Ted Dunning wrote: > Also, if you choose to switch to a different translator at some point, it > is likely that they will use the previous translations as the base for a > translation memory even if humans are doing the translation. That counts > as the project using the text to train a translation engine.
I don't think that counts. It's one thing to teach a local model that X means Y in another language. It's another to teach the global model that X means Y in another language. If "my translation app" takes your X->Y and uses it to apply to the next application it sees, then it's opening itself up to some really bad poisoning models. Because there's a lot of garbage that will be uploaded into translation engines. I'd be shocked if anyone actually did this. And yes, I do maintain translation tools. My tools certainly wouldn't do this. I maintain translation tools because I've seen the quality of translations, and they're awful. The goal of such a restriction is to prevent someone from using this output as a basis for making another generic translation tool. If someone takes a document from Spanish, and uses Bing to translate it into French, Microsoft is not going to complain if someone later takes that (French) document and translates it into Italian, no matter who does the translation. They're only concerned if someone takes the mapping between Spanish words and French words and uses it on an unrelated corpus / to improve their handling of unrelated corpora.
