Our second blog in the current series is on the use of machine translation. Tim Branton discusses the current state of automated translation and why it may – or may not – be applicable to your business’ needs.
Ian:We’ve all seen comically inept and bizarre examples of machine translation that have been created when copy is entered in to Google Translate, for example. Is machine translation improving, and should businesses consider the use of it more often?
Tim:There’s a long history of machine translation going all the way back to the 1950s when it first emerged as a concept, and I imagine that at that point it was assumed that it would be fully done and dusted by about the 1970s – but of course that didn’t happen.
There have been three key phases of machine translation, but it’s only been within the past five years or so that there have been really significant breakthroughs. Phase 1 was what is called ‘rules based’, which was based on the mistaken assumption that you could write a set of rules that would allow you to combine those rules with a huge dictionary, for instance from English to German, which would allow you to automatically translate content. What was found in practice was that languages are just far too complex, with too many differences within each language’s structure and syntax, and that there are too many exceptions in each language which precludes creating rules.
Phase 2, originating in the ‘90s, was called ‘statistical’ machine translation, using what was essentially a statistical guess as to what was most likely to be the correct translation in a particular scenario. This led to some improvements and machine translation got better, but it still wasn’t really useable at an enterprise level; this was what was employed in the early days of Google Translate.
To come back to the question “is it getting better?”, the answer is an emphatic yes, it’s really getting a lot better.
The big breakthrough (Phase 3) has happened with the explosion of computing power that we’ve seen more recently and the introduction of cloud computing, particularly from the late noughties onwards. Available computing power has increased so substantially that Google and others have been able to go a big step beyond the statistical approach with ‘neural machine translation’, or NMT. This is using machine learning as a way of taking a lot of previously translated content and using it to train computers with something called a ‘machine translation engine’. The machine translation engine has already ingested all of these previous translations and begins to learn many different ways of translating things that are similar to your content.
This does a better job for languages which are syntactically similar and it does a better job for certain types of content, such as technical documents and legal documents where there is lots of repeated or very similar content that uses consistent terminology.
It performs worse with language combinations, such as translations of English to German, because the syntax is so different between the languages, and it does quite a poor job in dealing with creative content; introducing humour is a good way for it to go rather horribly wrong.
To come back to the question “is NMT getting better?”, the answer is an emphatic yes, it’s really getting a lot better. You can prove this to yourself by typing something in to public facing Google Translate, where you’ll see that quite often you get rather spookily good results. This simply reflects the fact that something very similar or almost identical to your copy has already been human translated.
The important point to understand with NMT is that human translation is the foundation of machine translation. It’s machine learning, and all machine learning is based on work humans have already done, with the machine determining “OK, this seems to be the way this is done”.
Ian:So this is effectively a really substantial database that the machine can apply to the content you enter, comparing like to like?
Tim:It’s doing something more complex than that because the machine is actually learning from all of the content that it’s been fed, but that process obviously isn’t done in a single pass. You start with a trained translation engine using your existing translations, and you then go through a process teaching it further. The engine is told “that was wrong” and the machine will learn more from that, learning from the corrections of its translations; it will learn from the approved terminology that’s introduced. Very often if the term is in English and a search is made for a translation of that term in to German you’ll find there are multiple translations of that one word, but in a particular context there may only be one correct translation and the alternative translations that are presented are completely wrong in the context relevant to your business. The ability to teach the machine things about context is really critical.
With the large body of existing translations available to it, the machine is able to make better guesses as to what the correct choice is for a particular term and it’s better able to make choices about an appropriate syntax for each sentence that’s being translated.
Ian:With this awareness of the vastly improved ability of neural machine translation, how does a customer know if it’s an option for his needs?
Tim:The simple answer is: try it. Talk to us or someone else who offers machine translation and see what kind of results are achieved.
There are a few things which will help determine whether it’s likely to be of value to you. An obvious point is that the larger the amount of content that you’re translating on an ongoing basis, and the larger your corpus of existing translations, the more it’s likely to make commercial sense as far as budgets are concerned.
Another factor is how narrow your subject matter is. For example, if you are a software company focusing on a specific area – let’s say CAD or accounting software – you’re likely to get better results with a relatively small corpus of existing translations.
The larger your corpus of existing translations, and the narrower your subject matter, the more likely you are to get good results from NMT.
