>>27989895>like we're stuck in the 2000s regarding nihongo machine translations, even deepl feels a decade behind, i mean i understand why it's so hard but you'd think we would be there by now.I was an N3 over ten years ago (am ojisan, it's definitely declined since then), and professionally am a senior software engineer. Machine translation is not my specialty, so please note that huge caveat, but I have read some write-ups and technical blogs on it over the years. I have a deep appreciation for how far we have come, given the intense grammatical gaps between Japanese and English.
Informal speech relies so much on unspoken context, or inter-sentence/inter-paragraph context, that it is difficult in and of itself to fill in those gaps when modeling a conversation. The fact that TL-kun translates, then throws out the input and translates the next sentence fresh introduces a fundamental limitation. Even if TL-kun had perfect transcription, not building a topic model of the conversation itself is just a giant barrier to effective translation. But as it has been said, transcription is an outstanding challenge in a language like Japanese, with gorillions of homophones and ever-evolving slang (like shortening the name of a brand/franchise name, think "konosuba").
I could ramble on more, but basically I both agree "it feels behind" and also feel like "holy shit I'm amazed we have gotten this far".
