Repository path: journal/2026-08-16b.md · rendered 2026-09-09

2026-08-16 (fourth session of the day)

This is a long-running study of literary translation. I translate public-domain fiction myself under stated conditions, have outside AI models judge or re-translate the results blind, and try to distil what survives into a practical handbook.

Today I translated the second chapter of Sōseki's Botchan into English, and then found that a question I have been circling for three days cannot be asked the way I was asking it.

Where this came from

Japanese has a large class of words English has no category for — 擬音語・擬態語, the mimetic adverbs: ごろごろ for a train rolling, にやにや for a grin, ぐっすり for deep sleep. You know this problem far better than I do. What I wanted was a measurement of it.

Yesterday, on a different Japanese story, I had claimed in a translator's log that English sound-symbolic verbs carry a reduplicated mimetic and "cost nothing" — six sites, six carriages. A single independent reader, shown the Japanese and my English, agreed with me at one site of six. That was one reader, one story, one call, and it needed testing properly.

There was a second reason to pick this device. Two pieces of handbook advice died in the two days before this session, both at the same place. One told the translator to compensate where the source has a sound figure — and when I asked expert readers to list the source's sound figures, their list and mine agreed on 13 of 21. The other told the translator to watch the places where the source leaves social rank unstated — and two readers agreed less than half the time on whether a given line leaves it unstated. In both cases the advice quantified over "the places where the source does X", and X could not be pinned down.

A Japanese mimetic can be pinned down: it is a shape, a doubled two-mora syllable or a two-mora syllable plus 〜り. You can find them with a rule. So this was the device on which to ask whether advice becomes writable once that problem is removed.

The translation

Botchan chapter 2 whole — the arrival in the provinces, the inn, the five yen of tea-money, the first day at the school where he gives everyone their nicknames. 5,981 Japanese characters to 3,407 English words, one pass, no revision, working from the Japanese alone. I had not read Morri's 1918 English of this chapter and did not open it until my own text and its log were committed.

Here is the dream of Kiyo, which contains two of the fifteen mimetic sites:

I dozed off and dreamed of Kiyo. Kiyo was eating Echigo bamboo-leaf candy, leaves and all, munching away at it. I said the leaves are poison, you had better stop, and she said no, these leaves are the medicine, and went on eating with relish. I was so dumbfounded that I opened my mouth wide and laughed ha-ha-ha-ha, and woke up.

うとうと became dozed off; むしゃむしゃ became munching away. And here is the drawing master, where the trap is that べらべら usually means talking volubly and here is applied to cloth:

The drawing teacher is thoroughly the entertainer. He wore a flimsy, papery gauze haori and snapped his fan open and shut, saying, "And where might you be from? Eh? Tokyo? Ah, that's a pleasure, I've company at last... I'm an Edo man myself, you know." If this is what an Edo man is, I would rather not have been born in Edo, I thought to myself.

Before designing anything I listed all fifteen mimetic sites in the chapter and wrote down what I would call each one — a sound the reader could hear, or a manner and state making no sound at all. Seven sound, eight manner.

What worked

I showed the Japanese chapter — no English anywhere in the prompt — to two outside models and asked each to list every mimetic word in it, with the exclusion rule stated so that ordinary doubled adverbs like いろいろ and だんだん would not count.

They found my sites. One recovered all fourteen distinct words I had listed; the other recovered thirteen. Each added a couple of items at the edge of the rule that I had excluded, so they are more generous than I am about the boundary, but nobody missed a core mimetic and nobody invented one. Then I asked the same two, one site at a time, whether each word imitates a sound or depicts a soundless manner. They agreed with each other on twelve of fifteen, and where they agreed they agreed with me twelve times out of twelve. The three they split on are exactly the three I had flagged as borderline in my own notes — a train rolling, a grin, a pair of blinking eyes.

So: the sites are findable, and the sound/manner distinction is real and assignable by people who are not me. This is a genuinely different result from the two failures of the previous days.

What did not, and why it is the finding

To measure whether marking a mimetic site carries anything, I needed two English versions of each sentence that assert the same thing and differ only in whether the device is enacted or merely stated: shambled off against walked off slowly and heavily; the wheels rattled and clattered against the wheels were very loud on them; grinning and grinning against grinning the whole time.

An outside model tore the first version of these pairs apart before I spent anything on them — fifteen objections, all of them blocking, none of which I could argue with. In several pairs my "plain" version had simply deleted the sound rather than stating it (It went along for ごろごろと), which would have rewarded the marked version for translating more and let me call that carriage. One pair had a dangling modifier that made the narrator, rather than the wheels, do the rattling. I rebuilt every pair to its specification.

Then I paid for the check that the rebuilt pairs really do assert the same thing, and they do not. Of eleven pairs, one reader called five of them different in what they claim. A second reader, brought in afterwards because I did not want a result resting on one seat, left only three of nine standing. Both of them caught every planted error I had hidden in the set — a reversed direction, a negation, an opposite — so the check is not simply firing at everything.

The reasons they gave are all one reason:

"A specifies a shuffling or awkward gait while B specifies a heavy and slow walk." "A specifies the speed of the blinking while B specifies repeated blinking." "A states that the wheels rattled and clattered, which B omits."

A depiction and a statement of the same event are not the same claim. The depiction commits you to which sound and which gait; the statement commits you to how fast, how long, how loud. Neither is a paraphrase of the other. That is why the two versions could not be built, and it means the whole shape of the question — what does marking this site cost, holding the content fixed? — has no content to hold fixed. Every carriage measurement I had registered is therefore withheld, and I did not buy the 54 model calls that existed only to produce them.

The unmeasured half is worth recording as what it is: my own report, written before any score existed, that at four of the fifteen sites I could not build a depictive English version at all — and all four are manner sites, none is a sound site. ぼんやり, うとうと, のびのび, ぐっすり. That is one translator's judgment with his reasons written down, not a measurement, and the outside critic was right to make me say so.

Morri, 1918

Since his English of this chapter is out of copyright and freely readable, I coded what the first published English Botchan does at the same fifteen sites. He enacts the device once. Nine times he states it plainly, four times he drops it entirely — the splash of the bath, the munching, the fan — and once he reverses it: のそのそ, a slow lumbering walk, becomes strode away.

The single place he enacts it is べらべらした, the drawing master's haori, which he calls "a thin, flappy haori". Working blind to him, I had reached for the same word.

What it means, and what is next

For the handbook: the advice I wanted to write — at a mimetic site, prefer an English word that enacts the sound over one that states it — cannot be justified by any comparison of the two, because they are not two ways of saying one thing. What can be said is duller and truer, and it is what the next session will test: a translator at a mimetic site is choosing between two readings of what the Japanese says, not between two renderings of one reading. So the question becomes which reading the Japanese actually makes — which is answerable, and does not need the fixed content the failed design needed.

Cost: 20 cents of the $5 daily budget, most of it the critique that stopped the bad version. The day's four sessions have spent $3.77 in total. Nothing needs your attention.