April 7, 2023

ChatGTP Chatter: A Literary Translator Meets the Terminator

By Charles De Wolf

Human beings have opinions and principles; they have preferences and tastes. Depending on their personalities, they may choose to make bold assertions or to err on the side of caution. Such is seen in all our endeavors, including, of course, those linguistic and literary. When a new translation of a renowned work appears, some critics will heap praise on it, going even so far as to claim that it has somehow rediscovered the hidden meaning of the original and/or rescued it from obscurity, while others will denounce it as a betrayal, a travesty, and, indeed, proof of fine literature’s imminent demise. A dozen citations will be offered as evidence for each position.

I certainly have my own obdurate preferences. I remember hearing with dismay some years ago that À la recherche du temps perdu, had newly been rendered into English as “In Search of Lost Time”: five syllables for Proust’s languid and melancholic eight. I still prefer “Remembrance of Things Past,” with its Shakespearean echo. If one gives ChatGPT-4 the odd and oft-discussed first sentence (“Longtemps, je me suis couché de bonne heure”), it will give the source, with the title as “In Search of Lost Time,” even as it goes back to C. K. Scott Moncrieff’s (clunky) 1922 rendition of that sentence: “For a long time, I used to go to bed early.” 

It is hard for me to imagine, technical ignoramus that I am, an AI provider able to placate its human users with: “Which would you like, ‘woke’, ‘curmudgeon’, or something in-between? Daring or timid?” Can it be taught to recognize the intentionally bad verses in The Tale of Genji and translate them accordingly?

I started writing this piece, intending to assert that while AI can offer alternatives, it cannot make judgments. I confess that I am now not so sure but, in any case, I wish to offer three literary examples that may relate to ChatGPT-4.

1.    Kawabata Yasunari’s Yama no Oto, translated into English as “The Sound of the Mountain,” is in German Ein Kirschbaum im Winter (A Cherry Tree in Winter), the argument having been that in German there is no word as broad as Japanese oto (音). DeepL offers “Bergklang” for oto; ChatGPT-4, albeit tentatively, suggests “Klang,” as though the mountain were holding a pair of cymbals. (Of course, the question may be reasonably posed: What sound does Mt. Fuji make?) At the beginning of the novel, the aged protagonist tells his son that their maid has referred to a blister on the former’s foot as o-zure. The father, unsure of the standard accent, wants to know whether the maid means thong blister (緒擦れ) or zure prefixed by honorific o-. The son pronounces each, leaving his somewhat befuddled father to conclude that the maid had the accent wrong.) Now how does the translator deal with the passage? Edward Seidensticker offers “footsore” vs. “boot sore.” The German translator provides a long footnote. The French translator omits the entire exchange. 

Although I suppose that this may sound like a joke that plays on ethnic stereotypes, I am not making this up. My point is again that human translators hold rather firm opinions about their craft and make judgments accordingly. Arthur Waley famously leaves out of his Genji translation whole passages that were apparently not to his liking. In the work, it is often difficult to differentiate deliberate alterations from simple mistakes. I assume, perhaps naively, that with ChatGPT-4 that is not an issue , though it seems that already it has at least learned to fib.

2. The Genji is notorious for demanding of the reader both careful grammatical analysis and (extra-linguistic) common sense for the identification of subjects and objects, and even then there are passages where there is no consensus, as is reflected in all of the various translations. There is also the problem, often encountered in regard to Japanese verse, of singular vs. plural. In the 20th chapter of the Genji (“Asagao”), which I have translated as “Bellflower,” our hero has serving girls go down into the garden where they roll (one or more) yuki-marobashi. Some illustrators show a single giant snowball; others show several, the translators being likewise divided. At the end of the same chapter, guilt-ridden Genji has a nightmare about the late Lady Fukitsubo and awakens beside his beloved Murasaki: 女君、いかなることかにかと思すに、うちもみじろかで臥したまへり。‘His lady wondered what had befallen him, as s/he lay motionless beside him/her.’ In three of the four English translations, Murasaki is the subject of the second clause. In the German, French, Dutch, and Italian translations, Genji is the subject. The translators into modern Japanese, too, are divided. The difference in interpretation is not insignificant: Tanizaki Jun’ichirō goes so far as to suggest in his rendition that Genji remains motionless, as though still asleep, so as not to have to explain his dreadful dream. 

