New Google Translate languages | Hacker News Reader

New Google Translate languages

techcrunch.com

66 pointsanjalik23 comments

The neural network translation is so fickle, it can fail to accurately translate a three digit number. It gets worse with larger numbers, with errors by orders of magnitude.

https://translate.google.com/#en/ja/six%20hundred%20sixty-si...

https://translate.google.com/#en/ja/eleven%20thousand%20one%...

Interestingly I wasn't able to reproduce it with English/Russian translation, maybe they hardcoded it somehow, or maybe neural network for Russian hasn't rolled out yet.

It worked for me in languages other than Japanese. Might be a language-specific problem.

Before using neural networks, Google used to translate "백조 한 마리" (one swan) into "100000000000001 swans", because the Korean word for "swan" can be also interpreted as "one hundred trillion". So, "one swan" = "one" + "one hundred trillion" = 100000000000001. And then stick "swans" onto it for good measure.

Translation is hard.

I've been using it for English/Russian for the past couple years, haven't noticed any change... it's still pretty awful and makes the same mistakes.

They are supposed to transfer to neural translations today might not have transfered over yet. I used to follow a Russian forums because the developer of the software I used was on them. Would be interesting to see if there are any changes to the translations.

Looking at the alternative suggestions on both of those is interesting. For the first, the dash gets interpreted as a "to" as in "six hundred sixty to six". For the second, there's actually a correct translation also suggested!

This title is misleading. The languages Google "added" have already existed; what Google has done is run translation sentence-by-sentence instead of by phrase-by-phrase.

Is the algorithm (including e.g. network topology) published somewhere?

Alternatively, is there some open/published work with comparable performance?

Well, yeah: https://arxiv.org/abs/1609.08144

Note that for comparable performance, you probably need not only the network topology but also the training data...

"In addition, the Neural Machine Translation system learns over time and improves, resulting in better and more natural translations the longer it works."

Dumb question (maybe?). How is it learning since it's not getting any feedback on what is the right and wrong answer?

It is getting feedback, you can correct its results.

But who does that on any scale?

Occasionally, I submit my translations/validations via Community page: https://translate.google.com/community#en/ky

I would, if they'd not store my contributions locked away and assume ownership. Same with Google Maps, they have a contribution feature which I'd use if you could download Google's map under a similar license to OpenStreetMap.

For Gmail and stuff I get it, but translate and maps are data projects, not software projects, which take hard work by thousands of people to build. It's extremely hard to build a competitor, and so I refuse to contribute to the existing de facto monopoly.

There was one time that I tried to submit a correction to Google Maps anyway, not out of generosity but just to fix a navigation glitch, and the team that handles that data was quite rude about it.

I was trying to point out that if you're in Cambridge, MA and you voice search for "transit directions to Harvard Square", you should get directions to the well-known public square and commercial hub, not a mall parking lot in Charlestown.

After I tried to convey this through an interface that had no option for "you're failing to give directions to this extremely relevant place", they testily told me that the mall data was 100% correct. A team of people who had probably not been to the Boston area questioned who I was to tell them what is or isn't named Harvard Square.

Eventually I got it fixed by Google's real bug report process: bringing it up in conversation around people who worked at Google.

I am really hopeful, that some day they will give a link to download your contributed translations. But anyway, this service is one of the ways to help your language to prosper.

> give a link to download your contributed translations

Well, that wouldn't really help anyone. They would need to publish all translation data to help anyone make a competitor, which is not in Google's interest.

It is being fine-tuned, not retrained I guess. Which boils down only changing some layer weights and keeping others constant.

A billion users, one edit at a time.

I would assume that it's not learning from feedback from its own translations, but just from more and more text on the web.

So "is still learning" just means that it's still adding to its corpus of examples over time.

AIUI, part of the system is unsupervised, i.e. is trying to learn "representations" for data independent of any translation task with a right/wrong answer. Unsupervised learning is a hot topic in machine learning.

Presumably they can feed it human translations of the same text and let it find the correspondences. With their book scanning activities they must already have a good data set to work with.

Reply on news.ycombinator.com