Shunrō Oshikawa, ChatGPT, and the Birth of the "Pre-Translator"
How LLMs could help open up whole new worlds of literature
Author’s Note: This post is not an endorsement of replacing human translators with AI.
Last week, I had a bit of a moment and decided that I wanted to translate a story by Shunrō Oshikawa, one of the fathers of Japanese science fiction. Shunrō Oshikawa mostly wrote pulpy “boys’ adventure stories”, but many of them bear traces of turn-of-the-century sci-fi writers like Jules Verne. His most famous story, Kaitei Gunkan (The Undersea Warship) is a pretty clear homage to Verne’s 20,000 Leagues Under the Sea.
As far as I can tell, Kaitei Gunkan is the only work by Shunrō Oshikawa that has been translated into English. The rest of his body of work is basically unknown outside of the Anglophone world. This is mostly due to the fact that his work often does not hold up to modern literary standards, partially due to his early death, and partially due to the fact that he had some political views that would raise eyebrows today.
Since I’ve been on a foreign SF kick lately, I wanted to translate Oshikawa as a sort of literary experiment. My discussion with Regina Kanyu Wang and others convinced me that knowing another language is only part of translation. A translator must also be able to make informed trade-offs about preserving different aspects of the text, and it helps to be able to think out of the box about what kind of translation might resonate.
Obviously these skills will come easier for someone fluent in both languages, but that does not mean that a given bilingual reader will possess them. Translation is a unique practice that can reveal things about the translator— it can challenge their dominant narrative voice or underlying assumptions that they unconsciously apply. No matter how strong their command of both languages is, only a reader who is paying lots of attention to these factors will make an effective translator.
This all made me very excited to try my hand at translation, but there was one small problem: I’m only fluent in English. I figured I wouldn’t be able to get much out of the intended benefit of translation if I was just using Google Translate. But then I remembered something: isn’t ChatGPT actually surprisingly good for translating stuff?
Perhaps I could realize a small part of my dream of translation after all.
Google Translate was obviously a paradigm shift for digital translation, but I would argue that ChatGPT has been just as significant— not because of increased accuracy, but because of increased context. Essentially, Google Translate can only translate at the sentence level, and it must be deterministic— every sentence translation must always produce the same result. These individual translations tend to be more true than those generated by ChatGPT, but the basic principle is quite limiting.
Meanwhile, ChatGPT is highly adaptive, using the context of the surrounding text to “chunk” each sentence and paragraph into translation that are more likely to reflect the author’s intent. This context can be increased further by adding more background information about the text, or even talking to the text itself. I can literally ask ChatGPT about any potential mistranslations or weird connotations— a killer feature that clearly gives LLMs a huge advantage over any pre-existing translation tool.
Of course, all this still wouldn’t be enough to allow me to translate the heights of experimental Chinese literary fiction. But Oshikawa was an ideal candidate for this exercise because his work is relatively pulpy and straightforward. I wouldn’t be running into too many situations where I had to preserve the nuance of some multilayered reference to a proverb loaded with cultural meaning. But at the same time, it was written so long ago that some aspects would remain suitably challenging.
The translation was a lot more draining than I was expecting. I really agonized over a lot of choices, like whether I should translate the name of this archaic flashlight-like device as “flashlight” or “electric lantern”. I also decided it would be appropriate to take some fun creative liberties, like adding a pulpy exclamation point to every section title, and giving Tōsuke a folksy American voice rather than trying to preserve his tone more literally. In total, it probably took me something like 10 hours.
Even though the experience was a slog, but I still got a lot out of it. I did feel like I was able to get really close to the text in a unique way, and there was something magical about knowing that I was perhaps the first English speaker to ever read the text— I felt like I was playing a unique role in the preservation of human culture. My result was by no means professional quality, but I am still proud of how it improved upon the Google Translate version. And hey, an okay translation is better than none!
At the end of it all, “The Great Moon Race!” was… pretty bad. Even though some of its flair was almost certainly lost in translation, you don’t need to be fluent in Japanese to recognize that this story is no “Flowers for Algernon”. Baron Akiyama is cartoonishly evil, and the romantic conflict that ostensibly sets up the story is not mentioned at all after the opening section. Also, the story kind of just ends. ChatGPT mentioned that the end of my translation was “a bit flat”, before admitting that this is mostly just the original text being bad and not an issue of translation.
And this raises an interesting question— if so much of the quality of the text was apparent here, might literature enthusiasts be able to “read” world literature without knowing multiple languages? Obviously AI-assisted translation doesn’t have much practical purpose for actual publishing— there’s no one at the top to verify whether or not the translation is even accurate (hallucinations are sure to exist), and LLMs are still far away from displaying the nuanced cultural knowledge of a professional translator. But a professional-quality translation isn’t necessarily needed to curate foreign texts, especially if they can be sent up to a professional translator later.
Ultimately, we as a society have perhaps not fully grappled with the extent to which language barriers have lowered in the past decade. It is not too difficult to visit some foreign part of the Internet (like Douban) and find lots of information about highly acclaimed works of science fiction and fantasy that have never been translated. Not because of lack of merit, but simply because there are not enough bilingual readers to go around! Now that LLMs exist, even the monolingual readers could probably find an untranslated work that they enjoy quite a bit.
I’m not entirely sure what the implications of this are, but I strongly suspect that LLMs could end up being far more consequential for world literature than is currently acknowledged. Perhaps the next American sci-fi classic is already out there in Polish or Finnish, just waiting to be reborn. If anyone reading this post has had interesting adventures on foreign Internet, I’d be interested to hear about it in the comments!
You'll enjoy a similar exploration of translation and LLMs by Ken Liu, here: https://bigthink.com/high-culture/ken-liu-ai-art/
When I was a teenager (I think? or maybe I was in my 20s, I dunno), I had a burst of trying to translate some Spanish poetry I wanted to read despite not speaking the language - just really painstakingly with a dictionary. Hadn't thought about that in years, but even though it was slow and painful, it really was fun!