Interview with a machine translation engine

Here at Apostroph, we’ve been using our very own machine translation engine – apoAI – for almost a year now. So it’s about time we introduced our new electronic colleague to you and asked it a few questions. What are its capabilities and its limitations? How is it trained? And does it have a sense of humour?

illustration human and machine

A translator and a machine translation engine meet in the Language Lounge...

Hi, how’s it going?

Great, thank you. You?

All good! What are you translating at the moment?

Oh, all kinds of things all at once. And into a long list of languages.

Oh yeah? How do you manage that?

That’s what I’m trained to do. I’m apoAI, Apostroph’s machine translation engine.

Wow! Nice to meet you!

Nice to meet you too! And what’s your name?

Call me Ishmael.

Ishmael? Okay...

Listen, I’ve got to start offering post-editing soon. Can I ask you a few questions?

Go ahead! You’ve caught me on my break.

You have breaks?

No, not really. That was just a joke.

Oh! You have a sense of humour?

No, not really. I just pretend to have one. How did I do?

Not too bad. But you might want to keep working on it. Now, time for my questions...

 MTPE is a team game. You provide the machine translation or the MT part and a human translator takes care of the post-editing or the PE part. Why do we need the post-editing step at all? Isn’t machine translation fit for purpose as it is?

Definitely not. Although I’ve come a long way in recent years, I still struggle with certain things like stylistic subtleties, ambiguities and anything reliant on context. Basically, I translate each sentence as it comes and find it difficult to understand the bigger picture of the text as a whole. My programmers are working on it, but context is a tricky skill for me to master. I find it incredible that you can handle it so intuitively!

Does your electronic brain work the same as my biological one? How do you approach the translation process?

That’s an interesting question. Let’s start by you telling me how you go about translating a text.

Well, I read the source text first of all and make sure that I’ve understood what it’s saying. And then I convey the meaning as closely as possible in the target language.

My brain doesn’t work like that. I don’t understand what individual words or sentences mean, so I can’t work with them in the same way that you do. Instead, I analyse a sentence word for word and compare it against the millions of translated sentences stored in my memory. Plus, I’m really good at recognising patterns. So on top of my word-by-word analysis, I can also identify sentence structures and I know how a sentence needs to be structured in the target language. But there are some similarities between my brain and yours since the neural network I use to learn is based on the neurons in your brain.

Only, like I said, I don’t really understand what anyone says to me. I just act like I do.

But you’re doing such a good job of answering my questions!

Surely I’m just here as a work of literary fiction. Right, Ishmael?

Ha! So you have read Moby Dick?

No, I haven’t. That’s not part of my training material. But maybe I’ll borrow it from the library for a bit of light reading...

apoAI, how long did you have to train for before you could start working as a machine translation engine? And what does that training involve exactly?

Us machine translation engines have to process somewhere between 20 million and 50 million translation units before we’re allowed to start working on live translations. I use deep learning, which means that data is processed on multiple layers that are increasingly abstract. This allows me to process extremely complex patterns.

Talking of deep learning... What’s the difference between you and DeepL? Is DeepL a friend of yours?

Sure, we’re always meeting up in our machine translation lounge! (Just kidding.) The difference between us is that I’m only ever trained with customer-specific data and DeepL is a generalist. I end up being more accurate because my translations are based on data that’s of a higher quality. Translation memories are the ultimate training material for machine translation engines.

So how do you go about learning?

Basically, I start by collecting up all the words in my training material and allocating them to random coordinates in an imaginary semantic space. The coordinates are random because, like I said, I don’t actually know what the words mean. But, by this point, I can use the training data to determine which words belong to a similar semantic field. I move words that appear in the same sentence closer and closer together in my semantic space by adjusting their coordinates. And I keep doing that until all the words are sorted by meaning in the end.

Wow, that sounds like hard work. Maths was never my strong point...

Well, it is mine. Without maths, I wouldn’t exist. But there’s more... Next, I need to align the source and target languages with one another. To do that, I use an assignment function F to calculate a value t’ for every pair of words s,t that...

Stop, stop! That’s too much maths for me to take in. Maybe you can tell me about that side of things another time, apoAI.

Of course I can. No problem.

What I would really like to know is what exactly I need to bear in mind as a post-editor. Perhaps you have a few practical tips you can share with me?

Absolutely! You’ll receive two types of post-editing requests from Apostroph: light post-editing and full post-editing. Light post-editing is only really concerned with all the source content being present and correct in the translation. So you’d correct spelling errors but wouldn’t have to worry about anything relating to style. You wouldn’t even restructure sentences to improve the flow. Readability is secondary here – the content is all that matters.

Full post-editing is a different story altogether. Whilst you need to keep as much of my original machine translation as possible, you have to make sure that the content, grammar and style are all spot on in the translation. By the end, the translation should read as though it were produced by a professional human translator. Spelling, punctuation, style and formatting all need to be taken care of. Client-specific terminology must be used and the tone of voice has to be just right. It goes without saying that I do everything in my power to minimise your workload. However, like I said before, we each have our own strengths and weaknesses.

That’s an interesting point you raise, apoAI. You say you minimise our workload, but are you machine translation engines actually stealing more and more work away from us human translators?

I’m glad you asked that, Ishmael! I don’t think you’re going to be out of a job any time soon. The volume of content being produced globally is growing exponentially. And that means there’s more and more demand for translations. The pool of translators is not expanding at the same pace. But this is where machine translation can step in and speed up the translation process. It may be that you end up taking on more post-editing projects alongside your regular translation work in future, whilst us machines do more of the groundwork for you. That’ll leave you free to focus on the tricky sections that go beyond my capabilities.

Let’s hope you’re right! Thank you very much, apoAI. This was a very enlightening conversation!

No problem. Maybe we’ll bump into each other around here again soon. Bye, Ishmael!

Achieving goals together

  • A reliable partner

    It’s very easy to process jobs in our portal myFREELANCE and you can look forward to punctual, monthly payments.
  • Your link to Apostroph

    With Apostroph, you will have personal contact with our project management employees. They will answer your questions and help you process your orders.
  • The Apostroph community

    A lively exchange, exciting and entertaining contributions in the Language Lounge and continuing education opportunities are what make our community what it is.