Google’s ‘Babel fish’ heralds future of translation
In Douglas Adams’s famous Hitchhiker’s Guide to the Galaxy series of science-fiction books, interstellar species use Babel fish — “small, yellow, leech-like” creatures that feed on “brain-wave energy” — to translate speech in real time.
A team of developers at Google is working on the real thing, using statistical models to translate different languages, including Afrikaans, on the Web and on mobile phones, using voice input and output as well as text.
TechCentral sat down with Google Translate research scientist Ashish Venugopal at Google’s headquarters in Silicon Valley last week and asked him about the stumbling blocks to effective real-time translation and the future of the technology. This is an edited transcript of that interview.
TechCentral: How many languages does Google Translate now support?
Ashish Venugopal: There are 63 languages supported. That’s a lot of languages. How do we get all that data in there? If we tried manually to give the system those languages, it would be a hopeless task. The only possible way we could do this is to harness the power of machine computation. We build statistical models that are automatically training themselves and learning all the time. As people translate new content on the Web, our systems pick this up and it adds the words. The system is constantly reading and analysing the Web. It’s a statistical approach. The idea is that once we learn the essential model of how to speak a word, and we can apply that to every word. We haven’t memorised every word.
Are there any difficult languages that make it hard to get translation right?
Yes, there are some incredibly tough languages. If your language is very different from English, for example, then it will be very difficult to translate it to English. We use English as an intermediate language and so if you were translating from Russian to Japanese, we’d translate the Russian to English and then to Japanese.
When we talk about a “tough” language, it’s one that is really different compared to English. There are languages that are very different in multiple dimensions.
The first question to ask is, what is the order of words from in one language compared to English. In English, we’d put the subject first, then the verb and then the object, whereas the Japanese have the subject first, then the object and finally the verb. We have to teach computers how to recognise this reordering pattern.
We don’t tell the computer how to translate every sentence. We give it general patterns to look for. When it sees new data, it uses those patterns, matches that to data and then comes up with a model that it uses to translate sentences.
When we say languages are harder, they’re harder because of the ordering of words, they’re harder because there may be different notions of what a word even is. In English, you say you put the phone on the table — “phone” and “table” are objects and “on” is an additional word that explains what’s happening. In other languages, the “on” could be glued onto the word “phone” or “table” and we have to teach the computer that “on” could be connected to the object or be separate from it.
All these issues get easier when there’s more source data. We launch languages when we feel they are adding value to somebody. We have “alpha” or experimental languages where we were just able to launch the system, as opposed to it being fluent and correct. The alpha languages tend to have less source data available online.
What are the main stumbling blocks to this technology and what will be possible in the future?
We are really reliant on the source data. The first stumbling block for a new language is, is there data on the Web? Once there’s enough content on the Web and as we build our system … on average it works really well. On average, you’ll be very impressed with it. But every once in a while you’ll be irritated with it.
Because of the statistical approach, you may enter something and get some crazy translation. What we are trying to do is limit those crazy translations and ensure in all cases we are providing a reasonable translation.
This really comes from the fact that this is a statistical system. We’ve built it so you can literally put anything into it. We will translate anything you give us. It might be good or it might be bad, but on average it will be quite impressive.
What we are really working on now is clipping the bottom end of the cases where we make mistakes. We see these issues in languages that are very different compared to English. Russian, for example, adds a lot of information to words and they get longer and longer and when we translate we sometimes make mistakes there.
In the future, in a reasonably short time, we will take machine translation for granted, as part of our everyday lives. I mean that from an 80-20 standpoint, where 80% of the use cases we’ll be able to address effectively. The last 20% will be incredibly hard. That speaks to the fact that machine translation won’t be a substitute for a human translator.
No one is going to take an important political speech and put it into machine translation to publish it in 20 different languages. Our goal is not to create artificial intelligence; our goal is to provide an 80% solution where you’ll be able to understand the political speech’s point, but not it’s rhetoric, not it’s beauty necessarily.
Is the future of this technology instant voice translation using devices like mobile phones to facilitate real-time translation of conversations?
We can do that already, but not simultaneously. It’s not an immediate goal. It’s a matter of where we are focusing now. There’s still more work to be done on the quality side before we can start to develop this continuous form of operation.
Will you continue to do translation in the cloud (online on servers) or will it move down to devices like phones as they get more powerful?
We make all our decisions purely based on quality. We want to ensure the highest quality translations are delivered to our users in the shortest possible time and that’s leaning towards the cloud for now, but that might change.
What sort of computing power does Google Translate require?
We use the full power of Google’s search engine. The reason Google Translate exists is because of the investments made in search. We sit on top that search infrastructure.
Do you have a team of linguists working all over the world?
We have a team of statisticians, all working right over there [points and laughs]. It’s less linguistically orientated. There are linguistic ideas that influence our decisions. To give you an example, when I was working on the last set of Indian languages that were launched, I didn’t use any linguistic knowledge; I used Wikipedia and my grandmother. So, it’s Wikipedia, my grandmother and statistics. That’s what we use to put a language together. — Duncan McLeod, TechCentral