Facebook's AI Just Set a New Record in Translation

And was already submitted: https://news.ycombinator.com/item?id=17886827

How did we paint ourselves into this corner where the only way for our websites to exist is to run a different person's arbitrary code on our visitor's devices every time they open a page?

Chrome's protection only works in cross-origin iframes [1] and has been in beta for years. I haven't checked in a while but can't find a source that confirms that it went live.

Forbes serves a large portion of their ads in same origin iframes and so is not fully covered by this protection.

[1] https://blog.chromium.org/2017/11/expanding-user-protections...

Sure, but I think that would mean it would learn multiple meanings and the pattern groups that use each meaning.

Perhaps a more appropriate example (just something I thought of, idk if it would work exactly like this) would be like "antiviral" going to "an" "ti" "vi" "r" "a" "l". It could connect "an" and "ti" and associate it with the meaning of "anti" as a prefix along with connecting "vi" and "r" and giving a possible association of "virus". Finally, it could combine "a" and "l" into the suffix "al" and recognize that meaning.

Again, just my two cents. Not particularly sure if this is how that works.

They reference this paper on Byte Pair Encodings http://www.aclweb.org/anthology/P16-1162

they tokenize text and then learn an embedding for those tokens using (in their case) both the source and target lanuage. This presumably captures regularities for languages that are not very different and has the benefit of having a small dictionary.

An easier way to understand it is in the context morphology: word prefixes and suffixes mean things, and words have common roots.

For example, polymorphism could be decomposed into poly-morph-ism. Antidisestablishmentarianism, which is unlikely to appear much in the corpus, becomes anti-dis-establish-ment-arian-ism. Now the system can learn how to reuse "anti-" or "establish" from other examples more easily than trying to learn the full word's meaning from the one or two examples it might see in the corpus.

BPE is a clever way to induce these sort of decompositions automatically without any linguistic annotation, making them useful in multilingual settings. Other languages are much more morphologically rich than English, and there it really benefits.

This is similar to the word hashing used in deep structured semantic models. [1] This works by representing a word as k-hot vector of trigrams that form the word - it is able to generalize better to unseen words than one-hot whole word embeddings probably due to word morphology as the other answer in this thread suggests.

[1] https://www.microsoft.com/en-us/research/wp-content/uploads/...

Is Urdu a romance language? Or related closely to English ?