: of those, the pairs whose target side contains the look-alike;
• by_corpus: the same three numbers per corpus (tatoeba=373/301/4;europarl-v8=…).

How to read them:

1. Co-occurrence is not translation. A look-alike in the same sentence pair is often an unrelated word (general manual §5: English actually and French actuellement in one Europarl sentence are almost never the error). A high n_bad is a reason to read the rows, not a verdict. Word alignment would separate the cases; the tool does not align.
2. A false friend is confirmed when the recommended renderings dominate and the look-alike is rare (library → bibliothèque, decade → décennie). A partial false friend shows up as a large n_bad where the shared sense is common (control → contrôle in EU French), and then the entry names the senses that match.
3. A false enemy shows up as the look-alike being the majority rendering (sensitive → sensible).
4. The corpora have registers. Europarl and DGT are EU institutional French and English; Tatoeba is short everyday sentences, partly translated through a third language; the Bible pairs are two translations of a third text, the Kulish/Puluj Ukrainian Bible of 1905 in its own spelling. A count speaks for its corpus. DGT repeats many segments, which inflates its counts.
5. Two runs, same bytes. The tool reads the corpora in the given order and breaks ties by that order, so a rerun gives identical files.

The rows are sorted by n_src (column frequency_rank: 1 is the source word met most often in the pair's corpora), so the traps a translator meets most come first; the manuals keep that order inside each domain.

Examples. The tool picks, for each row, the shortest quotable sentence pair (25–220 characters) that contains the source word and a recommended rendering but not the look-alike, from the corpora that may be quoted (Tatoeba rows with a known author, Europarl, DGT, the Gutenberg pairs with identified translators, the public-domain Bibles); learn/test-only corpora (EMEA, JRC-Acquis, ECB) are counted, never quoted. A given example_id is verified instead: it must exist and match (example_ok = yes). Every picked example was then read by eye, and a pick where the rendering does not translate the word in the entry's sense was replaced or removed; each manual's Data and method section says how many. example_status is corpus for a verified id, constructed for our own translated sentence (below), needs_example otherwise.

3. Examples we make

Many entries have no corpus example: the Ukrainian corpora are small, and legal, medical and IT words are rare in Tatoeba. For those, and for popular words where a better example teaches more, we write our own pairs with parallel tr-run (crate README):

1. the Codex teacher writes, per queued entry, one source sentence of its own for the entry's sense and context, its translation, the rendering used and a one-line note, through mova-teacher (mova-teacher — the shared Codex teacher transport): Terra for difficulty = plain, Sol for hard (hard:<reason>) entries;
2. gates: schema, script, one sentence, length bounds, the source contains the headword forms, the translation contains a recommended rendering and not the look-alike; Mova's parse of the English and French sides (findings, never rejections); for the hard entries, a second Terra call that back-translates and judges the pair;
3. accepted pairs get ids ours-tr:<pair>:<n>, are stored in <pair>.ours.tsv beside the list (our own data, Elamur Cosmic License) and become the entry's example (example_status = constructed); parallel choices-md renders them.

Run of 2026-11-06 (queue of 740 entries; teachers gpt-6.1-terra and gpt-6.1-sol, medium effort): 711 pairs accepted, 29 rejected, none left in the queue. 435 plain entries by Terra (424 accepted), 305 hard by Sol (287 accepted); 57 calls (23 Terra and 17 Sol generation calls, 17 Terra verification calls). Rejections are mostly gate patterns that did not list the inflected form or the rendering the teacher chose (19), then the reviewer call (10). A rejected entry still has example_status = needs_example and stays in the queue. Every pair of the two pilot batches (40 entries) was read by eye, and a sample of the rest; the other pairs passed the gates and, for the hard ones, the second call, but were not all read. The reviewer is a weaker model than the writer for Sol rows, so a pair that is subtly wrong in a hard case can pass.

4. How Mova uses the lists

• Today: the lists are data and documentation. The manuals render them; the counts show where translators go wrong and where a look-alike is legitimate.
• Planned: a lexical check in the Eel's translation rounds tr-<src>-<tgt> and in Mova's post-editing checks: when the source contains a listed word and the target contains its look-alike, the check flags it with the entry (kind, senses that match, recommended renderings), weighted by the entry's measured n_bad / n_src and by the sense the parse and the frames give the source word. Partial false friends need the sense first, which is why they point to the context lists. A check that flags must also be shown to stay silent on the senses that match (a negative control: contrôle des frontières must not be flagged).
• The Nightingales (one per language, naturalness models from Wikipedia) will judge the frequency and register entries in the target language alone.

5. Files and columns

List columns: src_lang, src_lemma, tgt_lang, tgt_lemma_lookalike, kind, senses_that_match (- for none), right_renderings (form (sense) items separated by ;), example_id, note (the trap, with a constructed ✗ rendering and the ✓ one), then domain, src_match, good_match, bad_match (patterns: | between alternatives, spaces between consecutive tokens, * for a prefix), difficulty, and the measured columns n_src, n_good, n_bad, by_corpus, example_ok, frequency_rank, example_status.

Domains: everyday, academic, law-eu, medicine, it, politics, economy, latin-french, grammar, and for Ukrainian russian, surzhyk, polish.

Rebuild (counts dated 2026-11-06; corpora in parallel):

parallel choices en-fr.tsv new.tsv new.examples.tsv en fr \
  $MOVA_DATA/corpora/parallel/en-fr/{tatoeba/pairs.tsv,europarl/pairs.tsv,dgt/pairs.tsv,gutenberg/pairs-good-build.tsv,bible/pairs-web-lsg.tsv} \
  $MOVA_DATA/corpora/parallel/en-fr/{emea,jrc-acquis,ecb}/pairs.tsv@count
parallel choices-md new.tsv new.examples.tsv en-fr.md new.md 2026-11-06
# en-uk: tatoeba + bible/pairs-web-kulish.tsv; fr-uk: tatoeba + bible/pairs-lsg-kulish.tsv

Counts (2026-11-06).

Every example id was checked again on 2026-11-06 by a separate scan of the corpus files, outside the tool: all exist, with the same two texts. The difficulty marks were set partly by rule from the kind and the domain (every Russian-channel, surzhyk and Polish entry is hard:register, for instance) and are a first sorting for the planned teachers, not a judgement entry by entry.

6. Limits

• Counts are co-occurrence in sentence pairs, not word alignment; the patterns are hand-written and can over- or under-match (a prefix catches a related word, an inflected form is missed).
• The English–Ukrainian and French–Ukrainian corpora are small and everyday (Tatoeba) or old and religious (the 1905 Bible): many Ukrainian entries have no measurement, and the Russian-channel traps cannot be measured in them at all.
• The lists cover the traps we judged most useful; they are not complete, and their judgements on Ukrainian russianisms and surzhyk follow the folder's stance (usage, register and house style decide), not a purist list.
• The planned lexical check does not exist yet. Our constructed examples are model-written and model-translated; the hard ones were reviewed only by a second model call.

The Elamur team

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