IME conversion errors: the henkan slips that spellcheck never sees

Nobody types Japanese directly. You type sounds, either as rōmaji or on a kana keyboard, and an input method editor (IME) turns those sounds into kanji and kana. That step is called henkan (変換, "conversion"), and it is where a whole family of errors is born. Type いがい and the IME has to guess whether you meant 以外 (igai, "other than") or 意外 (igai, "unexpected"). Type ほしょう and it has three common candidates, all real words, all pronounced the same.

The guess is often right. When it is wrong, the result is a correctly spelled word in the wrong place. Spellcheck sees a valid dictionary word and moves on. The sentence reads smoothly at speed, because your eye recognises the shape it expected. These are the slips that survive two read-throughs and turn up in the printed brochure.

This piece looks at where conversion errors come from, which pairs cause the most trouble, and a routine that catches them more reliably than "read it again carefully".

Why spellcheck has nothing to say

An English spellchecker works by looking each word up. "Recieve" is not in the dictionary, so it gets a red line. A Japanese conversion slip never produces a non-word: the IME only ever offers words from its dictionary. 意外の点 is wrong in context but every character in it is correct, and 意外 is a perfectly good word.

Grammar-aware tools can catch some cases, for example when the wrong word breaks a fixed collocation. Most of the time, though, telling 保証 from 保障 depends on what the sentence is about, and that is a question of meaning, not spelling. A person has to read it.

Japanese has two kinds of trap here, and it helps to keep them apart:

• 同音異義語 (dōon igigo): different words, usually Sino-Japanese compounds, that share a reading. 機会 (kikai, "opportunity") and 機械 (kikai, "machine").
• 異字同訓 (iji dōkun): one native Japanese word that can be written with different kanji depending on the sense. はやい is 早い for time ("early") and 速い for speed ("fast").

The first kind is a choice between two words. The second is a choice of spelling for one word, and the line between the senses can be fine. The Japanese language subdivision of the Council for Cultural Affairs, the advisory body of the Agency for Cultural Affairs (文化庁), has published a report of usage examples for these pairs (「異字同訓」の漢字の使い分け例), and many style guides follow it.

The pairs that cause the most trouble

These are worth knowing by heart, because they come up in ordinary workplace and web writing all the time.

Notice that in the business pairs, both candidates are words a reader of a work email expects to see. 在庫のご紹介 and 在庫のご照会 are both plausible phrases, and only the situation tells you which one the writer meant. That is exactly why no list can replace reading for sense.

The 異字同訓 group adds verbs and adjectives: 暑い / 熱い / 厚い (atsui: hot weather, hot to the touch, thick), 会う / 合う (au: meet a person, fit or match), 計る / 測る / 量る (hakaru: time and numbers, length and depth, weight and volume), 変える / 代える / 替える / 換える (kaeru: change, substitute, replace, exchange). Some of these depend on the house style as much as on the dictionary, so check what your client or publication prefers before you mark them.

Other ways conversion goes wrong

Homophones are the famous case, but the IME has other habits worth knowing.

The wrong segment boundary

The IME has to decide where one word ends and the next begins before it can convert anything. The textbook example is ここではきものをぬいでください, which can come out as ここで履物を脱いでください ("please take off your footwear here") or ここでは着物を脱いでください ("please take off your kimono here"). Both conversions are grammatical. Only one is what the sign meant.

Segment errors are easy to fix while typing: in most IMEs, the left and right arrow keys move between segments and Shift with an arrow key makes the current segment longer or shorter. They are hard to spot afterwards, because the result is fluent Japanese with a different meaning.

Rōmaji slips that become kana errors

If you type in rōmaji, your keystrokes follow pronunciation, but Japanese spelling does not always. The particle は is pronounced wa, so typing konnichiwa gives こんにちわ instead of the standard こんにちは. Typing o for the object particle gives お instead of を, as in 本お読む for 本を読む. And ん can swallow the next vowel: type kanni hoping for かんに and many IMEs produce かんい, because nn has already become ん and the i stands alone. These are kana-level slips and the IME will happily convert them into something else again.

The piece on typing Japanese covers input methods and IME settings that make some of these less likely.

Prediction and learning

Modern IMEs predict. After two or three kana they offer a whole word or phrase, and pressing Tab or an arrow key takes it. That saves a lot of typing and also makes it easy to accept the candidate one line above the one you wanted.

