How To Tell If MT Post-Editing Is Right For You

MT post-editing is here to stay. If it’s faster or cheaper than human translation, there’s a case for it. The stakes are high, but the budget and quality risks are real. Here’s how to turn the odds in your favour.

It’s official: it’s OK to admit you’re considering machine translation.

Everyone else is. Even your translation agency is coming round to it. Of course, we all saw it coming. It’s been on the cards for years. Now, the technology and quality have improved to such an extent that even Google Translate isn’t terrible anymore.

But machine translation (MT) systems weren’t designed with translators or creative content in mind. They’re aimed at businesses producing so much content that human translation isn’t viable. It’s too expensive, too time-consuming, too slow – or all three at once.

So before you decide whether MT post-editing is right for your content, here are a few things to bear in mind.

MACHINE TRANSLATION: THE BASICS

Machine translation is a fast-moving field. It pays to keep up with the latest developments.

There are three main approaches to teaching a machine to use language:

1. Rules-based: grammatical systems take centre stage
2. Statistical: huge volumes of text provide the basis for pattern analysis
3. Neural: the new kid on the block mimics how biological neural networks operate

Some systems, like Moses, are open source, while others like Google Translate are proprietary. Either way, most systems use translation memory (TM) functionality to increase re-use of validated translations.

But machine translation is never a free ride. Every MT system requires specialist knowledge to use effectively. Plus, although output quality is continually improving, each system has its own blindspots. Rules-based MT lacks style. Statistical MT can be inconsistent. And neural MT struggles with rare words.

The success of an MT post-editing project doesn’t just depend on your system’s quality and the training data you’ve used. There are three other important factors.

1. SOURCE/TARGET RELATIONSHIP

Languages can be grouped together in families – like the Romance languages. MT works best when source and target languages are closely related.

For example, English has a subject-verb-object sentence structure. So does French. This makes the translation easier to process than Japanese, where the word order is usually subject-object-verb. (Although frankly, there are many reasons why Japanese is difficult for machines.)

2. SOURCE TEXT SUITABILITY

It’s also important to consider whether the source text is suitable for machine translation. MT is most effective when it’s being used for repetitive text like software documentation and product descriptions. These work best when the source language is carefully controlled for consistency. The less variation there is between texts, the less the machine needs to learn, and the more reuse you get out of your translation memory.

At the other end of the scale, user-generated content is completely unstructured. Each person writes differently and they might not follow formal grammar rules. In this instance, a high-quality machine translation will be almost impossible. However, if the reader won’t wait for a human translation, a rough gist can be a good compromise.

Booking.com was an early adopter of neural machine translation. The hotel aggregate site has 29,181,235 listings in over 40 languages. Human translation is mission impossible. Instead, the company’s language teams concentrate on editing, leaving the machines to do the heavy lifting on property and room descriptions.

Hotel descriptions with controlled language are perfect for a hybrid MT/TM system. But for customer reviews – where speed trumps quality – unedited machine translation is an option.

3. THE HUMAN FACTOR

So where does this leave human translators?

They’re a crucial part of any machine translation project. As post-editors, they review, edit and re-translate the text as needed to meet agreed quality standards.

Of course, those ‘funny’ MT errors are a lot less hilarious when it’s your job to sort them out. To make the switch from translating to post-editing, a change of mindset is needed. It means no longer seeing MT as the bad guy, but as another tool in your translation arsenal.

Your human team is also critical in identifying repeated mistakes in the raw output of the translation. This feedback allows your MT provider to incrementally improve the output quality. And to save your post-editors from the nightmare of repetitive change syndrome.

If MT post-editing doesn’t sound quite right for you, interactive translation prediction (ITP) can be a viable alternative. ITP acts more like an auto-complete feature than an MT engine. For translators using tools like SDL Trados Studio, this approach might feel familiar. Studio already combines productivity aids like fragment recall, AutoSuggest and adaptive MT. Time will tell which approach wins out.

COST CONTROL

Low output quality equals slow post-editing. Slow post-editing means higher costs and a later delivery date. So far, so obvious. But how can you predict the level of work needed when faced with thousands of words?

The short answer is you need to take a long view. Look at SAP. The business software producer understood that input quality determines output. So they built their own MT system, using their terminology and language. This allows them to offer automated, verified translations directly to their customers.

Of course, such an investment isn’t within reach for smaller or younger businesses. But jumping into MT without testing means all bets are off. An unwavering focus on quality ratings and throughput is essential to keep the show on the road. Testing must begin long before the first project. Only give the green light when the output consistently meets quality and throughput targets. Your budget will thank you. (So will your post-editors.)

FINAL THOUGHTS

Machine translation is so much more than Google Translate. It relies on language resources including glossaries and translation memories. It’s even a default component in many translation platforms.

At Wordbank, we use machine translation as part of a menu of translation services. We understand that sometimes it’s the only workflow that makes strategic sense. But we always show our cards and put human expertise at the heart of our work.

Have you got an intractable problem with translation volumes and costs? We know how hard it can be to get an MT post-editing programme off the ground. Get in touch – we’re always happy to help.

Photo by Lenin Estrada on Unsplash