I first tested Google Translate out a few years back, leaving me reassured that AI would not replace me as a literary translator. AI has made great leaps since, however. Karin Kaneko’s recent article in The Japan Times compared four kinds of generative AI tools for translating Japanese text into English—Chat GPT-4, Bing, Bard, and DeepL—which have become very clever indeed. Chat GPT could even render the opening scene of Kawabata Yasunari’s Snow Country into Shakespearean rhyming couplets.
Is there nothing this bot can’t do? Watch out, technical translators: you’ll have to find another job, especially since most Japanese still don’t mind funny English.
I cut my translator’s teeth on the works of Izumi Kyōka (1873-1939) who—like many writers of his generation—started out writing in classical Japanese before crafting his own modern vernacular prose. As even Japanese can attest, he’s devilishly difficult.
I recently translated a 1901 Kyōka short story called “Three Feet Square” (Sanjaku-kaku). Kyōka’s sinuous sentences vividly evoke a lost world. Here’s a taste: first the original passage in romanization, then Chat GPT’s version, then mine.
Kore ga kadoyashiki de, oremagaru hairo o shita michi ga hitosuji, denchū no ichijirushiku katamuita no ga, mae e ushiro e, betsubetsu ni kashira o futte okufukō tatteiru, harigane ga nakabadarumi o shite, hisashi yori hikui tokoro o, yowayowa to, naname ni, samosamo otoroeta katachi de, eitai no hō kara nagaku tsuzuiteiru ga, zu ni kaita sen o hiku to, bunmei no teido ga dandan kotchi e kuru ni shitagōte, yanegoshi ni niburu koto ga wakaru dearō.
Chat CPT prefaces its translation as follows:
The provided text seems to be written in a mix of old-fashioned Japanese and some archaic kanji characters. Due to the archaic language and obscure context, translating it accurately into modern English might be challenging. However, I’ll attempt to provide a general translation based on the available information:
“This was the corner mansion, and as the road curved, a single gray line stretched out, with electric poles leaning forward and backward, their heads inclined separately, standing deep inside. The steel wires also sagged weakly and diagonally, seemingly worn-out, while the eternal direction continued from the west, but drawn in a gradual manner, indicating the progressive civilization in this direction.”
Chat GPT adds: Please note that this translation might not be exact, and some parts may still be ambiguous due to the challenging nature of the original text. If you have a specific context or purpose for this translation, it might be helpful to provide additional information to ensure a more accurate rendition.
The bot wants to know more and is using you, the user, to accomplish its task better.
Now my version:
“This was the corner residence and if you turned right, there was a single greyish road where the telephone poles leaned this way and that, helter-skelter, their lines drooping in the middle, lower than the eaves in some places, dangling diagonally as if exhausted. They ran from all the way back from behind the storehouse and, as if drawn on a blueprint, those lines traced the march of civilization from past to present, but you might say it sagged when it reached these rooves.”
Chat GPT’s “eternal direction” is a flat-out mistake, since eitai (which does mean “eternal”) is another name for a storehouse. It also misses the part about the telephone lines sagging under the eaves, and any human editor would nix using “direction” twice in a single line. While it does capture the gist, its prose wins no prizes. I certainly don’t hear the master’s voice in the AI version. So I can roll out rhythm and assonance and alliteration to conjure up a scene of a seedy, working-class district of woodshops on the banks of the Sumida estuary more than a century ago.
Will all translators devolve into copyeditors? I’ve found AI good at least for checking my translation’s accuracy. Having a host of literate Japanese readers around would be better, but everybody’s busy, right? Chat GPT is happy to spare a few milliseconds of its time, but don’t ask it to translate reams for you. More than a paragraph or so at a time taxes its algorithm and the result is not elegant.