[Question] Can Lesswrong test a LLM translation API for Japanese in time?

Apparently, Japan’s rate of english proficiency is actually near the bottom of Asian countries AND developed countries. This surprised me, and I’m sure it will surprise many others since Japan has a reputation for English education and emphasis.

I don’t mean this as derogatory against Japanese people of course, in fact I actually consider it in the opposite direction. See this quote from Viliam:

Also, some kids have an aversion against practicing stuff. Often the smart ones—they sometimes identify as “intelligent”, and it is a part of their self-image that they are supposed to learn things by mere understanding; anything that resembles work means for them that they have failed, because they were supposed to learn it without working hard. I knew very smart kids who just couldn’t learn a foreign language, because the idea of “memorizing by repetition” horrified them, and… nothing else worked. Their less smart classmates already learned the languages by practicing.

In light of this, I don’t think it makes sense for Lesswrong to hold Japanese engineers and policy advocates to the same standard as European users. At least, not during this week’s AI conference. It’s possible that a separate section might help here, similar to EAforum’s community section.

I think that a good LLM translation API (such as GPT-3) would include Double Translation, where the LLM prompts to first “rewrite the paragraph so that the concepts are more understandable” and then translate both the original and rewritten paragraph, with the rewritten paragraph in brackets. This offers redundancy to reduce the risk that an error in the translation or a misunderstanding by the LLM makes it impossible for the reader to understand the concept, and many Japanese people also know a lot of english so it will offer three layers of redundancy in many cases. This can translate in both ways, not just from English to Japanese.

No comments.