I’m not talking about translations of basic dialog such as “How are you?” or “Nice to meet you.” from English to let’s say Japanese, Mandarin, French, Spanish & whatever. Rather the topic dwells more on translation apps or LLM’s failing to capture the nuance of a conversation since words can connotate different definitions as well as being understood differently when applied.
Take the word 蛇行 (だこう) which is defined as amongst the lines of “meandering” but when compounded with 蛇行運転 for example, the definition leans closer to “erratic driving” but most online translatiors mess this up on saying “driving in a serpentine manner” which sounds stupid. No wonder why Japanese is hard for LLM’s to work with.
These tools are as good as the information fed to them, so I’d presume they weren’t fed nuanced situations. Now why would that be, I have some theories, either the people building the machine didn’t think/remember/find such materials to feed it, or these materials are hard to come by.
Reminds me how, also in Japanese, most dictionaries online I found, and even some people I asked about, couldn’t tell the difference between 今日 and 本日. What helped me was having enabled the Japanese-French dictionary setting in a multi-language dictionary I checked despite barely understanding French.
I feel like the fundamental reason is that words have multiple definitions. 蛇 does mean snake, so “driving in a serpentine manner” kinda makes sense, even though it’s not the primary definition. I feel like it’s kinda the wrong tool for the job. If you want to understand the nuance, you could look at a multilingual dictionary that gives multiple definitions. Jisho.org does define 蛇行運転 as “erratic driving; driving in a zigzag.”
Not a great example because that’s the only definition it gives, but suppose I want to look up the word 浪人. Impressively, Google translate gives “Ronin (a student who failed the university entrance exam and is studying to retake it)” but a dictionary like Jisho will give me that PLUS the original meaning: “ronin; masterless samurai.”
Well, the lesson here is to use the right tool. DeepL, Google Translate, and Microsoft Translator all get this one right on just the term alone, and I’m sure even fringe translators would get something with a little more context right–a sentence like, 証言者は、事故を起こした運転者が蛇行運転していたと言いました。
Use Jisho instead for terms, or one of the JMDict-wrapped extensions like Yomitan.
For the same reason AI is crap at mimicking human speech, human art and human reactions.
Languages are profoundly human. They’re not just a code.
Which ones have you used / use the most / have the most gripes with? Other people might be able to point you in the right direction for a good one.
DeepL and Google trnpanslate are the two I know, and I only use deepL now. Don’t remember the names of any others.
As for actually why, well I seem to remember google translate was focused on nuance but seemed to fall off or outright regress in quality since LLMs have become a thing.
They should come up with one box for literal meaning and contextual meaning. You could highlight bits and then have the interface translate just that section. The software is woefully inept for so useful a service 🥲