They also learn. If you once picked 意外 when you meant 以外, many IMEs move 意外 to the top of the list for いがい next time. One mistake becomes a habit of the machine. If you find the same wrong word turning up again and again, clear that entry from the IME's history, or reset the learning data in its settings, rather than just correcting the text each time.

User dictionary entries (単語登録) deserve a look too. An entry registered with a typo or a wrong kanji will reproduce the error every single time you type that reading.

Set phrases broken into parts

Four-character compounds and fixed expressions are often converted in pieces, and each piece can go wrong. 一堂に会する (ichidō ni kaisuru, "gather in one place") becomes 一同に会する because 一同 is the more common word. 絶体絶命 (zettai zetsumei, "desperate situation") becomes 絶対絶命 because 絶対 is so frequent. 危機一髪 becomes 危機一発. 五里霧中 becomes 五里夢中. The checker on the front page flags a number of these set phrases; the legend of issue types lists the kinds of things it looks for.

A before and after

Here is a short notice, typed quickly, with four conversion slips. Read it once before you look at the corrected version.

新しい機会の保障期間は、購入日から一年です。始めてご利用になる方は、上記意外の点についても窓口までお問い合わせください。

Every word in it is real. Now the corrected text:

新しい機械の保証期間は、購入日から一年です。初めてご利用になる方は、上記以外の点についても窓口までお問い合わせください。

("The warranty period for the new machine is one year from the date of purchase. If you are using it for the first time, please also contact the counter about points other than those above.")

The four changes: 機会 to 機械 (a machine, not an opportunity), 保障 to 保証 (a product warranty), 始めて to 初めて (the first time), 意外 to 以外 ("other than"). Read at normal speed, the first version looks fine to most people, which is the whole problem.

A routine for catching them

Reading slowly helps, but it does not help enough on its own. What works better is to change how you look at the text so your eye stops filling in what it expects.

1. Keep a personal list. Write down the pairs you or your writers actually get wrong. Most people have a handful of repeat offenders. Before you sign off a document, search for each word on your list and check every hit in context. A search for 意外 takes seconds and finds every instance.
2. Search by reading groups. For the big three (以外/意外, 保証/保障/補償, 初め/始め), search for both spellings, not just the wrong one. Sometimes the slip is the other way round.
3. Check compounds in set phrases. Search for the first two characters of the four-character compounds you use often and confirm the other two.
4. Use the candidate notes. Many IMEs mark homophones in the candidate list and show a short usage note when you move to them. If you are unsure while typing, read the note instead of guessing.
5. Reconvert to check. Most IMEs can reconvert a word that is already in the text (再変換). Select it and use the reconvert command (on Windows with a Japanese keyboard, the 変換 key) to see the other candidates, often with their notes.
6. Print it, or change the font. A different layout breaks the familiarity that makes slips invisible. Even switching the editor to a different typeface or to vertical text can be enough.
7. Run a rule-based check. The checker on the front page catches several fixed-phrase slips and a lot of the surrounding noise (doubled particles, mixed character widths), which leaves your attention for the meaning.

One method does not help much here: reading aloud. It is excellent for particles and rhythm, as the piece on reading aloud explains, but homophones sound identical by definition. 以外 and 意外 are the same aloud. For this kind of slip you need your eyes on the kanji.

Questions readers ask

Do native speakers make these mistakes too?

Constantly. Conversion slips are typing errors, not knowledge errors. A writer who knows perfectly well what 補償 means will still send 保証 when the IME offers it first and they are typing fast. That is why checking for them is a normal part of proofreading Japanese, whoever wrote the text.

Is kana input safer than rōmaji input?

It avoids the rōmaji-specific slips, such as こんにちわ or お for を, because you type the kana directly. Homophone errors happen equally with both, since the conversion step is the same.

Can I switch prediction off?

In most IMEs, yes, in the settings, and some people who type fast prefer to. It will cut down accidental picks from the prediction list. It will not stop homophone errors, which come from the main conversion.

Should I just write the tricky words in kana?

Sometimes that is the right call, and many style guides open certain words into hiragana anyway. But it changes the register and the look of the text, so it is a style decision for the whole document, not a fix for a single slip.