What is improved during machine translation post-editing?

With machine translation becoming increasingly popular, the concept of post-editing is entering people’s vocabulary. But in which ways can post-editing improve a machine-generated translation? What are its limitations? We take a look at the whole process below and explain what is improved during machine translation post-editing – and what isn’t.

Post-editing – what does it actually involve?

There are two levels of post-editing: light post-editing and full post-editing.

What is improved during light post-editing?

Light post-editing corrects errors of:

• Spelling: a word or a name has been spelled incorrectly or a spelling has not been regionalised (e.g. the US spelling was used when the target language should have been UK English)
• Grammar: e.g. a singular verb has been used with a plural noun, or the text is unclear as to which noun an adjective is accompanying.
• Punctuation: hyphens, commas, full stops, question marks and the like being used incorrectly or in a manner that creates confusion.
• Mistranslation: the machine translation software has misunderstood the source text and the translation means something different than the source text.

This has the advantages of being very quick and budget-friendly, but it is not suitable for all situations. The resulting text may still have a clunky, artificial-sounding style. Light post-editing is more suitable for simpler texts and for when all you need is to know the meaning of a document.

What is improved during full post-editing?

In addition to correcting all the errors covered by light post-editing, full post-editing also looks at more ‘subjective’ and ‘human’ factors’:

• Industry- or client-specific terminology: the client may prefer for a certain term to be used, consistently, in its translations, or one particular translation may be the industry standard.
• Tone of voice: if the source text is written in a formal, serious style, the translation should be as well. If the source uses colloquial expressions and slang, so should the translation. Unless this would fall foul of…
• Cultural aspects: for example, the text might refer to a person who is very famous in the source country but unknown in the target country, so the human translator will need to find an appropriate local equivalent. This is an aspect of localisation.

Is post-editing always a good solution?

Although it might sound from the above that machine translation post-editing follows the same workflow as the conventional, human translation process, the end result will not be exactly the same. Machine translation post-editing produces ‘a product comparable to a product obtained by human translation’ (the ISO 18587:2017 standard), i.e. of similar but not equivalent quality.

Post-editing is also not suitable for all types of text. If a text uses a lot of metaphors, humour, figures of speech and complex or abstract concepts, it will take a human translator to understand it properly. The translator doing the post-editing will therefore have to rewrite the translation from scratch, making the machine translation an unnecessary step in the process. The same is true of marketing translations, which take a much freer approach and focus on conveying the overall message rather than the meaning of each individual word.

There is also the issue of time. Whereas a human translator may use an incorrect word when translating a term, they will usually stick with that word throughout the translation, as they think it’s the right one. This makes proofreading easier, as all you have to is search for the wrong term and replace it with the right one. Machine translation, on the other hand, is liable to use a different wrong term every time when translating something, so the human post-editor will have to change every such mistake manually. This, combined with the need to rewrite some sentences entirely, can mean that machine translation post-editing takes longer than regular human translation and proofreading.

Getting the best out of machine translation post-editing

With this in mind, there are ways of getting the best results out of post-editing. This involves optimising quality at each step of the process, beginning with the source text. The fewer grammar and spelling errors the source text has and the more consistently it uses terminology, the easier it will be for the MT engine to translate. This might involve ‘tidying up’ the source text in advance. Try also to keep sentence length to around 20 words: anything longer is liable to confuse the translation engine.

The next step is to use a good MT engine. Neural machine translation (NMT) tools like DeepL and Google Translate are seen as the future of machine translation and can minimise the number of corrections the post-editor subsequently will have to make. Some MT engines are also better in certain language combinations than others.

Being transparent at each stage of the process is very important. Clear guidelines with the post-editor for the level of editing required are essential. It’s important for everyone to understand how machine translation post-editing works and the sort of quality one can expect from the result.

Finally, follow up the post-editing process with a review of the results, and give feedback to the makers of the MT engine. This will enable them to refine the engine itself and thus provide better translation output in the future.

SwissGlobal specialises in PEMT solutions and we would be happy to advise you on your translation projects. Get in touch with us today.