It seems like something that, on paper, should be very easy to accomplish. Is it just down to the fact that certain things cannot be translated directly from one language to another?
In: 1
I think its important to recognize that language is not exactly a 1 to 1 process. Even though we have the words that may translate directly, often, this leads to cryptic and grammatically poor translations because how you would say something in English may be radical different than, say, in Hindi.
And this gets at the crux of it: no one really wants a simple word to word translation, we want meaning (or semantic) translation. And thats where it gets hard.
Most machine translation now days is done with machine learning. Specifically, they are done with Transformer networks. Data scientists will give this network A TON of text in both languages. For example, Google released a really popular model a few years ago called BERT, which was trained on the entirety of WIkipedia and any unpublished books they got their hands on. The point of doing this is so that *the network begins learning patterns in language* . This is why we sometimes get weirdness in translation: because the model actually has no understanding of the rules of any language, *it simply learns the patterns present in the text of those languages*.
I remember seeing a post somewhere on reddit a whole ago where someone was using grammarly and it recommended that the author shorten the phrase “he had not left” to “he left’nt”. Although this isn’t an example of translation, it illustrates my point: grammarly had no knowledge of how English grammar works, instead it was simply aware of the pattern of contractions and tried to apply it.
The same things happen in translation. The machine doesn’t know the actual rules of either language, its simply pattern matching between the two languages. And when those patterns don’t apply, the machine can make perplexing, sometimes humorous mistakes like “left’nt”.
Latest Answers