Weak machine translations: fix the drafts Weblate's automatic translation leaves, and stop new ones

tiagoagueda commented

2026-10-15 16:56:02 +00:00

Owner

Many catalogues hold weak machine drafts. Fix them in Weblate, and stop new ones from piling up.

Where they come from

Every Weblate component runs the Automatic translation add-on: LibreTranslate, query state:empty, mode fuzzy, threshold 80. Each new source string gets a LibreTranslate draft in every language. Many of those drafts are poor:

placeholders translated, dropped or broken (%(compte)s, {compte}, SECRET KEY);
literal or plainly wrong meanings ("Trigo de pão" for Breadcrumb, «Prométhée» for Prometheus, «Ce vaisseau» for This ships inside Postulo);
the wrong variety or register (Brazilian forms in pt-PT, «tu» in French);
repetition loops in zh-Hant (相關相關相關…);
leftover markup such as [X] and [hariri chanzo].

Drafts count for the release gate (#347), so none of this blocks a build, but users see it.

How big it is (2026-10-15, after #893 merged)

Left out of the compiled catalogues. messages.py check excludes these drafts from the .mo files, so users see English for them:

1,684 in total: 1,151 whose placeholders differ from the source, and 533 that drop a count other numbers need.
By language:

Language	Excluded drafts
sw	528
ur	322
ar	140
sq	129
eu	100
ca	71
tr	60
zh-Hant	50
sk	40
ro	31
bg	26
zh-Hans	24
ga	23
cs	17
nb	16
fi	14
el	14
others	under 10 each

Compiled but poor. A sample audit (150 strings per language, 2026-10-13) found these shares of drafts wrong in meaning:

11%: ca, lv
7%: et, eu
5%: bg, de, it
4%: da
3%: cy, es

Never reviewed at all:

zh-Hans: all 4,734 strings are machine drafts.
zh-Hant: about 3,650 machine drafts. Its empty strings and 289 broken repetition loops were already rewritten.
sr-Cyrl: 2,113 older machine drafts.
The other 0.4.0 languages: everything except about 140 strings each.

Already done, outside this issue:

Reviewed and corrected: pt-PT, fr-FR, fr-CA, es-MX, gl, id, hu and lt, plus the 2026-10-13/15 fill.
Rewritten from scratch as drafts: bn, ar and hi.
Written as hand drafts, needing native review: the seven new languages (az, be, kk, cnr, nn, rm, fo).

What to do

Stop the inflow. Pick one:
- turn the add-on off;
- raise its threshold so it only copies translation memory;
- keep LibreTranslate for languages with no translator, but add a check so a draft that changes placeholders is never saved.
Today a draft that breaks a placeholder is saved as needing editing, and nothing flags it.
Fix the excluded drafts first. Do every language in the table: restore the placeholder set and the count, or rewrite the string. These are what put English on screen.
Review the 0.5.0 languages, worst first. Order by the audit: ca, lv, et, eu, then the rest. Use the same glossary-first method as the 2026-10-13 passes: a binding term sheet per language, then batches that check against it. Keep each result as needing review until a native speaker approves it.
zh-Hans and zh-Hant. Retranslate from the zh-Hant glossary (Taiwan usage) and a new zh-Hans one. Patching these drafts isn't enough: they mix scripts and get core terms wrong (應用程式 for a job application, 上市 for a listing).
Add a guard. A check (messages.py check, or a Weblate quality check) that reports per language how many drafts are excluded, so the number can only go down.

Not part of this issue

These are filed or tracked elsewhere:

Shared source strings that need a context. "Apply" means both applying a filter and applying for a job; "From", "Other" and "Sent" are also shared. Translators can't fix these, so they need a separate core issue for pgettext.
The "Industry" retranslation in fr, pt and es after #897. Left to the maintainer.