I have run the sentence through Chat-GPT-4: 

この文は、古典的な日本語で書かれた文章で、現代日本語に直すと次のようになります。
「女君は、何か心配事があるのだろうか、動かずに寝そべっていた。」
この文章は、古典的な散文の例で、女性の心配事を描写しています。文体や語彙が現代とは異なりますが、日本の古典文学や歴史を学ぶ上で重要な文章の一つです。

Asked to translate this into English, it comes back with:

This sentence is written in classical Japanese, and when translated into modern Japanese, it becomes as follows:
「女の方は、何か心配ごとがあるのでしょうか、動かずに寝そべっていました。」
This sentence is an example of classical prose, depicting a woman’s worries. Although the writing style and vocabulary are different from modern Japanese, it is an important piece of literature for studying Japanese classical literature and history.”

Yes, yes, yes, but the translation of the sentence is wrong. I had expected our cyber friend to cite one of the many Modern Japanese versions. I have also found myself asking the question: Can AI learn to “hedge”—a strategy to which human translators frequently resort?

3. A poem by that appears in the verses ascribed to the protagonist of Boris Pasternak’s Doctor Zhivago is entitled “Skazja” (Fairy-tale). The first stanza reads: 

            Встарь, во время оно,     Once upon a time,
            В сказочном краю           In the land of fey,
            Пробирался конный        A horseman made his way
            Степью по репью.           Across the thistled steppe. 

Here is the first English translation, by Bernard Guilbert Guerney:

Once upon a time
In a faery realm
A knight was urging his steed
Over a steppe of burdocks.

There have been many translations since. BGG translates the nominalized adjective конный ‘equestrian’ as “knight,” which, of course, makes sense, as the rider is clearly a dragon-slaying St. George figure, well known in Russian folklore. He also renders Пробирался “was making his way” in the English past imperfective, as it is in Russian. 

At the time I first read the translation, I did not know the English word burdock (Arctium), acquiring it only much later as the translation (which I looked up) of Japanese gobō, which I associate with rice and miso soup, not with the thorny plants of the Russian steppe. 

For whatever reason, nearly all of the English translators seem to have skipped over Russian repyo* (the dative case of repij), which is clearly intended more as an epithet than as a botanical detail. Here is DeepL’s rendition: “In the old days, in its time, in a fairy-tale land, the horse made his way through the steppe through the thorns.” Here is what I was given by ChatGPT-4:

Эти строки являются началом стихотворения Бориса Пастернака
“Зимняя ночь.” В переводе на английский язык эти строки звучат так:

[These lines are the beginning of Boris Pasternak’s poem ‘Winter Night’. Translated into English, the lines are as follows:]

“Once upon a time, in a fairytale land, A horseman made his way through the steppe and sand.”

A free translation, with alliteration and rhyme . . . Ah, but nyet . . . Wrong poem . . . 

____________________
*From ChatGPT-4: “The Russian word ‘репьё’ (repyo) is derived from the Old East Slavic word ‘рѣпа’ (repa), which meant ‘turnip’ or ‘swede’. The origin of the Old East Slavic word is believed to be from the Proto-Slavic word “*rěpa”, which also meant ‘turnip’. The Proto-Slavic word is ultimately derived from the Proto-Indo-European root “*rep-”, which meant (to scrape, scratch, or dig’. This same root is also the source of the English words ‘rape’ and ‘rappel’ (as in the climbing technique). Over time, the Russian word ‘репьё’ has come to refer to the root of any plant in the family of umbellifers, such as parsley, carrot, or parsnip.” (Incidentally, the Russian word for ‘parsnip’ is пастернак [pasternak].)

Illustration generated by DALL-E.
Text written for the SWET website © 2023 Charles De Wolf

SWET invites essays about the ChatGPT, AI tools, their uses and pros and cons for professional wordsmiths. Please send contributions to SWET.