Editing Machine-Generated Texts

Machine gibberish? No, thanks!

Machine translation systems certainly have been improving in recent years. But being able to imitate human language is not the same as possessing human-like intelligence. Machines do neither feel nor think like us. They don’t know empathy, they don’t giggle about silly dad jokes, and they don’t wince when you yell at them. This lack of emotions and lived experiences often makes their texts sound lifeless and dull.

In some scenarios, this isn’t a big deal. A simple user manual or a city council’s latest ordinance isn’t expected to win comedy awards. However, if a machine-generated translation elicits unintended laughter and makes your company go viral for all the wrong reasons, this is a problem.

To make sure this won’t happen to you, I offer a range of so-called MTPE services. These are in line with the relevant ISO standard and follow the principles outlined in my AI policy.

Overview

What is “MTPE”?

MTPE is short for Machine Translation Post-Editing and refers to the improvement and polishing of machine-generated translations. Sometimes you may also find the abbreviation PEMT, short for Post-Editing of Machine Translation, and if you’re dealing with German texts or business partners, the abbreviation MÜ for “maschinelle Übersetzung” might pop up here and there. And because this isn’t confusing enough yet, certain marketing teams of translation agencies and “AI” vendors have now introduced the new term Artificial Intelligence Post-Editing (AIPE). But that’s just another instance of too much hype. At the end of the day, “AI” systems are still machines – there’s no need to muddle the terminology further. Thus, I’m sticking to the established abbreviations, MT and MTPE.

These services come in two types: “full MTPE” and “light MTPE”. The former means doing a complete revision of the machine-generated translation with the goal to create output that is “accurate, comprehensible and stylistically adequate [and] indistinguishable from human translation output” while trying to “use as much of the MT output as possible.” This definition and additional requirements are set in the ISO 18587 standard: “Translation services – Post-editing of machine translation output – Requirements” (2017).

In contrast, the much simpler “light MTPE” approach only aims to facilitate information gisting. A typical example: In an international company all employees should be able to read and understand internal policies regardless of their native languages. In such a scenario, the translations don’t need to be stylistically adequate, but they should still be comprehensible and accurately convey the main insights of the source material.

IMPORTANT: If your text is meant for publication, “light MTPE” is the wrong approach!

With the usual tools: 80 % translation + 20 % polishing

With MTPE method: 20 % translation + 80 % polishing

A common misconception is that MT content will only need “a few quick fixes.” Sure, when you have a very simple text or only need a rough translation of internal files for informational purposes (“light MTPE”), this could work. But the word “rough” is key here: Poorly edited MT suffers from incorrect terminology, loses the source text’s subtleties, and often fails to reflect the original author’s style and tone of voice. Your readers and customers deserve better!

This is why “AI” rarely saves time

For years, language professionals have been using tools that speed up the translation process and lower costs for end clients without compromising quality. This type of workflow is known as Computer-Assisted Translation (CAT). Typical CAT software comes with termbases (to ensure consistent translation of terminology), so-called translation memories (to allow the efficient re-use of previously completed translations), and various little helpers such as an auto-suggest feature (to avoid the repetitive typing of similar words) – in short: CAT software can be quite a time saver!

Now, when clients pre-translate texts with an “AI” system – whether it’s ChatGPT, Amazon Translate, DeepL or a similar product – and hope to reduce translation costs to a third or even a tenth of the regular rates, it often ends with disappointment. Because I will then have to compare the MT with the source text, move or replace words, fix inconsistently translated terminology, and get rid of silly mistakes a professional human translator would never have made in the first place. (For some examples, see “Human vs. machine”.)

The bottom line is that a translation done by a skilled pro using CAT software, keyboard shortcuts, and other tools doesn’t take much longer than improving the flawed output of a machine. Plus, in cases where it makes sense, I could actually integrate an “AI” translation engine into a CAT workflow myself (always in line with my AI policy and only after consultation with you). Thus, it is important that you manage your expectations and don’t expect unrealistic discounts.

Suitable Content

Certain types of texts can indeed benefit from machine pre-translation and be processed faster. For example, when I started out as a translator, I often handled product datasheets with highly repetitive content. This is just the sort of monotonous typing that machines are welcome to take off my hands. Other typical examples include:

User manuals
Recipes and similar instructions
Common descriptions for hotels, flights, rental cars, restaurants, etc.

