Machine translation (MT) has become indispensable in the translation industry, and translators are increasingly becoming post-editors as well. Slovenian translator Vito and French translator Guillaume discuss their views on MT and the experiences they’ve had with artificial intelligence.
How long have you used MT, and why did you start working with it?
Vito:I got my licence for Google Translate in 2011, so I’ve been familiar with it for over a decade. At the time I’d been working as a translator for seven years. In the early years I endured lots of “teething troubles” in the profession, and I realized that as a linguist you need to invest in technology, training and the back office to remain at the cutting edge and stand out from the crowd. For example, we had TRADOS – it cost lots of money, but we could use translation memories to translate quicker and more reliably, which meant we could be more flexible on the price. The switch to MT came later, but the case for it was the same.
Guillaume:I’ve used MT since summer 2017. In my case, I found out from colleagues that this new technology was now on the market, and I was desperate to try it. At the time I was working on a long technical text relating to mechanical and systems engineering, and I wanted to see if MT could be used for these kinds of jobs and if it really could help me translate faster.
You’ve now done lots of MTPE (machine translation + post-editing) jobs, so how does post-editing differ from translating? Which do you prefer?
Guillaume: I think the two processes are fundamentally different. When you’re translating, you have to “invent” the entire sentence structure from scratch. When you’re post-editing, by contrast, the sentence structure, style, grammar and specialist terms are all there already for you to review and correct if necessary. But the amount of research involved is the same. I definitely like both methods, but you have to be aware that MT can’t – or shouldn’t – be used for every project.
Vito: For me, post-editing is like proofreading. And if I had to describe the difference between proofreading and translating, personally I find proofreading tougher: you read and read, and everything’s OK, but you have to stay alert! But if the text you’re proofreading is badly written, it’s a slow process – sometimes it takes longer than if you were translating it yourself. Machine translations were awful in the early days, but now they’re somewhere between good enough and very good. I’ve known for a while now that my role in the process is becoming less important and less valuable.
Could you give us a rough estimate of what percent of your jobs are now MTPE jobs?
Guillaume: I’d say it’s now 50:50, but it’s moving more and more towards MTPE.
Vito: For me, at the moment it’s 20% human translations and 80% MTPE jobs, and in both cases they tend to be short texts.
How do MTPE jobs affect your earnings, bearing in mind that they help you work faster?
Vito: In my experience, I can deliver jobs more quickly, but for less money. I need between five and ten times as many jobs before I earn the same overall. But I don’t have the time to constantly look for new jobs.
Guillaume: Luckily, I haven’t yet noticed any difference. Provided that the texts are suitable for MTPE, my experience is that I’m about 40% quicker. If post-editing is paid fairly – i.e. 60% of the translation price – there’s almost no impact on what I earn thanks to the higher number of MTPE jobs I can take on. But one problem I often have is clients who assume that a particular text can be post-edited when it’s not suitable for it. They sometimes underestimate the level of specialist expertise required or the amount of research involved in order to translate the text. And the inevitable result is that these kinds of jobs aren’t worth my while financially.
Which engines do you use? Are there differences in the quality of the output?
Guillaume: I work almost exclusively with DeepL, and I have my own Pro licence. Otherwise I use the output given to me by my clients in the respective CAT tool.
Vito:I only use Google. Clients have recently started using LanguageTool more and more, but I find it hard to compare them.
What are the engines good at and less good at in your language?
Vito: Slovenian is a language with relatively complex grammar. None of the MT engines have been able to handle the Slovenian dual yet, and they often struggle with verbs in context, as they give them a “masculine ending” even if the feminine or neuter form is needed. The formal/informal second person form gets mixed up when translating from English to Slovenian in particular, the gerundive form is used far too often rather than a subordinate clause, and the long genitive chains regularly used in German cause problems. Another issue worth mentioning is the random use of synonyms instead of translating terms consistently, especially with LanguageTool.
Guillaume: For French, machine translation engines – mainly DeepL, which as I say is what I use most of the time – handle the syntax and spelling really well. But they aren’t as good at the style and punctuation. And of course they can’t read between the lines, so they’re completely out of their depth when it comes to cultural idiosyncrasies and the nuanced meanings and connotations found within a text. Plus, DeepL translates from English, which means if the source text was in German, that text is translated into English first. Machine translations from German tend to be less good, so I have to spend noticeably more time post-editing them.
What do you think are the risks with MT?