If you’re spending a million dollars a year on translation then you really should be looking at machine translation, but there are other factors that should influence your decision as to whether it’s worth looking at. I mentioned earlier the importance of the type of content being translated: if technical documents are being translated you should be looking at neural machine translation, but if it’s creative marketing content, right now, you’re much less likely to get results that are acceptable.
If you’re in a scenario where you have a choice between machine translation and not translating something, which very often is the case, what choice should you make? There are arguments that in some cases it’s safer to leave content untranslated because you have confidence in your original English content. For content in a highly regulated sector it could be a serious mistake to do anything other than a full quality translation, but in most cases content which is consumer facing for which you have a choice of utilising machine translation or not translating, the consumer would almost certainly like to see a machine translation rather than no translation.
This leads on to a second point, regarding user expectations. People are quite used to seeing machine translated content, and as long as it’s not presented as human translated content (i.e. you’re trying to fake it), people will accept machine translation that’s not quite ‘right’. For example, when reading product reviews in places like Amazon or Air BnB which are translated to your language by machine translation, what you as a consumer are interested in is whether the review says something is good or bad, or whether an Air BnB apartment was clean or dirty – a sophisticated, syntactically perfect translation isn’t needed. To be able to see a rough machine translation that tells the reader whether the reviewer was happy or not with the apartment or the product is the core of it. By not providing this information you’re not dealing with that doubt the consumer has at the point of decision – so why wouldn’t you do it?
Ian:What’s the significance of the idea of ‘criticality’?
Tim:I mentioned the question of whether or not content relates to a highly regulated sector, and there is also a question as to whether something is commercially critical to you, and whether there is a safety aspect or a legal aspect to it.
To go back to the user reviews scenario, by definition that content is written by an end user and it isn’t your content; all you’re trying to do is make your end user’s content intelligible in German in a rough way for your other customers. If something is critical it requires that the appropriate effort be made.
It’s not just a digital choice between machine translation and human translation because in practice the majority of machine translations are enhanced with the work of a human editor
Another important thing to be aware of regarding machine translation is that it’s not just a digital choice between machine translation and human translation because in practice the majority of machine translations are enhanced with the work of a human editor. That approach gives you a head start because you hopefully end up with a good first pass at translating the content, which the human editor will then go through and correct. You’re ultimately aiming for something which is as close to a high quality human translation as possible; the better trained your machine translation engine is means the machine translator will get the human editor a long way towards achieving that full quality human translation.
The flip side to this is that if you start with difficult content for a difficult language, and you haven’t got a big corpus of existing translations, it creates a scenario in which the human translator that the machine translation is put in front of will likely say “this is so far wrong it’s much easier for me to scrap the machine translation and start from scratch”. There’s always a judgement that needs to be made: is it good enough in certain instances to put that content directly in front of your end user, or is it something that really requires editing by a professional translator – if it does need a human editor, the machine translation needs to be good enough that it saves the human translator a significant amount of time, or what was the point in doing the machine translation?
Ian:Finally Tim, “how much can it save me?” is a key question that will be asked by just about anyone contemplating using machine translation.
Tim:I was talking at the start about this immense computing power that we have available to us; it’s got a lot cheaper but there is still a cost associated with training a machine translation engine. This depends on how much content you’ve got, as the more content you’ve got the more it’s going to cost you, and also the more it’s going to save you. You should almost think of it as a capital cost up front of getting going with your custom machine translation engine, introducing it to your existing translations and your existing terminology.
The translations from that point will be provided at something like a 95%+ saving versus human translation. If you then introduce human editing into the process that will introduce a significant amount of cost into it. In that scenario you may be saving let’s say 40%, but that’s a very variable number as it depends on type of content, and on how close you need it to be to a completely perfect translation at the end of the process, so there are numerous factors that will determine how much you actually save.
If the machine translation isn’t good enough there was no point in doing it because it will take the human translator as much time as it would have if she had just sat and done it. You’ve also introduced a risk factor in to the process; any time you present a translator with something that’s already been translated, particularly something that’s syntactically correct, there’s always a danger that there are mis-translations kind of hiding in plain sight.
So, to conclude: a translation really has to be good enough that it provides a jump ahead and is giving a human editor something that’s going to save her a significant amount of time versus translating from scratch.
One final thought: pretty much all the biggest corporations on the planet are using machine translations, but even Google uses human translation for the most important and tricky content. So the real point is, get a strategy that works for your business.