Generic T&C documents
Simple course/e-learning content
Menus, error messages, and similar components of software/apps

Of course, translations that are generated automatically should undergo a thorough review to make sure both content and form are fit for purpose. And once again, please note: A professional translator’s toolbox already includes suitable tools for these types of texts, which means current “AI” systems will rarely yield significant gains in productivity.

When MTPE causes additional costs …

Problems typically arise when machine translations aren’t used in an expedient way but only serve as a shortcut to save money. Ironically, the total costs can increase due to a poorly planned MTPE strategy!

Consider the following example: An e-learning company had hired me for an MTPE project in which English course content was to be machine-translated and then iteratively improved by multiple language service providers. The translations would be made accessible to paying (!) users of the platform right away. Despite the German version being labeled as a beta feature, this was problematic from the start: The content was supposed to be suitable for a premium offering but shouldn’t be polished before publication? Not a good idea.

The next obstacle came in the form of vague requirements. During the first round, translators were told to fix “only major errors” that might cause frustration among learners. But this is subjective. Some people get annoyed by nested sentences, others are bugged by inconsistent terminology or constant switching between formal and colloquial style. In addition, the client was hoping for a throughput of up to 2,000 words per hour – this doesn’t leave time for diligence.

In the second round, another translator would review and polish the initial translator’s work. Such double-checking is a good idea, but: The second translator would be new to the content, and there’d still be various mistakes the first translator didn’t fix due to being pressed for time and without clear instructions. Hence, this second person would have to start from scratch but with an even tighter allocation of hours.

Finally, the third round focused on clicking through the published German version on the e-learning platform to make sure everything was properly translated, formatted, and correct in context. However, missing context is something that should have been clarified immediately in the first round (without context, you wouldn’t be sure whether something was correct or a mistake). Unfortunately, one of the translation agencies hired by the e-learning company for this project turned out to be of the dubious kind. They failed to raise this issue and instead just skimmed all the texts.

Long story short: This e-learning company ended up paying repeatedly for the same service and wasting a large chunk of money.

Let’s do the math for illustration. Assume a typical course on the platform has 10,000 words, and the translators all charge $50 per hour. Considering the throughputs desired by this client, we get the following effort required:

– 1st round: 5 hours (2,000 words/hr)
– 2nd round: 3.3 hours (3,000 words/hr)
– 3rd round: 2.5 hours (4,000 words/hr)

This results in: 5*$50 + 3.3*$50 + 2.5*$50 = $540.
If the client had opted for one proper round of full MTPE instead, with a feasible throughput of about 1,200 words/hr, they would’ve paid just 8.3*$50 = $415!

Payment and Terms

Similarly to other types of editing, I charge for MTPE services on an hourly basis, with my standard rate being €50/hour. When the MTPE approach is used in a meaningful way and with suitable texts, about 1,000–1,500 words per hour can be considered a realistic throughput (full MTPE). I’d be happy to take a look at your specific project to assess the potential effort and benefit of MTPE.

Payments can be made via SEPA credit transfer or, if you’re outside of Europe, via PayPal. My standard terms are 14 days, and my Terms & Conditions apply.

Decades of research – still not done

The history of machine translation did not begin with Google or OpenAI. In fact, the first practical considerations regarding the automatic transfer of texts from one language into another go back almost a whole century! In the 1930s, an engineer and a scholar, Georges Artsrouni and Peter P. Troyanskii, independently filed a patent application for a mechanical translation machine (book rec: Early Years in Machine Translation).

However, it became clear this was a rather lofty goal, which would require at least some foundational research first. But despite a good amount of enthusiasm in the 1940s and 1950s that brought about various scientific papers, conferences, and new academic chairs, the mood would quickly change. For a simple reason: Human language is complex, multi-faceted, and often ambiguous. It cannot be described and translated using only a basic set of deterministic rules.

And so this research field evolved and went from purely rule-based approaches to a combination of statistics and language models to predict word sequences, until it eventually embraced machine learning, neural networks, and big data for training more powerful models. Of course, all those decades saw significant progress. Long gone are the days of jumbled Google Translate output riddled with unidiomatic phrases and grammar errors. And yet, after so many years, machines still need help from qualified humans to create precise and consistent translations tailored to the target audience.

Frequent Questions

Your question isn’t answered here? Please, also check my FAQ page or drop me a line via my contact form. Thank you!

...

…

...

…

...

…