Google Brings Offline Neural Machine Translations For 59 Languages To Its Translate App (techcrunch.com)

Google is rolling out offline Neural Machine Translation (NMT) support for 59 languages in the Translate apps. Some of the supported languages include Arabic, Chinese, English, German, Japanese, Spanish, French, and Korean (TechCrunch has a full list of the languages in their report). From the report: In the past, running these deep learning models on a mobile device wasn't really an option since mobile phones didn't have the right hardware to efficiently run them. Now, thanks to both advances in hardware and software, that's less of an issue and Google, Microsoft and others have also found ways to compress these models to a manageable size. In Google's case, that's about 30 to 40 megabytes per language. Users will see the updated offline translations within the next few weeks.

Comments

pivot language?

[epine]

Is English considered to be the pivot language, or do all of these models product the same intermediate representation?

Rather useless article, with no shred of a deep understanding, whatsoever.

I'm guessing you run the input model from language to IR, and the output model from IR back to language, so you need to have at least two models to use this app. (I suppose you could translate from English to IR and back to English again, for perverse joy.)

Only I haven't read anything about training multiple machine translation models with a shared IR. That strikes me as technically difficult, and I would have thought I'd have seen some loud crowing out there, had it been achieved (it's now been a couple of months since I gave the Internet a good shake on machine learning, and things move fast).

Re:pivot language?

[Lenbok]

I'm not sure to what extent it relates to the specific offline translation modules in the translate app, but a while back the Google Research blog had a post on multi-lingual machine translation models (and that let them do translation between two languages for which they didn't have direct translation training corpus). So at least in that case, there is just a single translation model rather than separate input and output models that go to and from an IR.

https://ai.googleblog.com/2016...

Re:pivot language?

[AmiMoJo]

I don't have that detail but I can tell you my experience as a user of machine translation for 15+ years.

Originally it only really worked on formal documents, and even then only produced something you could barely understand. The biggest issue seemed to be that it didn't understand context at all.

Google made some early improvements in making the translated text sound more natural. They also managed to fix a lot of common phrases that didn't quite fit the standard grammar model and thus didn't used to get translated properly. Apparently they did that by using the web as a resource for natural language and by allowing users to submit corrections.

Then AI started to be used. Baidu were the first I think and their Chinese/English translation was a huge improvement over everything else. It seemed to work slightly better going from English to Chinese though, and when Google released their AI updates not long after Chinese to English became nearly perfect.

It's actually incredible how good it is now. Often the resulting translation is not only accurate and seemingly context aware, it sounds like something a person might actually say. You don't have to think about what you are writing either. Before you had to be careful to phrase things so that the software could understand it, but not any more.

There are still some issues, like the way Japanese newspaper headlines often get translated as if it was a person speaking about their own experience (e.g. some houses were flooded, but the translation is "my house was flooded" because the software assumes that context), but for conversations between two people it's like Star Trek or something.

Re:pivot language?

[AmiMoJo]

Try using Google Translate and Bing Translate on a random story from srad.jp. Srad used to be Slashdot Japan before the name change, and the story summaries are written in an informal tone similar to Slashdot ones (but better edited!)

Google:

Ministry of Economy, Trade and Industry, Internet mail order order due to erroneous judgment of smart speaker will not be contracted

First of all, about electronic commerce using smart speakers, smart speakers have the ability to order orders to net mail orderers by voice. However, when an order occurred due to misunderstanding or misunderstanding, guidelines on how to handle that order were not shown. In this revised bill, it is clearly stated that "contract through AI speaker has not been established" for misrecognized orders, and it is said that businesses must properly deal with these problems. Also, even if the ordering party makes a mistake, if the system is such that confirmation is not made for the order, the ordering party may be able to argue the invalidity of the contract.

You can see some trivial mistakes, like how two different words in Japanese are translated into the same English word "misunderstanding", but the meaning is clear and things like the ministry name are correct and the sentences are actual English.

Bing:

Due to misjudgment of transdermal production Ministry said, smart speaker e-store purchase contract would suppose

There are features for e-commerce using the first source to speaker smart speaker audio to Internet mail order companies order allows. But the positives and say mistakes in order occurs, the order what to do of the guidelines was not shown. We're in this amendment, and describing "the agreement through the AI speakers has not been established" in probable order operators must respond properly on these issues. In addition, says is possible if the system check do not for the order if the buyer did mistakes, the officials can claim contract invalid.

It's like something out of the 90s era Bablefish. Not only is the interpretation of the original Japanese poor, but the resulting English sentences are broken too.

maybe

[phantomfive]

I'm not sure the new model is actually better than the old model. In recent months, I've seen it make bizarre mistakes, like translate "man" as "woman" in contexts where there was no room for mistake. Also it translated 10,000 as a million. Something is wrong with it.