Is Machine Translation Solved?

A user of Japanese-English machine translation rated ChatGPT at roughly 80 out of 100 and DeepL at about 60 for one-shot paragraph translation, arguing the technology is very good but not "solved." The user, posting on LinkedIn, said a true solution would require real-time, streaming speech and extremely short utterances that let people communicate across languages without consciously compensating for translation. The user also noted that less common languages such as Sanskrit still have a long way to go.

Honestly curious about what others think on this. Cross-posted from llm nlproc machinelearning neuralempty | Li Ling Tan https://lnkd.in/p/g8jYxA9B Hmm… As a user rather than an MT researcher, my feeling is that machine translation is already very good, but I would hesitate to call it solved. For context, I mostly use Japanese and English in both directions, and rarely use other language pairs. If I judge fairly harshly and limit it to a one shot translation of about one paragraph, my very rough personal score would be around 80 out of 100 for ChatGPT and perhaps 60 for DeepL. Both can of course do better with some help. In particular, with ChatGPT I can usually improve the result a lot through conversation, context, and revision. But if I am evaluating the translation technology itself, I think it is fair to be stricter about the first attempt. My personal definition of “solved” would also be much broader than ordinary text translation. I would want something closer to complete communication across languages, using whatever communication channels are available. Real time translation, extremely short utterances, and streaming speech are all interesting boundaries to me. If a system could handle those naturally enough that people no longer had to consciously compensate for the fact that translation was happening, that would feel much closer to “solved” in a practical sense. The difficulty is that I am not even sure this remains machine translation in the academic sense. Context, vision, speech, memory, personalization, interaction, and feedback can all become part of the system. Even today, many LLM based workflows already feel difficult to describe as simple source text to target text translation. So I do expect substantial progress to continue. What I am less sure about is whether the most interesting future progress will still be called machine translation. It may increasingly look like a more general system for communication across languages, and personally I think that possibility is more interesting than simply making a human translator faster. it still has way to go for not so general lang like Sanskrit.