Similarly, I would pay a lot of money for a Spotify or Netflix alternative where a knowledgeable human looked at my activity every month or so and created a hand-tailored recommendation list, since those products seem pretty bad at nuance in ways that even just simple "do you like this for reason A or reason B" questions would answer.
I know how the sausage is made at some non-Google places, and there's a common institutional blind-spot where the algorithms people don't want any human-driven chocolate getting in their peanut butter, and lazy product people are perfectly content to let any junk the algorithm throws up be the math guys' fault. Sadly, the world is full of edge cases, and sometimes you need one-off rules. (There's a parallel thread going below where people are aggressively defending the fact that the translate product as implemented is bad at certain use cases. Baffling. At the end of the day, are you happier telling users they're doing it wrong than given them what they want?)
Edit: I did not want to complain about the translation quality nor show my sense of entitlement. You have authoritatively stated and used a some jargon to tell us it doesn't use English when translating between languages, so I simply wanted to understand more about this process because, on the surface, it looks like English was involved in my specific example.
I can't say I can point to any sort of evidence of this, so take it with a pinch of salt, but my gut feeling is that the people in Google who are responsible for GT are somewhat fanatical about making purely statistical machine translation (and, in particular, deep-learning machine translation, a.k.a. neural machine translation) work. They are most probably perfectly well aware that it is possible to improve their system by using some rule-based fall-backs, or some background knowledge, but they have decided that they dont' want to use those and instead want to do machine translation from scratch and from data only, without any expert input.
So it sucks, like all pure systems suck.
This is like complaining that Word isn’t as good as vi, which is decades old, for text editing. Sure they could bolt a text editor onto Word if they wanted to, but why bother? That’s not what people are looking for when they use the product. And if that’s what you want, there are many other choices that are much better.
I would start by determining the number of words in the query. If it's equal to 1, the user probably wants to translate a single word, maybe give him the 1:1 dictionary result. If there's more, the other words possibly provide context, so let's use the sophisticated thing.
https://www.bloomberg.com/news/articles/2018-02-02/translati...
GT Engineers don't want a pat on the back. They want to do better.
Which makes a lot of sense when you realise that all the languages in the language pairs that GT's interlingua covers will have many, many more texts going to and from English, than to and from any other language. Attempting to derive a many-to-many language model from an English-dominated corpus is very likely to lead to an intermediary model that is basically English-to-many.
In fact, if you read the relevant paper [1] the dominance of English in tables etc jumps out at you.
And if you look at how Google figured out its Neural Machine Translation system has created an interlingua, that is ... visualisation. So it's not like they have some kind of principled way to demonstrate that this interlingua they advertise actually exists. They just look at some pretty images and interpret them to say that it does.
Even this visual "evidence" as they call it, is from a many-to-many language model trained on Japanese, Korean and -you guessed it- English.
__________________
No, it doesn't. It goes via an intermediate representation which is common for all the languages they translate. The meaning of a phrase is represented as a "meaning vector", or more recently, a matrix. The problem is that it is hard to translate rare words, because when the neural net decodes the next word, it needs to consider the whole dictionary at once (the problem of too large softmax function slowing down the process). So they need to focus on a subset of words. There are various solutions, such as hierarchical softmax, but they are still difficult to train. Google translate uses a ton of compute to create a model as is. They probably can't offer everyone a version that is 2% better but uses 100x more resources.