Night457 wrote: ↑Thu Sep 14, 2023 12:32 am
The biggest criticism I have read of ChatGPT is that it is a brazen liar: it sometimes produces complete nonsense with utter conviction, and those who do not know better will find it thoroughly convincing because the results seem reasonable enough. In this case, someone who understands both Polish and English will recognize where the machine translation fails badly. If the goal is to translate a work of entertainment, then in the large scheme of things losing some accuracy to save time and effort is not such a big deal. People just need to realize that is happening, that's all. It is simply a question of whether or not a knowlegeable person is willing to put the time into correcting the idiot machine. Often they have other things they want or need to do. And I think in many cases, someone who is NOT a polyglot simply wants to get on with the job and make a subtitle that serves the purpose of expressing a meaning that is "close enough".
I think it is obvious what I consider the
ideal kind of translation, but I am also a realist and accept the limitations of available time, energy, and human knowledge. I do not even understand much of what exactly people are going on about in my own language, so I will not stress too much if I don't fully understand what is being said in other languages too. I just will try and not get in any fistfights over those misunderstandings, and everything will be fine.
Here is a fun thing that happened to me. I told ChatGPT to translate a Russian subtitle to English. However, I did not realize the Russian subtitle was saved with an improper encoding. As a result, ChatGPT was sent an unreadable file to translate. And translate it did! It just pulled 90 minutes of dialog directly from its ass - it made up every single damn line from nothing.
However, ChatGPT's tendency to make up what it doesn't know isn't always a bad thing. I think one of the things we pick up on when we watch a film with machine translated subtitles is bad grammar, for example. ChatGPT is less likely to generate bad grammar because it will make up details when it can't make sense of the input. And its inventions usually sound plausible enough to fool you. I also think using ChatGPT on Whisper subtitles works pretty well because of how ChatGPT can sometimes guess words where Whisper misheard and created an error.
Ultimately, I think ChatGPT is going to get really good at translations. Right now it's limited because 1) most of us, including myself, are using the older 3.5 models and not the new stuff and 2) the vast majority of the training set is English (over 90% of the training set for GPT 3.5) and the AI will need to be trained more in other languages to produce comparable results.