This is a patent on a very particular form of translation model that handles rare words, i.e. this paper that all the authors are on: http://arxiv.org/abs/1410.8206
What you wrote is literally the exact meaning I intended with the title, but now reading some comments I see many people are interpreting it as "Google atttempts to patent deep neural networkS for machine translation", (notice the plural) which has a vastly different meaning. I guess I assumed that this audience would know that it wouldn't be remotely possible to patent a broad technique that's already so widely published.
I think that even though the application is for a specific architecture, that this is still worth knowing about, since LSTM is a such well known technique for dealing with sequential data like sentences, and since so far virtually all progress in AI and ML has been driven by academia and has remained open.
I also think it's important to note that pretty much any interesting machine translation task will contain rare words, so even though they've framed it in terms of a specific task, it's one that's actually extremely broad. Machine translation is not so hard when you have tons of data and when you've seen every word many times in combination with its translations. The only really interesting case is when we have to use background knowledge and context to infer meaning. Since natural languages are notoriously ambiguous, this happens all the time. So this app may be broader than it first appears.
Also the patent application claims priority to a provisional patent application filed October 24, 2014, which was in all likelihood filed in view of the paper being submitted on October 30, 2014.
Okay, that would allow it - the gap is slightly (1 week) less than a year, and the paper is published by Google employees who presumably are in the patent; the publication would still disallow someone else from patenting it even within that one year.
If memory serves, Google did this with Word2Vec as well, which is a bit sketchy.
If you open source something, there should be a reasonable expectation that it is contributed to the commons. Otherwise a lot of people will build off your stuff, which you can then turn around, file a patent, and claim infringement.
I think that even though the application is for a specific architecture, that this is still worth knowing about, since LSTM is a such well known technique for dealing with sequential data like sentences, and since so far virtually all progress in AI and ML has been driven by academia and has remained open.
I also think it's important to note that pretty much any interesting machine translation task will contain rare words, so even though they've framed it in terms of a specific task, it's one that's actually extremely broad. Machine translation is not so hard when you have tons of data and when you've seen every word many times in combination with its translations. The only really interesting case is when we have to use background knowledge and context to infer meaning. Since natural languages are notoriously ambiguous, this happens all the time. So this app may be broader than it first appears.