Why Japan is building its own version of ChatGPT

[John Stroman]

Dear Fellow 'Yakkers,

Below is a link to a short summary article in Nature with the above title. I thought this information might be interesting to both E>J and J>E translators.

https://www.nature.com/articles/d41586-023-02868-z?utm_source=Nature+Briefing&utm_campaign=8398e3a1df-briefing-dy-20230914&utm_medium=email&utm_term=0_c9dfd39373-8398e3a1df-49147323

One hurdle that AI must overcome is the complexity of the Japanese writing system both in writing prompts and in generating output that is acceptable to native Japanese speakers such as the correct use of 敬語.

John Stroman

----------------

[Geoffrey Trousselot]

I hope they think to include all historical writings. And I hope they include all other languages as well. It would be amazing if it included everything that has been digitalized to date.

I don't think we are far away from having digital multi-lingual avatars that will usher in multilingual business environments, open up the entire country to tourists and immigrants, where the exchange of information between languages will become so smooth, the idea of a language barrier will relegated to a potential risk in the case of a solar flare. 

Geoffrey Trousselot

--
You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/honyaku/CAJVTx96Kk-%2B2yjOgz9Y3c3ah8CN9_CvK1X9p2cM%2Bd3TfX2bsCw%40mail.gmail.com.

[Kevin Johnson]

Are we not missing the point this article is underscoring? It's literally stating that ChatGPT is bad at Japanese, so bad in fact that it's worth "hundreds of millions of dollars" of R&D to major corporations to try to fix this problem.

With respect, the conclusion I would draw from that is not "we're really close to a post-language era now." In fact, given that the article says "for an LLM to be useful and even commercially viable, it needs to accurately reflect cultural practices as well as language," the implication should be clear that ChatGPT is not even commercially viable in Japan right now due to its poor understanding of the Japanese language.

It's also important to note that the article is really talking about Japanese-to-Japanese capabilities, not translation capabilities. In fact, it lists as the main problem with ChatGPT that Japanese queries go through a process of Japanese->English->English->Japanese that produces poor results. That means its current translation capabilities are poor.

Their goal seems to be to produce something that works entirely in Japanese, not something "multilingual." It actually seems to be monolingual, using Japanese-only sources instead of translations. If anything, they likely want rock solid human translations as data to feed their Japanese-only LLM.

So to summarize, they're saying ChatGPT's ability to understand Japanese is so bad it needs heavy investment from corporations and universities to try to fix those problems before it's commercially viable or even useful.

Whether the magic of technology will someday overcome ChatGPT's current Japanese ineptitude is an open question anyone can speculate on, and lord knows we do a lot of speculation in this group.

But my experience with Japanese tech firms in 2023 says they're not very good or innovative (no capital) and unlikely to succeed where Silicon Valley failed. I hope they do, if only because Japan as a country has fallen so far behind technologically and economically that it's more 一億総貧困 than 一億総中流 (saw this phrase again in the Japanese news just today). Japanese salaries have been startling low for a long time and people are suffering as high energy costs and the cheap yen on top of that makes importing harder. Anything to bring money flowing back to Japan is a plus for people working with Japanese.

Kevin Johnson

[Tom Gally]

While it is true that ChatGPT does not perform as well in Japanese as it does in English—not surprisingly, as Japanese text probably accounted for less than 1% of its training data—it is also true that there has already been strong take-up of ChatGPT in Japan. Since June or so, I have seen many references in the Japanese press to companies, government agencies, and educational institutions adopting it. While some of that is no doubt the result of hype, I am sure that most have tried it first—in Japanese—and have found it useful.

The public discourse about AI in Japan seems to include more emphasis on positive applications and less concern about the potential threat to employment than does the discussion elsewhere. My explanation is that AI is seen here as a solution to the chronic low productivity of white-collar workers and the rapidly shrinking working-age population, without being a threat to the jobs of current full-time employees, who enjoy strong job security.

Just this morning, I talked for more than an hour with a professor at a university in Tokyo (not mine). The president of his university had called him in last week and asked him to come up with a proposal for integrating AI broadly into the university’s curriculum, and the professor wanted ideas from me. The president—a scientist—apparently uses ChatGPT frequently himself and is convinced of its value, and he believes that AI can be used effectively for education. I’ll be interested to see how that plays out. I personally think that AI can be a powerful tool for individualized learning. I’m not so sure, however, how well it can be integrated into the educational system in Japan.

Tom Gally

Yokohama, Japan

--

You received this message because you are subscribed to the Google Groups "Honyaku E<>J translation list" group.
To unsubscribe from this group and stop receiving emails from it, send an email to honyaku+u...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/honyaku/CAMk1scgWZUaCVboZ4bdR%2Bxgx1yCmjQu6tbBh0ES%3DUjs2cZcWQQ%40mail.gmail.com.