Machine-translated Japanese: what to check by hand before anyone reads it

Say a small company needs its help pages in Japanese by the end of the week. The English goes through machine translation, the output looks clean, and someone who reads Japanese is asked to "give it a quick look". This piece is for that someone.

The difficulty with machine translation output is not that it is full of obvious mistakes. It is that it is fluent. Sentences end properly, particles sit where they should, and the text has the rhythm of real Japanese. The errors that matter hide inside that fluency: a negation that vanished, a polite form aimed at the wrong person, a term that changes from one page to the next. None of these jump out on a quick read. Each needs a deliberate pass.

The checks below are grouped by what goes wrong. Each has a short way to search for it and a before and after example.

Negation that went missing or changed scope

This is the error with the worst consequences, so it comes first. English negation is spread around the sentence (not, no, never, unless, without, fail to, prevent). Japanese negation usually sits at the end. Machine translation output sometimes drops it, moves it, or turns a partial negation into a total one.

Source: "Do not unplug the device until the light turns green."

Before: ランプが緑色になるまで、デバイスのプラグを抜いてください。 (Unplug the device until the light turns green.)

After: ランプが緑色に点灯するまで、電源プラグを抜かないでください。

Source: "Not all features are available on the free plan."

Before: 無料プランでは、すべての機能が利用できません。

Most readers will take this as "none of the features work on the free plan". The source means some do not.

After: 無料プランでは、一部の機能をご利用いただけません。

Source: "Your data will be deleted unless you renew."

Before: 更新すると、データは削除されます。 (If you renew, your data will be deleted.)

After: 更新しない場合、データは削除されます。

How to check: make a list of every negative in the English source, including the quiet ones like "unless", "without", "fail to" and "no later than". Then find each one in the Japanese. Look for ない, ません, ず, 禁止, and prefixes such as 不, 未 and 非. Any English negative without a Japanese partner is a finding until proved otherwise.

The wrong level of politeness

Machine translation tends to choose a politeness level segment by segment, and it does not always know who is speaking to whom.

Humble forms for the customer's actions

Kenjōgo (humble language) lowers the speaker's own actions. Applied to the reader, it is a real error, not a style point.

Source: "Please see the attached file."

Before: 添付ファイルを拝見してください。

After: 添付ファイルをご覧ください。

拝見する is the humble verb for the writer's own looking. For the reader, the respectful ご覧になる is needed.

Source: "When you submit the form, you will receive a confirmation email."

Before: フォームを送信させていただくと、確認メールが届きます。

After: フォームを送信すると、確認メールが届きます。

Mixed sentence endings

Because segments are often translated one at a time, a page can switch between です／ます and だ／である from one paragraph to the next. The checker flags mixed endings, and the legend of issue types explains when a mix is deliberate (quotes, headings, lists) and when it is a break.

あなた everywhere

English uses "you" constantly. Japanese user documentation mostly leaves the reader out.

Source: "You can change your password in Settings."

Before: あなたは設定であなたのパスワードを変更できます。

After: パスワードは［設定］から変更できます。

A page full of あなた reads as translated, and in customer-facing text it can sound cold or even rude. For a longer list of business politeness traps, the keigo email checklist covers double keigo and させていただく overuse.

Katakana that no one would write

Machine translation output is often too willing to turn English words into katakana.

Some of these loanwords are used in Japanese, and in some industries they are normal. The question is whether a Japanese writer of this kind of text would choose them. For a consumer help page, usually not.

Long vowel marks are the second katakana problem. The same document may give サーバ in one place and サーバー in another, or ユーザ and ユーザー. Neither form is wrong, but mixing them is. Pick the form the client already uses, or the one their style sheet sets, and search for both.

Also check proper nouns. A product name that should stay in Latin script may have been transliterated, or a word that should be translated may have been left in English.

Terms that drift

Machine translation does not remember that "account" was アカウント on page one, so on page four it may be 口座. "Sign in" may come out as サインイン, ログイン and ログオン in the same set of pages. "Cancel" may be キャンセル, 取り消し or 中止.

