Using AI to Sync Japanese-to-English Translations: The Approach and the Patterns That Did Not Work
Introduction
This site treats Japanese as the source of truth, with English versions managed as translations. As the number of articles grew, managing translations manually was no longer practical, so I built a translation sync system using AI. This article covers how the system works and the problems that arose, along with how I addressed them.
How the Translation Sync Works
The current translation sync follows this order.
- Create and finalize the Japanese version of the article (
src/content/blog/ja/) - Translate it into English using the i18n-sync skill (
src/content/blog/en/) - Review the translation and fix any problems
- Publish after confirming the quality of both the Japanese and English versions
The sequence of “finalize Japanese first, then translate into English” is important for maintaining Japanese as the source of truth. I did not adopt the reverse approach of writing in English first and translating into Japanese afterward, because content drift was more likely to occur that way.
Patterns That Did Not Work
Three problems came up repeatedly during operation.
Over-Interpretation Caused the Nuance of the Original to Disappear
A section I wrote in Japanese as “points to be aware of” was sometimes translated into English as “pitfalls to avoid.” The meaning is close, but “pitfalls” is a figurative danger expression that this site’s style policy avoids. In addition, the Japanese expression “〜と考えられます” (which conveys that something is the author’s interpretation, not a definitive claim) was sometimes translated as a definitive “this is.” The intent I expressed in Japanese — marking the reasoning as my interpretation — disappeared from the English version.
To address this, I added to the translation instructions: “Maintain a form close to the original rather than interpreting freely” and “Preserve expressions that indicate the author’s speculation or interpretation (such as might, I believe, in my view) in the corresponding places.”
Changes to the Heading Structure Changed the Flow of the Article
In one article, where I had three H2 headings in Japanese — ”## Background of the problem,” ”## How I handled it,” and ”## Result” — the English version consolidated them into a single H2: ”## Background and Resolution.”
When the heading structure changes, the logical flow of the article changes. “Background,” “response,” and “result” are each sections with distinct purposes, and consolidating them removes the signposts that help readers follow the information.
To address this, I added an explicit rule to the translation guidelines: “Match the heading levels and order to the Japanese version. Do not consolidate, omit, or add sections.”
Terminology Was Translated Inconsistently Across Articles
The site-specific term “harness” was translated as “framework” in one article, “scaffold” in another, and left as “harness” in yet another.
When the same concept is translated differently across articles, it becomes harder for readers to understand terminology consistently across the site. For search purposes, inconsistent translations of the same concept also make it harder to find related articles.
To address this, I created a list of terms excluded from translation. Site-specific terms such as “harness,” “frontmatter,” “SKILL.md,” and “CLAUDE.md” are now maintained as-is in the English version, even when they appear in katakana in the Japanese version.
How I Now Confirm Translation Quality
After translation, I check the following.
- Whether the heading levels and order match the Japanese version
- Whether expressions indicating the author’s speculation or interpretation are preserved in the corresponding places
- Whether site-specific terminology is used consistently
- Whether “we,” “our,” or “us” is being used as the first person for the author in the English version (the correct form is “I”)
Summary
When using AI for Japanese-to-English translation sync, three problems tend to arise: over-interpretation, changes to heading structure, and inconsistent terminology. Making explicit rules in the translation guidelines — prioritize close translation over free interpretation, maintain heading structure, and exclude site-specific terms from translation — helps reduce these problems.