The concept of NMT has radically changed the translation industry over the past few years. Thanks to deep learning, today’s NMT engines are capable of storing huge amounts of information about words and sentences from training texts and then using this for the translation process. The result is fluent and increasingly natural-sounding translations. But what modes of communication and types of texts are suited to machine translation and which aren’t? And will the resulting translation require LPE (which corrects only the most serious errors) or FPE (a more in-depth revision process)?
Let’s start off with four instances where NMT is suitable:
NMT is a good option for texts that are intended to be used within a company, such as internal emails, minutes and intranet news. LPE is sufficient for internal purposes; all that matters is that the text can be understood by everyone who reads it.
Texts created by users or customers – such as product reviews, feedback and comment sections under posts on company blogs – can also be translated with NMT. These also only require LPE to serve their informative purpose.
NMT allows you to efficiently handle large amounts of text. FPE should be used when it comes to fundamental product information such as ingredients or key data. This requires a more extensive edit because, for example, foreign units of measurement need to be checked. Product information should not be confused with product descriptions, which are written in a far more descriptive and emotionally appealing manner, and are therefore not suited to NMT.
NMT is also ideal for user manuals that are purely informative. As is the case with the other examples, an AI trained via deep learning provides the necessary basis for these translations. This is the only way to guarantee factual correctness. User manuals also require FPE to ensure that they are free from errors.
Now let’s talk about the instances where machine translation is not the best way to go:
Creative texts, especially slogans and claims, are often riddled with puns, idioms, irony and other rhetorical devices, with the aim of appealing to the customer emotionally. In its current technological state, NMT is not capable of reproducing such stylistic strategies. Marketing texts include newsletters, blogposts and general posts related to a specific campaign. These kinds of texts require more than just a simple translation: they call for a transcreation, a creative adaptation of the text aimed at the relevant target market.
Although a translation memory can be fed with specialist knowledge in the form of a terminology database, the technology isn’t quite advanced enough to correctly translate the convoluted sentences found in complex texts. This is partly due to the issue of text cohesion: machine translation can fluently translate individual sentences, but doesn’t take into account the relationships between them. It lacks the context of the whole document, which is why it can’t consistently translate technical terms across multiple sentences. Machine translation is also incapable of using specific terms selectively or explaining them when they appear for the first time.
Websites and apps
A website needs to be easy to navigate, from the homepage down to the very last subpage. The same applies to apps on devices. Textual links need to be created and a certain user experience established – yet another area that machine translation can’t deal with in its current state.
Speeches and presentations
The written versions of speeches, e.g. for a company event, business presentation or pitch, are similar to marketing texts in terms of the rhetorical devices they use. Emphases, irony and creative puns would be lost in machine translation – and with them, the impact of a speech.
In conclusion: NMT is the right choice for texts that don’t need to appeal to emotions. The standard of quality expected of the target text is therefore lower, as it is intended purely for informative purposes. The quality of machine translation always depends on how thoroughly the technology has been trained: the machine first needs to be fed source texts and trained with them for the deep learning to be successful. For creative texts and texts with a specific structure, whether within one document or across several, you should leave the translation to a human translation expert.
Cover image via Pexels (CC0)