For interface text, the rule is stricter: a button label in the instructions must match the label on the screen exactly. If the Japanese interface says 保存, the help page cannot say セーブ.

How to check: pull 10 to 30 key terms from the English (product features, button labels, recurring nouns) and search the Japanese for every variant of each. A simple table of term, chosen Japanese form and rejected variants also becomes the start of a glossary for the next round.

Numbers, dates and units

Scale words cause the classic errors. English counts in thousands and millions, Japanese in 万 (10,000) and 億 (100,000,000).

• "1.5 million users" is 150万人, not 15万人.
• "2 billion yen" is 20億円.
• "250,000" is 25万.

Dates need the source's locale. "06/07/2025" is 7 June in American usage and 6 July in British usage, and the Japanese 2025年6月7日 or 2025年7月6日 will pick one with no hint that a choice was made. If the source gives a weekday, it settles the matter: 6月7日（土） is a Saturday in 2025.

Units and currencies are a brief question. Were miles, inches or Fahrenheit meant to be converted? Should dollar prices stay in dollars? Machine output leaves them as they are, which may or may not be what the client wants.

Digits themselves can come out in a mix of half-width and full-width forms (3件 next to ３件). The piece on full-width and half-width characters covers the cleanup.

English shapes inside Japanese sentences

Some output is correct but carries the structure of English.

• Pronouns. それは, これは, 彼, 彼女 appear where Japanese would leave the subject out or repeat the noun.
• Long modifiers. An English relative clause becomes a very long noun modifier before the noun. Split the sentence instead.
• Punctuation. Half-width commas and full stops, straight double quotes instead of 「」, and exclamation marks in calm instructional text.

Before: これは、あなたが作成したファイルを、他のユーザーと共有することができる機能です。

After: 作成したファイルをほかのユーザーと共有できます。

Brackets, spaces and Latin text

Help pages and interface text mix Japanese with Latin script all the time: product names, file types, Wi-Fi, PDF. Machine output handles each segment on its own, so the conventions wander.

Button and menu names are the clearest case. One page puts them in 「」, the next in ［］, a third leaves them bare or keeps the English in straight quotes. The verb wanders too: クリック in one step, 押す in the next, 選択 in a third, for the same action.

Before: "Save" ボタンをクリックし、「閉じる」を押してください。

After: ［保存］をクリックしてから、［閉じる］をクリックしてください。

Spacing around Latin words is a house style question. Some styles put a half-width space between Japanese and a Latin word (PDF ファイル), others write them together (PDFファイル). Either is fine on its own. A document that does both looks careless, and a search for one form will miss the other.

How to check: decide the bracket style, the verb for each action and the spacing rule before you start, write them at the top of your term list, and search for the forms you rejected.

The same checks in the other direction

Japanese to English machine output has the mirror image of these problems. The negation at the end of a Japanese sentence can get attached to the wrong part of the English.

このボタンを押さないと、変更は保存されません。

Before: "Pressing this button will not save the changes."

After: "Your changes will not be saved unless you press this button."

The output also has to invent subjects that Japanese left out, and it guesses. The routine in reviewing a Japanese to English translation works through those passes in detail.

An order of work that fits in an afternoon

1. Read the Japanese on its own first, without the English. Mark anything that sounds translated or unclear.
2. Do the negation sweep with your list of English negatives.
3. Do the term sweep with your list of 10 to 30 terms.
4. Check every number, date and unit against the source.
5. Read once more for politeness only: one voice, the right forms for the reader, no humble verbs aimed at the customer.
6. Run the text through the checker for mechanical issues such as width mixing, half-width katakana and mixed endings.
7. If anything is going to be printed, published or sent to many people, get a final read from someone whose first language is Japanese.

Step 6 comes late on purpose. A rule-based checker catches surface problems well and knows nothing about meaning. It cannot see that a negation went missing or that 拝見 points at the wrong person. Those are the checks that only a careful human reader can make, and they are the reason the "quick look" needs to be a slow one.