MEINRAD's Blog

Machine translation for technical documents: can it work?

Written by Meinrad Reiterer | 8 April 2021

Irina Höll MA, MSc is a technical translator, interpreter and language engineer. She explains why machine translation (MT) can be both a blessing and a course, how to introduce it and use it for technical documentation, and what you can do to make your texts better suited to MT engines.

 

You’re a technical translator, a trained technical editor and an expert in machine translation. You’re also a lecturer in the Technical Documentation course at the Joanneum University of Applied Sciences in Graz, Austria. Could you tell us a bit about your professional background?

I trained as a conference interpreter and translator. Since 2013 I have been an In-House Linguist at Häupl-Ellmeyer, a patent law firm in Vienna, and I’m also a freelance translator and interpreter for German, English, Spanish and Swedish as well as a consultant language engineer. As I was getting more and more requests to translate technical operating instructions, in 2017 I decided to enrol for the “Technical Documentation” course at the Joanneum in Graz. My master’s thesis focused on machine translation and technical documentation, and it eventually turned into a sort of machine translation handbook for technical editors. Ever since, I’ve been advising clients, translation agencies and other linguists on how to go about introducing MT.

 

Machine translation has become widespread in recent years, and it’s now an established part of life for the technical editors and translation teams at many businesses. Would you say machine translation is a blessing or a curse?

Lots of my translator colleagues get hot under the collar about this, but I’m pragmatic about it. It’s new technology, and it’s here to stay. As a translator, there’s no use trying to fight it or pretending it doesn’t exist. We’re a very long way off being made obsolete by machine translation – machines can’t yet replace humans when it comes to communication. MT is only a curse in the sense that it’s now so widely accessible.

 

In what way is that a problem?

Basically, everyone can now use machine translation. And they use it without thinking about it. Then they wonder why the translations they get are so bad, or why using it doesn’t actually save them anything like as much time and money as they thought. Suddenly, this technology has a bad reputation. So many people don’t realize that machine translation simply isn’t suitable for every text.

 

So which texts are suitable for machine translation?

In my experience, texts where the purpose is to engage as much as to inform the reader – marketing texts, for example – are not suitable for machine translation, because there’s too much post-editing involved. That includes all types of advertising and marketing texts, texts that speak to the reader on an emotional level in some way, feature specific cultural references or contain sensitive content. By contrast, operating instructions are usually well-suited to MT. In terms of the language they use, they’re ideal: the phrasing is (or at least should be) simple, and the aim of the text is simply to inform. But what matters most in technical documentation is that MT should never be used without post-editing. Otherwise, you’re just shooting yourself in the foot.

 

What do technical editors have to focus on when producing texts which will later be machine translated?

If you want to use machine translation for your texts, ideally you should think differently in terms of content creation for your technical documentation. There’s a golden rule for the language you use: keep it simple, clear and consistent. The more modular and repetitive it is, the better, especially when you remember how translation memories make it extremely easy to reuse texts. Make sure you spell everything correctly and phrase it sensibly, which means keeping sentences short – it’s been shown that the quality of MT outputs gets worse as sentences get longer. Use the active rather than the passive voice, and resist the temptation to use humour, unusual words, dialectal variations and abbreviations. It’s also better to use fewer pronouns and more definite articles. That might sound obvious and straightforward, but in practice you need the discipline and self-reflection to rigorously implement these aspects and produce a text that will make the translation process go more smoothly. Authoring systems are worth considering, as they can help make texts more suitable for MT engines.

 

The managers of a business might decide that from now on machine translation must be used for all documentation in order to save money. What’s your advice for technical editors and in-house translators in that situation?

I recommend taking a fundamentally positive approach. Tell those managers: “OK, we can try – but only under certain conditions!” Before going ahead, everyone should familiarize themselves with the issues involved in MT, such as which texts are suitable for it and which aren’t (ask your translation agency if you’re unsure). Then analyse your own content, and think about the economics and about what you’re aiming for. Do you need high-quality, publishable texts, or do staff just need a rough idea of what internal memos are saying? And don’t forget the risks. Managers can have unrealistically high expectations in terms of the time and money saved and the quality of the texts that will be produced, and this is one of the biggest obstacles that often makes the introduction of machine translation and post-editing a failure in practice.

 

Why is it important to carry out a risk assessment first?

Lots of things can go wrong with translations, and with machine translations in particular. At best, bad translations will make people laugh or shake their head in disbelief. Often, they’ll damage a company’s image – after all, the translation is part of the product, and if that part of the product is poor, people’s opinion of the product as a whole will go down. Even more seriously, the customer might not be able to commission the machine and can’t produce anything, or they won’t maintain it properly and end up damaging it. But the absolute worst-case scenario is personal injury. Manufacturers are liable for poor-quality translations, and that goes for machine translation just as it does for human translation. With an even higher risk. So I really can’t say this often enough: you should not use machine translation without professional post-editing. MT without post-editing means you won’t meet the necessary standards. You don’t know what the engine will spit out, and neural engines in particular frequently leave bits out, add bits in and distort the content of texts. And you won’t notice those errors if you simply use the output as is.

