Topics

Migrating Machine Translation from Google Translate to Claude API

  • column

About two months ago, we switched our article translation from Google Translate to an LLM.

Why did we choose Claude? How does the system work? And why Claude over Gemini, which is less expensive?

The experiment is still ongoing, but here is a summary of the current setup and our thinking behind it.

Google Translate can, of course, translate as well.

To be clear from the start, Google Translate did not fail us.

It is fast, and for everyday use it is more than sufficient. We had no major complaints so far, but when translating entire articles, simply swapping Japanese for another language felt far short of what current AI can deliver in terms of accuracy.

Read sentence by sentence, the output is not wrong. But when read as a whole article, it can feel slightly unnatural, and the same term may be translated differently partway through. Headings and copy that rely on letter-by-letter typographic design could also break entirely.

It's a must that the translation doesn't just convey the meaning, but also reads naturally as an article, right?

So, a bit late to the party, but I've finally switched to machine translation using an LLM.

Last year, we expanded into multiple languages using Google Translate, and our DR dropped from 30 to 19.

Translation quality isn't the only reason we switched to an LLM.

When we expanded the article into multiple languages using Google Translate, we made a number of changes to the SEO setup as well, but because we added eight languages all at once, link equity was severely diluted.

Ahrefs showed our DR falling from 30 to 19. 😨

That's a drop of 11 points.

It fell quite dramatically.

Of course, it's not as simple as saying the DR dropped just because we used Google Translate. The real issue is the SEO structure of the multilingual site itself, including URLs and internal links. We urgently need to rethink the entire multilingual setup, including the translation method and how translated articles are managed, right!

DR should be back above 30 now, but since we plan to run some more experiments soon, I expect it'll drop again. w

Going up and down...

After comparing them, Claude came out on top for our articles.

When we switched to LLM-based translation, we looked into several models and had each one translate actual articles.

We mainly focused on the following points.

  • Whether the translated text reads naturally
  • Whether it can maintain context in long articles
  • Whether it can keep terminology and notation consistent
  • Whether it understands the relationship between headings and body text
  • Whether it follows the specified translation rules
  • Whether it avoids adding content that isn't in the original

In our comparison, Claude produced the best translations for our articles.

This does not mean Claude is the best translation model across the board.

These results come from testing with our own articles and evaluation criteria only. The outcome may vary depending on the type and length of the text being translated and the writing style required.

We chose the Claude API this time, prioritizing natural-sounding article text.

If cost alone is the concern, Gemini is cheaper.

Claude delivered better translation quality, while Gemini is cheaper if cost is the only factor.

We have not committed to using Claude indefinitely. For now, we will start operations with Claude, which delivered the best translation quality!

If Gemini can meet the required quality, we will switch to it later!
Alternatively, we are considering using multiple models depending on the type of article.

That is the kind of setup we are considering.

LLMs change quickly, so the model that performs best today may not be the best six months from now.
Both model performance and pricing can change.

Use the best option for now, and switch when circumstances change.

I think that kind of distance is a good approach for using AI these days!

Call the Claude API only for articles that haven't been translated yet

The Claude API is not called every time an article is displayed.

Translated versions of articles are cached at build time, and saved translations are returned from the next request onward. The Claude API is only called for articles that have not yet been translated.

If articles were translated every time they were accessed, API fees would be incurred each time. Load times would also increase.

Translate articles only when they are needed, and reuse each translation once it has been made.

This approach keeps both API usage costs and display speed under reasonable control.

Retranslate articles when they are updated

If translation results are cached, you also need to handle cases where the original article is updated.

It would be a problem if the Japanese article was corrected but the translated version still showed the old content.

For that reason, when the source text changes, we don't use the saved translation. Instead, we retranslate it using the Claude API.

We're still fine-tuning these details as we use the system in practice.

Summary

So far, we're happy with our choice of Claude for translation quality. That said, depending on cost and accuracy, we may certainly switch to models such as ChatGPT or Gemini at some point.

The experiments will continue! We'll cover the details around SEO in a separate article!

About the author

Though I'm the president, I'm always a hands-on counterpart. I take joy in understanding new technologies and in those moments when something becomes more convenient, and I love being completely absorbed in the work on the ground. I get excited about future technologies and want to keep enjoying new experiences no matter how old I get.

Tadanori Morimoto

Project Manager / Director / Founded 2007

View this staff member's articles

Reliable team structure and responsive support are our strengths

Liberogic's experienced staff actively drive projects forward, which has earned high praise from our clients.
We make sure to properly assign project managers and directors to keep the entire project running smoothly. Our style of allocating resources to the right people, rather than adding full-time commitment unnecessarily, helps prevent cost increases. We are also known for our speed, from understanding the scope of work to preparing and submitting estimates.

Note: Please be advised that we do not actively engage in SES-style on-site staffing assignments.

We support nearly all major project management and chat tools, including Slack, Teams, Redmine, Backlog, Asana, Jira, Notion, Google Workspace, Zoom, and Webex.

Have a web challenge? Let us know.

Case Studies