Help Needed to Improve Japanese Translation Quality

Aparna 25 Reputation points
2025-06-18T06:10:14.92+00:00

I'm using Azure Text Translator to translate English text into Japanese, but the results are not accurate. Could you please suggest how to improve the translation quality?

Here are some issues I'm facing:

Incorrect context in some translations

  1. Extra or unnecessary words being added - for example, translating S40 to S40の

Unnatural translation – for example, “LHD” remains the same, but “RHD” is translated as "右ハンドル" (full form)

Azure AI Translator
Azure AI Translator
An Azure service to easily conduct machine translation with a simple REST API call.
489 questions
{count} votes

Accepted answer
  1. Saideep Anchuri 9,500 Reputation points Moderator
    2025-06-21T17:09:40.4866667+00:00

    Hi Aparna

    using Azure AI Translator for initial translation and then verifying the output with a Large Language Model (LLM) like OpenAI, before storing it. This hybrid workflow is increasingly popular for multilingual applications where custom training data is limited.

    1. Translate with Azure AI Translator

    Use the Azure Translator Text API or Document Translation API to translate content. It supports over 130 languages and retains document structure for files like Word and PDF.

    • You can integrate it via REST API or SDK.
    • It supports language detection, transliteration, and custom glossaries.
    • For documents, use Azure Blob Storage to manage source and target files
    1. Verify with OpenAI

    Once translated, pass the original and translated text to an OpenAI model (e.g., GPT-4) and ask it to:

    • Evaluate translation quality.
    • Suggest corrections or improvements.
    • Flag cultural or contextual mismatches.
    1. Store Verified Output

    After verification:

    • Store the final translation in Azure Blob Storage or Cosmos DB.
    • Optionally, log the verification feedback for audit or retraining purposes.

    Kindly refer below link: document-translation-flow

    Thank You.

    2 people found this answer helpful.

1 additional answer

Sort by: Most helpful
  1. Alex Burlachenko 11,610 Reputation points
    2025-06-18T08:18:44.67+00:00

    hi Aparna, thanks for posting this question

    be sure u're using the latest api version. sometimes older versions have quirks. check the azure translator api docs https://docs.microsoft.com/en-us/azure/cognitive-services/translator/ for updates. also, try enabling the 'sentence length' and 'include alignment' options - they help with context accuracy.

    azure's custom translator feature https://docs.microsoft.com/en-us/azure/cognitive-services/translator/custom-translator/overview is gold one for fixing those industry-specific terms like 'LHD' and 'RHD'. u can train it with your own glossaries. yes it takes some setup but wow does it make a difference ))

    see those extra particles like 'の' popping up? classic machine translation issue. the fix? pre-process your text with simple regex to handle known patterns. something like

    text = text.replace(/(S\d+)の/g, '$1');

    if your source text has hidden formatting. tools like notepad++ can show invisible characters that mess with translations. this might help in other tools too )

    and about '右ハンドル' - that's actually correct japanese for RHD... but if u need the acronym, force it by adding 'RHD' to your custom dictionary. worth looking into how other localization teams handle automotive terms https://www.jat.org/glossary.php super useful resource!

    remember to break long sentences before translating. japanese has very different sentence structure so shorter inputs often give cleaner outputs. aha and don't forget to set the 'contentType' parameter to 'plain' if you're dealing with codes like S40....

    Let me know how at the end u solve that..

    rgds,

    Alex

    1 person found this answer helpful.
    0 comments No comments

Your answer

Answers can be marked as Accepted Answers by the question author, which helps users to know the answer solved the author's problem.