Guillaume: One major risk is that we pay less and less attention to the source text, and potentially don’t make enough adjustments to the target text as a result.Just because a target text sounds good doesn’t necessarily mean that it’s “correct”, i.e. that it’s an accurate translation of the source text. And especially when the results are getting better all the time, there’s a risk of being blinded by good output if you spend hours doing nothing but post-editing. I think another problem is that MT might tempt us to disregard the translation memories and term base entries, though personally I always look closely at them. Then there’s the issue of privacy. You have to be very careful and only work with Pro versions.
Vito: I think the main risk is that we translators will become redundant. And there’s an equally serious risk of the MT output no longer being post-edited at all.
Which types of text do you think MT is suitable for, and which types of text cause problems?
Vito: MT is definitely not suitable for literary texts, advertising and marketing texts, or video subtitles. But I recommend it for operating and assembly instructions.
Guillaume: That’s my experience too – I’d also say that MT is suitable for technical texts with minimal specialist jargon, IT texts and content from international legal sources which has already been translated umpteen times and is available online – EU directives, for example, have been translated and published online in various places. I’d even say that there are some texts, such as technical texts, which I wouldn’t like to translate without MT anymore.
MT is less suitable for texts where a message needs to be conveyed in a particular way, rather than just the bare content. You can forget about MT for advertising and marketing texts, and I keep finding that even newsletters are often difficult to post-edit as the MT output is more of a hindrance than a help. As a post-editor, it’s very tempting to just use the style produced by the MT engine rather than phrasing things “nicely”. And texts with content that isn’t translated or published very often, such as national legislation, are less well-suited for MT – these legal texts often can’t be translated in full as the laws of the target country may be different. The translator’s job is to find a legal equivalent and explain what the different terms mean.
For example, a bailiff in Austria (“Gerichtsvollzieher”) isn’t the same as a bailiff in France, where they have additional responsibilities such as carrying out surveys in the event of disputes. If you just translate the word, an Austrian wouldn’t understand why the bailiff is suddenly doing that particular job.
Which of your clients mainly order MTPE, and which sectors do they come from?
Guillaume: It’s mainly clients or agencies in the technology, medicine and IT sectors.
Vito: The percentage of MTPE jobs is steadily growing, both from smaller clients who aren’t as familiar with the issues involved in translation and from established clients and translation agencies.
What’s your data security policy? Are you given instructions, especially if you’re working with free MT engines?
Vito: I think this is an issue for clients to think about.
Guillaume: Direct clients have never asked for machine translations – so far I’ve only had MT jobs from translation agencies, where I see the MT output in the CAT tool straight away, so I don’t have to worry about data security. If I use MT with private clients, I only work with fee-based engines such as DeepL Pro, where I know data security is taken very seriously.
Have you completed training on what you need to do when post-editing? Are courses available?
Guillaume: At university, we didn’t talk about machine translation or post-editing at all, but that’s probably because at the time the technology was very new and not widely used yet – as far as I know, universities do now explore the issues involved in MTPE. So personally I had no professional training for it, though three years ago a translation agency ran a training course on the subject which I enjoyed taking part in. Since then it’s been a constant process of learning by doing. I’ve actually learned the most from reviewing post-edited texts. Sometimes I come across mistakes which have clearly been produced by the MT engine, and I think “I might have fallen into that trap too”. In other cases, the agencies give me guidelines for what they expect from post-editing.
Vito:The client usually gives me a list telling me what I need to do, and I follow that. I don’t think you need any specific training for post-editing.
Finally, let’s look at what the future holds: do you think that human translators will eventually disappear? Where will artificial intelligence take us?
Vito: Maybe I’m too pessimistic about it, but almost everyone now has Babelfish and Google Translate on their phone. MT is the first hint of what we can expect from artificial intelligence in future: it’s gradually becoming as smart as we are, and soon it will be even smarter. So I think AI and MT will have an impact on language services. We’ll have to wait and see where MT will take us, and to what extent human translations will still be needed. MT means bypassing the intermediate stages in the process of getting products from manufacturers to consumers, and the chain between clients and the service providers is becoming shorter and shorter. Agencies are part of this chain, just like us translators – will they survive?
Guillaume: I see it differently. I don’t think human translators will disappear any time soon. Machine translation is based on deep learning, which means the MT engines use existing data sets (published translations) and probability algorithms to try to predict what a sentence might mean. So these tools rely on mathematics, and they don’t understand what they’re saying – I reckon that will be the case for a long time yet. I think MT tools will become increasingly sophisticated and will be used more and more in our everyday work, but as long as there are texts where the message plays an important role, it’ll be a long time until we become obsolete as translators and reviewers.
Main image: © Adobe Stock