Peeking into the architecture used for Google's Neural Machine Translation | Hacker News Reader

Peeking into the architecture used for Google's Neural Machine Translation

smerity.com

81 pointseternalban9 comments

Hey HN, I'm author of the article. If there's anything confusing, I'd be happy to help :)

The GNMT architecture is likely to keep popping up in Google's papers due to the amount of engineering that's been dedicated to make it scale well. Other than chewing up massive machine translation datasets, from single to multiple language pairs, it's also been trained over the entirety of Reddit for a conversational model system.

Just this week a new paper[1] was released using the GNMT architecture which teaches a machine translation system how to translate between language pairs it has never been taught before by using bridging from other language pair knowledge. It's all pretty fascinating stuff :)

[1]: https://arxiv.org/abs/1611.04558

Great article, thank you for getting to the crux of the paper. Now that you have a base case working:

* Could you convert the parameters that govern the creation and interconnect of each layer into lisp and feed the whole thing into a genetic algorithm to evolve even better translators? Using GAs to evolve NNs was the new hotness in the late 90s and I don't know what ever happened with it. It would take on the order of 10,000 times more computation to simulate all the generations, but hey, it's Google.

* How hard would it be to generalize the GNMT for solving other types of problems where the output translation is the solution to a problem or what to do next? While reading the article, it occurred to me that the hidden layer S could be thought of as the meaning of the input data. If S is smaller than the input data, then you could treat it like a word feeding into another GNMT to process paragraphs, chapters, etc. I guess this would hinge on whether sentences have any relation to one another. In any case, would love to hear your thoughts on something like that.

I'm just an armchair hobbyist toiling away my career building CRUD apps, but I really feel that if we can get beyond the learning curve (no pun intended) and get to basic building blocks like this, then engineers will run with it and scale it very quickly to a passable implementation of hard AI.

Does anyone know of any good primers on building your own machine translation? My wife is from a small European country, and all machine translation I've used usually end up with incompressible results. My theory is that it's not that there is anything hard about the language, it's just they don't put as much effort into training this language as the number of users is small - so it ends up with poor results. As an experiment I'd like to try building my own, but I don't know where to begin.

https://www.tensorflow.org/versions/r0.11/tutorials/seq2seq/...

If the European country she is from is within the EU, you can probably get parallel corpus from the EU parliamentary translations. I believe there are other EU research projects which release useful corpus.

Sadly one of the largest tranlation corpus of multiple languages can't be used for copyright reasons i.e the user made subtitle files that are shared all over the internet. Every week a few hundred episodes of different shows are released in different countries that are translated by users into many languages. Now that I think about it an app that could transcribe the words spoken in a TV show and then would translate it using google translate would be a huge with the manga crowd. Though Japanese hasn't been added yet.

It's actually not at all clear if they couldn't be used. Things like Google's word2vec model are derived from copyrighted data and seem to be fine.

Perhaps distribution of the subtitles is illegal under copyright law, but I suspect the trained model would be fine.

It would be an interesting case anyway.

This is a directory of available multilingual corpora:

http://opus.lingfil.uu.se/

Poor results are probably lack of enough data rather than lack of enough researcher attention.

NMT requires massive amounts of data to give even okayish results.

If it's an EU language, then it shouldn't be caused by "they don't put as much effort" as there are a bunch of datasets and research projects that, as a policy, work on all 24 EU languages.

These systems aren't always publicly available for anyone (e.g. http://ec.europa.eu/dgs/translation/translationresources/mac... is freely available for gov't institutions only, https://www.letsmt.eu/Login.aspx has commercial pricing), but the freely available resources (e.g. EU corpora) tend to be well integrated in the main free MT systems (google/bing/etc).

What you see there is just a bit worse than the maximum currently achievable - you can get a better system for a particular narrow domain (which is often done in practical applications) by integrating domain-specific data and terminology, but for general purpose translation don't expect to get anything significantly better than e.g. google translate even for a small language unless you can invest many man-years of work - e.g. a major funded research project for that particular language.

Regarding "My theory is that it's not that there is anything hard about the language", there are two factors that you must keep in mind. First, the current (pre-neuron) mainstream approach of phrase based SMT works much worse if the target language is highly inflectional, as many European languages tend to be. Second, the main available large training data sets for European languages are mostly legal/government documents, so the systems work better for that domain but much worse on "common" topics since they're poorly covered in training data.

Reply on news.ycombinator.com