Machine translation

phil/translation-bundle

Machine translation proposes targets for the untranslated segments of a document. It is off by default, and its targets are never more than proposals: each is suggested, marked as machine output on the screen, never learnt by the memory until a person translates or reviews it, and credits nobody.

The port

A machine translation service is reached through an adapter named after it, which implements Phil\TranslationBundle\Application\Contract\MachineTranslator:

Method	Contract
`name(): string`	The name of the service, shown in errors
`translate(list<string> $sources, LanguageTag $sourceLanguage, LanguageTag $targetLanguage): list<string>`	One target per source, inline HTML, in the same order

The bundle ships no adapter and no SDK: the application writes the adapter of the service it chose, on Symfony's HTTP client, and names it in the configuration.

phil_translation:
    machine_translation:
        enabled: true
        translator: App\Translation\SomeServiceTranslator

enabled: true without a translator stops the container with a message; a translator without enabled: true stays off.

Translating

$count = $translations->machineTranslate('article:42', LanguageTag::of('fr'));

The segments of the prepared document that are untranslated are sent by batches of 50 (Translations::MACHINE_BATCH). Each target with words, in valid UTF-8, is proposed — suggested — and the row marked as machine output (SegmentRecord::isMachine()); a target without words is left out. The method gives how many segments were proposed a target.

MachineTranslationFailed is thrown when machine translation is off, when the translator throws — its exception kept as the previous one —, or when it does not answer one target per source; nothing is written then. NotFound is thrown for a document not prepared in the language. canMachineTranslate() tells whether a translator is configured.

On the screen

When a translator is configured, the page shows a button that proposes machine translations for the untranslated segments (POST /{document}/{language}/machine), then reports how many were proposed. Each target that came from the machine and that nobody has rewritten carries the mark machine translation. A failure of the service answers 502 with a message of its own, the same whatever the service answered: what an adapter's exception says — an address, a quota, a key — stays out of the page, and stays in the MachineTranslationFailed it chains, for the application's logs. Without a translator, the route answers 404.