That’s why it’s important to carry out a risk assessment before you begin using machine translation. Ultimately, the question to ask is this: is the money you might save worth the increased risk?

 

Let’s say your texts are suitable for machine translation. Could you grab the bull by the horns and start using MT straight away?

You could, but it isn’t a good idea. In many cases, as soon as business managers hear that MT can save their business money, they want to go full steam ahead with it. But lots of them don’t realize that a fair amount of groundwork is required in order to save money with MT while continuing to produce high-quality translations. So it’s important to closely analyse the current situation in terms of language management at your business before you give machine translation the green light. If it’s not very sophisticated – for example if you don’t have any translation memories or terminology management systems, or if you do but they aren’t maintained properly – my advice is to focus on human translation for the time being. There’s no point testing MT until you have solid foundations and smooth language management processes in place: if things aren’t working properly with human translation, machine translation won’t solve the problem. It might be cheaper in the short term, but over time you won’t produce the high-quality texts you need and you won’t save money either!

 

Has the role of technical editors changed through the increasing use of machine translation? Is it possible that translating documentation will come under technical editors’ remit in future?

It could happen, but I’m sceptical. I think operating instructions should always be written by native speakers. Of course there are technical editors who speak multiple foreign languages very well, and there’s no doubt that they’re more familiar with in-house company terminology than a translator who has never done anything for that company before. But that’s only really true for the first job, and translators learn incredibly quickly. So my view is that everyone should stick to what they know best – again, we’re talking about minimizing risks and maximizing quality.

 

Apart from the quality aspect, are there other risks if you’re looking to use MT to boost productivity when translating technical documentation?

Yes, there are a few. The biggest issue is cyber privacy. Again and again, people are far too careless when using machine translation. For example, non-native speakers who write operating instructions in English sometimes just copy a sentence or paragraph into Google Translate or DeepL when they can’t think of how to phrase something. That’s a bad idea! The text might contain sensitive product information, and often it’s the kind of information they’re forbidden from disclosing to anyone as part of their employment contract. But that’s exactly what they’re doing when they use free MT engines. Many people also use MT to translate e-mails from their colleagues – I’d advise against that too, because you don’t know in advance what information the e-mail will contain. So you should be very careful. It’s all too easy for information entered into free MT engines to end up as results from search engines or to be used as training data by other companies.

 

If you’re thinking about working on MT with a translation agency, how can you tell whether the agency has experience with machine translation? Where do you start?

One thing to check is ISO 18587 certification, the post-editing standard in conjunction with ISO 17100. Another is more simple: communication. Do they ask you questions? They should want to know which content needs to be translated, roughly how much they can expect, and what the risks are. And you should also find out which engine they work with, whether training data will be available, who’s responsible for maintaining them, and how long it will take. Do they have an in-house MT expert you can contact if you have any questions? Is there someone to guide you through the whole process, or are you on your own? Who are their post-editors, and what training and experience do they have? Can you send your texts as they are, or do you need to adapt or rewrite them for each job? A professional translation agency will be by your side at all times, and they should also offer to run test projects. Then you can judge whether MT gives you the quality and value for money you need, or whether you’re better off without it. Another possibility is to produce a style guide that helps technical editors write texts suitable for machine translation. And high-quality translation agencies will be willing to discuss the issue with your management – they can often supply cost-benefit calculations that help convince the decision-makers of the benefits of a professional approach to MT.

 

That sounds like a pretty time-consuming process. Is it?

Yes. Introducing machine translation always involves a lot of groundwork, and it may take a few months until suitable processes are in place and all the requirements are met. You can’t just start using MT and expect to get the same high-quality output as you would from human translations – it’s a false economy. But it will pay off if you take the time and make the effort to get the full process chain up and running first.

 

Apart from potentially having to adapt source texts, what else do you have to focus on before starting to use machine translation?

MT engines can now be trained, and one way to do that is with translation memories. TMs should be “spick and span” – i.e. they shouldn’t contain any terminology you don’t want. Otherwise post-editing will take longer. And of course that applies to the term bases too.

 

What other advice do you have for technical editors and in-house translators?

There are two main things to remember:

  1. Don’t shy away from the latest technology. Use it, and above all, make sure you understand it. That will pave the way for high-quality translations that give you full value for money.
  2. Don’t neglect the groundwork. Do the risk assessment, get your translation management processes in place, and clean up your translation memories. 

Apply a bit of common sense, be careful, adjust your expectations and shore up your knowledge of MT, and you’ll quickly go a long way towards successfully integrating it into your language management processes. You’ll soon see the benefits: better value for money, higher productivity, quicker turnaround times, shorter time to market, new customer markets, localization of previously untranslated content, faster and easier-to-understand communication between staff around the world... the list goes on. Sounds good, right?

Main image: © Irina Höll