Collect translations for measurement

Choose a supported source and preserve the text, revisions, and context a measurement needs.

Measurement can begin with translations your team already produces. Keep the repository or translation-management system as the place that maintains accepted text, and choose how to make that text available to a check.

The useful unit is an identifiable version. If a message changes after evaluation, the earlier answer still describes the earlier wording. Keeping source and target revisions together lets the team return to the evidence rather than reconstruct it from the latest content.

Choose the source for your workflow #

Starting pointHow to measure it
Crowdin activityConnect Crowdin, configure an event measurement, and inspect the resulting Radar event and answer. Radar access is required for its screens.
GitHub commitsConnect the GitHub App and configure source and target paths. Changes to supported localization keys on the default branch become events.
Zendesk Help CenterConnect Zendesk. Published article changes become source or translation events in their respective languages.
Local translation filesUse lingo score with a configured scorer and model. Results are written locally, outside Radar event history.
Lingo.dev translationsEach completed localization is already an event. Review Lingo.dev translations explains how to select engines, languages, and a sample.

For another source, contact Lingo.dev to arrange event collection before automating delivery.

Collect GitHub and Zendesk activity #

For GitHub, install the App for the intended repository and connect it to the Lingo.dev organization. Commit .lingo/config.json with the source locale, target locales, and source-file patterns. Those paths identify which changes belong to source text and which belong to translations. Inspect an event after a default-branch commit changes a configured localization key; one commit can produce events for several keys and languages.

GitHub collection reads JSON, JSONC, YAML, YAML with a root key, Android resources, Apple .strings, Flutter ARB, PHP, and TypeScript files. Repository changes explains how changes become events.

For Zendesk, authorize the Help Center connection and enable its languages. Radar records published article activity by language. Open the event and compare the article identity, locale, and text with Zendesk. Collecting activity does not itself trigger a translation Sync or publish a new target article.

Both connections keep accepted translations in their original system. A measurement examines their recorded events without becoming the publication step.

Import provider history before measuring it #

Choose the provider history you need:

  • GitHub: imports up to the latest 5,000 commits on the repository's default branch. Each commit is read with the .lingo/config.json saved at that revision. Commits before that configuration existed produce no events.
  • Zendesk: imports the current published versions of existing articles, sections, and categories in each language, dated by their last update. Earlier revisions and draft versions are outside this collection.

New GitHub and Zendesk connections start a history import automatically. Follow its progress in the connection's Settings tab. To start or retry an import later:

  1. In Studio, select the repository or Help Center under Connected, then open its Settings tab.
  2. Under Commit history for GitHub or Help Center history for Zendesk, select Import history.
  3. Follow the import's progress, which counts provider items queued for collection.
  4. Wait until the events you intend to measure are visible in Radar → Logs, then inspect their dates, locales, and text.

For a measurement that runs on provider events, its Backfill tab also offers Import events, with connected GitHub repositories and Zendesk Help Centers. Imported history is collected without running measurements. Once the desired events are visible, enable the measurement and start a historical evaluation for their interval.

If more imported events appear after the backfill, run it again with Only events not measured yet enabled. If the import fails, inspect its progress and retry from the same Settings tab. Check the measurement's backfill and results separately from the import.

Preserve the comparison #

Keep the source-system ID, content ID, target locale, and target revision together. For a meaning check, add the corresponding source revision. For an edit measurement, retain the draft from which the edited target was produced.

For example, in your own tracking record, delivery.promise, locale fr, revision 42 identifies a particular catalog entry. Revision 43 may change the promise.

Record the producer when the workflow supplies it: a human, model, machine-translation provider, or translation-memory reuse. Preserve an unknown origin as unknown. A provider user who approved a translation is not necessarily the person or model that wrote it.

Supply evidence for the chosen question #

The criterion determines what to collect. A spelling check can examine target text alone. A meaning check needs paired source and target, with context for ambiguity. A terminology check needs the applicable term and its usage boundaries. A placeholder check needs the expected placeholder contract, while edit distance needs the actual draft-target pair.

If a translation is missing its source revision, keep it among the unassessed cases for a meaning check and obtain the source before judging it.

Verify one collected event #

  1. Configure the selected measurement before producing test activity in the connected source.
  2. Find the corresponding event. Compare its locale, text, provenance, and available draft with your own record.
  3. Open the measurement answer and read the criterion beside the text. Check whether it returned an answer, returned N/A, or failed to complete.
  4. Keep the event and measurement identities with the evaluated revisions. Make any accepted correction in the system that owns the text.

For local scoring, inspect the saved result beside the input files and scorer configuration instead; that path does not create a Radar event.

For a repeated workflow, retain what was selected, collected, and measured. Measure translations from your pipeline covers collection and recovery, including retries and changed revisions.