Machine translation with post-editing

Automatic machine translation systems such as Google Translate and DeepL have achieved a high level of language proficiency with the help of artificial intelligence (AI). As a consequence, the importance of machine translation (MT) systems is increasing, especially for day-to-day translation requirements. A survey by Common Sense Advisory in 2016 found that, even back then, 80% of language service providers were already offering MT. While user-generated content can be machine translated without further intervention (which is the case for e.g. eBay or Amazon), more demanding tasks still require subsequent processing to improve the results. Called post-editing (PE), this is where a human edits and revises the translation generated by the machine. PE combines traditional translation skills with MT (which these days tends to be output from neural machine translation or NMT). PE brings the final target text to a reader-friendly, professional level.

Actually, another editing intervention can be carried out even before the text is translated by machine: it's called pre-editing. Pre-editing is the editing of the source text before MT. Here, the source text is read through in its entirety and checked for typing errors and the like: these are corrected so that the text is easily recognised by the machine. This is particularly worth doing where a text is to be translated into several languages.

Machine translation is the automatic translation of texts from one language into another by a software system. Google was offering a statistical translation system as early as the 2000s. The integration of artificial intelligence, in particular an approach called "deep learning", into the technology known as neural machine translation has now led to MT reaching an advanced level, such that it can be impossible to distinguish simple machine translations from human ones. Neural MT represents a remarkable advance in the effectiveness and relevance of machine translation applications. To evaluate translation quality, algorithms such as Bleu Score are used to measure the similarity of machine translations to human reference translations. The reliability of Bleu and other evaluation systems has come in for criticism, though: some professional systems in use today produce translations rated poor by Bleu, but excellent by experts. Accordingly, the search is now on for alternatives – which are difficult to find. Unlike evaluating something like hardware providing a precisely measurable performance, arriving at an accurate assessment of a translation is no easy matter. The issue is all very subjective, as evidenced by the fact that no two translators ever translate the same text the same way – apart maybe from short sentences.

Important aspects of PE
A test conducted earlier in 2020 found that the DeepL and Google Translate translation tools produce the best outcomes. That said, these tools can generate undesirable results such as double negatives and wrong translations of names, to name but two. Post-editors therefore need to know that the texts they are working on have been generated by a MT tool. They will then keep an eye out for completely different errors than if they were dealing with a human-translated text.

Another challenge of MT is that it often contains stylistic errors that are not immediately obvious: being linguistically polished and ostensibly correct, they are not apparent from a superficial read-through. Even the best translation machines can produce misleading results.

Do your texts require a high-quality translation because they're intended for a long shelf life or for external purposes? If so, our recommendation is to opt for PE to the ISO 18587 standard. However, PE is not simply proofreading. It involves comprehensive editing by professionally qualified and native-speaking proofreaders. The end result resembles a traditional translation: the style and terminology are consistent and tailored to your requirements.

Are you under time pressure, or are your texts for in-house use only? Then our recommendation is to opt for PE "light". This is "lighter touch" PE by professionally qualified and native-speaking proofreaders, yet still much more than simple proofreading. They ensure that your text is clear, correct and complete, but don't consider its style or client-specific terminology. PE "light" only corrects obvious mistakes and aims at a clear, though not necessarily stylistically appealing, text.

It's important to clarify what is expected of the target text. The intended purpose determines whether or not MT should be used – and, if so, the level of PE – or whether the translation should actually be carried out by a qualified translator. Depending on the type of text and/or language combination, PE can be more complex and therefore more expensive than a human translation carried out from scratch.

A post-editor or specialist translator can use the new standard "Post-editing in accordance with ISO 18587" for the "post-editing of machine-generated translations". This describes how the human translator works on the translation produced by the machine in order to obtain a linguistically correct final version.

Requirements for PE
MT and PE can sometimes save money compared to the traditional translation process. But they are perhaps of greater relevance for time-critical projects: MT with PE often allows large amounts of data to be translated more quickly.

The amount of time required for PE is an important factor in assessing whether or not the use of MT makes financial sense. PE is more complex and time-consuming than proofreading a human-generated text. Whether "light" or full PE is required for a machine-generated translation ultimately depends on the quality of the raw translation: use of a professional machine translation system is recommended to ensure the quality of the language. Ideally, this kind of system is trained on previous translations and available technical terminology.

PE "light"
For texts with a short shelf-life and internal communications, PE "light" can be enough to reach the target audience and convey the information. Furthermore, the quality of the source text plays an important role: incorrect or poorly written texts are less well recognised by the translation system and placed in the right context than common phrases.

PE according to ISO 18587
The ISO 18587 standard defines the requirements for the PE of machine translations. The standard is published by ISO (International Organization for Standardization) and is therefore binding worldwide. The purpose of the ISO 18587 standard is to provide greater transparency for consumers and users of translation services. Amongst its requirements are that language service providers deploy specialised translators (technical translators well versed in the subject area) for PE.

Texts edited in accordance with ISO 18587 are tailored to the end user and must precisely fulfil the function of the text in the target language and, if necessary, withstand the critical scrutiny of specialists in the relevant discipline.

To sum up, it would be fair to say that machine translation translates large volumes of text quickly and efficiently. But one thing quickly becomes apparent: since language is such a complex phenomenon, machines will probably never be able to translate completely error-free and in the appropriate style. Even the most powerful computer needs a human to understand, check and correct its output.


Yes, I am interested in receiving news about Apostroph. Please send me your newsletter.