Many catalogues hold weak machine drafts. Fix them in Weblate, and stop new ones from piling up. ## Where they come from Every Weblate component runs the **Automatic translation** add-on: LibreTranslate, query `state:empty`, mode *fuzzy*, threshold 80. Each new source string gets a LibreTranslate draft in every language. Many of those drafts are poor: - placeholders translated, dropped or broken (`%(compte)s`, `{compte}`, `SECRET KEY`); - literal or plainly wrong meanings ("Trigo de pão" for *Breadcrumb*, «Prométhée» for *Prometheus*, «Ce vaisseau» for *This ships inside Postulo*); - the wrong variety or register (Brazilian forms in pt-PT, «tu» in French); - repetition loops in zh-Hant (相關相關相關…); - leftover markup such as `[X]` and `[hariri chanzo]`. Drafts count for the release gate (#347), so none of this blocks a build, but users see it. ## How big it is (2026-10-15, after #893 merged) **Left out of the compiled catalogues.** `messages.py check` excludes these drafts from the `.mo` files, so users see English for them: - 1,684 in total: 1,151 whose placeholders differ from the source, and 533 that drop a count other numbers need. - By language: | Language | Excluded drafts | |---|---| | sw | 528 | | ur | 322 | | ar | 140 | | sq | 129 | | eu | 100 | | ca | 71 | | tr | 60 | | zh-Hant | 50 | | sk | 40 | | ro | 31 | | bg | 26 | | zh-Hans | 24 | | ga | 23 | | cs | 17 | | nb | 16 | | fi | 14 | | el | 14 | | others | under 10 each | **Compiled but poor.** A sample audit (150 strings per language, 2026-10-13) found these shares of drafts wrong in meaning: - 11%: ca, lv - 7%: et, eu - 5%: bg, de, it - 4%: da - 3%: cy, es **Never reviewed at all:** - zh-Hans: all 4,734 strings are machine drafts. - zh-Hant: about 3,650 machine drafts. Its empty strings and 289 broken repetition loops were already rewritten. - sr-Cyrl: 2,113 older machine drafts. - The other 0.4.0 languages: everything except about 140 strings each. **Already done, outside this issue:** - Reviewed and corrected: pt-PT, fr-FR, fr-CA, es-MX, gl, id, hu and lt, plus the 2026-10-13/15 fill. - Rewritten from scratch as drafts: bn, ar and hi. - Written as hand drafts, needing native review: the seven new languages (az, be, kk, cnr, nn, rm, fo). ## What to do 1. **Stop the inflow.** Pick one: - turn the add-on off; - raise its threshold so it only copies translation memory; - keep LibreTranslate for languages with no translator, but add a check so a draft that changes placeholders is never saved. Today a draft that breaks a placeholder is saved as needing editing, and nothing flags it. 2. **Fix the excluded drafts first.** Do every language in the table: restore the placeholder set and the count, or rewrite the string. These are what put English on screen. 3. **Review the 0.5.0 languages, worst first.** Order by the audit: ca, lv, et, eu, then the rest. Use the same glossary-first method as the 2026-10-13 passes: a binding term sheet per language, then batches that check against it. Keep each result as needing review until a native speaker approves it. 4. **zh-Hans and zh-Hant.** Retranslate from the zh-Hant glossary (Taiwan usage) and a new zh-Hans one. Patching these drafts isn't enough: they mix scripts and get core terms wrong (應用程式 for a job application, 上市 for a listing). 5. **Add a guard.** A check (`messages.py check`, or a Weblate quality check) that reports per language how many drafts are excluded, so the number can only go down. ## Not part of this issue These are filed or tracked elsewhere: - **Shared source strings that need a context.** "Apply" means both applying a filter and applying for a job; "From", "Other" and "Sent" are also shared. Translators can't fix these, so they need a separate core issue for `pgettext`. - **The "Industry" retranslation in fr, pt and es after #897.** Left to the maintainer.

tiagoagueda added the

internationalisation

label

2026-10-15 16:56:07 +00:00

tiagoagueda referenced this issue

2026-10-17 10:39:14 +00:00

SEC2-041: Markup inside translated text is trusted everywhere and no check keeps a translation's tags to the source's #983

tiagoagueda added this to the 0.5.0 milestone

2026-10-25 17:04:12 +00:00

tiagoagueda referenced this issue from a commit

2026-10-28 15:11:16 +00:00

Hold the drafts left out of the catalogues to a per-language ratchet (#910)

tiagoagueda referenced this issue

2026-10-28 15:11:18 +00:00

Hold the drafts left out of the catalogues to a per-language ratchet (#910) #1076

tiagoagueda referenced this issue from a commit

2026-10-29 10:23:48 +00:00

Hold the drafts left out of the catalogues to a per-language ratchet (#910)

tiagoagueda referenced this issue from a commit

2026-10-29 18:19:13 +00:00

Hold the drafts left out of the catalogues to a per-language ratchet (#910)

tiagoagueda referenced this issue from a commit

2026-10-29 18:20:17 +00:00

Hold the drafts left out of the catalogues to a per-language ratchet (#910) (#1076)

tiagoagueda referenced this issue

2026-10-31 12:27:59 +00:00

Merge Weblate's translations (#1058) #1175

tiagoagueda modified the milestone from 0.5.0 to 0.6.0

2026-10-31 15:55:19 +00:00

tiagoagueda commented

2026-10-31 15:55:19 +00:00

Author

Owner

Moved to 0.6.0 at the maintainer's word (2026-10-31), to release 0.5.0 today. The remaining machine drafts are Weblate review work; the 0.5.0 release gate counts drafts, so they do not block it.

Rows
Columns

Weak machine translations: fix the drafts Weblate's automatic translation leaves, and stop new ones #910

Where they come from

How big it is (2026-10-15, after #893 merged)

What to do

Not part of this issue