In the first part of our interview with Mattia Ruaro, STP’s resident machine translation specialist, we talked about machine translation (MT) in general: how it works, how it has been used at STP and what companies can do to train the MT engines they use.
In this second part, you can read Mattia’s thoughts on the newest development within MT technology, which has people predicting the end of translation as we know it: neural machine translation.
So, Mattia, what is neural machine translation? And what’s with the hype?
Neural machine translation (NMT) is essentially the same as statistical machine translation (SMT), but there is more of a “brain” behind it. NMT can potentially improve itself over time and learn on its own.
The vital difference is the amount of data an NMT engine needs – which is way, way more than a traditional SMT engine.
Essentially you have nodes that establish connections on several levels, such as the context and clause level. This makes NMT more flexible – it can analyse shorter bits of text, so the flow of the target output tends to be better.
We often joke that when you train a SMT engine, you’re training a machine. Neural is more like teaching a child a language – or bringing up a bilingual child! While the engine is learning, it makes plenty of mistakes along the way, of course.
How does NMT output compare to previous technologies?
The first thing is better fluency. The output from an NMT engine tends to be more idiomatic, meaning it reads more like natural language. More often than before, the engines are able to use an appropriate synonym or expression within the context of the sentence at hand.
Adapting to the immediate context helps a lot with languages like German or Danish that have complex syntax. Subclauses separated by commas are interpreted more accurately, for instance.
One key aspect of NMT is that it interprets morphology better. For example, a verb in the first person would usually be rendered as an equivalent verb in the first person. So, if the source says I write in English, the target would be j’écris in French, with the correct ending. If the engine cannot recognise the person, it will give you the next best thing, which is usually the verb in the infinitive (for example écrire). This is then easy to edit manually.
We talked about training MT engines before. How does training NMT engines differ from SMT and RBMT (rule-based machine translation) engines?
NMT needs a lot more data than SMT and RBMT. The biggest hindrance to adopting NMT in the first place is that smaller companies can’t find enough data. To get started, a NMT engine needs at least 10 million words of data.
By comparison, an SMT engine can be good as long as the data is good; you can get a decent SMT engine with as few as a million words.
So, NMT is much more about quantity over quality in this respect! Just to put this into perspective, our Finnish NMT engine has 140 million words right now.
Another thing is training the engine. NMT engines tend to resolve issues themselves based on data you add – they come up with rules. You can still add rules if you want, but sometimes this can be counterproductive – you risk doing too much, being too strict.
For example, a German to English translator at STP was wondering why the German-English engine was translating personal names. It turned out that these specific names were also all meaningful nouns (such as the surname Müller, which means “miller”). This means we had to consider the need for a new rule carefully, since the noun Müller (capitalised, like all nouns in German) might come up in a text about millers later.
In this case, leaving it alone and replacing the translated name manually each time was the easiest thing to do. It’s an easy mistake for the translator to spot. You see the error, you check the source and you fix the output accordingly. No one is expecting the output to be perfect.
Will NMT replace human translators?
A hundred times, no! A technology like this is only as good as the use you make of it.
I could imagine a situation where a company with several offices around the world would need internal communications, such as short messages from HR, translated very quickly. These could be run through a specialised engine the company has developed and trained for that purpose. The translation wouldn’t be high quality, but people would get the gist. But this would be internal communication and nothing customers would ever see – just for information purposes. Another example is using MT to translate large amounts of survey responses for market research purposes.
But this is not how it’s been used or how it is perceived by many. Many early adopters of machine translation have misused the technology, which has affected its reputation.
The key thing is to use MT output appropriately. Professional translators can use it as a tool. It has even been suggested that post-editing output produced by a MT engine could be a separate service one provides as a translator, as long as you know what you are doing.
Translators are not being replaced; it just that the way they work is changing.
Does NMT technology work differently with different language pairs?
It seems it has done, for some language pairs. For instance, English-Japanese is working quite well, which I find quite impressive. Nordic languages have not been concentrated on much, as they are smaller.
German output seems to suffer from the syntactic complexity and strictness of the language, and capitalisation is a huge issue. Romance languages seem to be working fairly well; NMT engines seem to cope with their verb paradigms and tenses.
Rather than the language pair, the issue is more the target language itself. Obviously Finnish has been a bit of a headache for us.
Why, is the grammatical complexity more important or is the issue the lexicon?
I think morphology is more important, the grammatical complexity within words. The engine will have a harder time discerning the different parts of a word.
The Finnish case system is a real challenge for the engines. Each case ending is a variable, and you need to consider this variable in every scenario. Finnish has 15 different cases and there are several possible endings for many of those cases, which means there are a lot of potential alternatives.
So far, I have only heard of one company making a Finnish engine work really well in the terms of the morphology and fluency. And that can only be achieved by specialising in one language.
How costly is NMT? Is it worth investing in?
Very costly. You need powerful servers to operate the amount of data we’re talking about. If SMT is driving a car, NMT is more like flying a jet – the fuel costs are much higher. It’s a lot more affordable now than it was before, though. Now more and more options are available, prices are falling.
In terms of cost-efficiency, I would say that, if used correctly, MT has the potential to really speed up translation in established workflows.
How secure is MT in general and NMT specifically? How can we be sure that personal data and other data is safe?
It is as secure as you want it to be. It depends on who deals with your engines and how. We have third-party technology, but we’ve checked their locations and their background.
We also clean the data to keep it secure so that no personal data gets used to train engines. Even Google no longer reuses the data you send back to them. For a while now, they have limited themselves to the data from Google itself rather than using the final output from the translators.
In other words, I think machine translation is very safe.
In the third part of the interview with Mattia, we will talk to him about the practice of machine translation post-editing and how translators can learn to edit the output from MT engines.