METM22 Chronicles: Simon Berrill

Translation or editing? Making MT your own

Simon Berrill challenged his audience to think twice before dismissing machine translation. He talked about how he went from being sceptical of the use and value of MT to incorporating it into his translation process.

Before he started using MT, Simon would produce a very rough first draft, a second draft to refine the target text, and a third version with a read-through for minor edits, spellcheck, QA, etc.).

When Simon analysed his use of time in this process, he discovered that roughly two-thirds of it was spent creating low-quality first drafts. Since he was looking for ways to craft even better translations without having to invest more time, it seemed logical to try to cut down on the time-consuming production of those preliminary target texts.

This is where MT proved helpful: after testing the MT engine DeepL with some of his translation work and finding that the MT output was not substantially worse than his own first drafts, Simon decided to let DeepL handle that first step. As a result, the time allotted to the different stages of the process changed significantly, with the first draft now consuming only 3% of the time. More changes were necessary for the MT output in the second step, but this was not an issue since the first step saved so much time (time that could also be used for producing an even more polished third – and sometimes fourth – draft).

Simon continued by giving us several Spanish-to-English translation examples to illustrate where MT delivers useful – and sometimes surprisingly good – results, and where it does not:

Benefits:

  • MT sometimes comes up with word choices the translator might not have considered.
  • With certain paragraphs, MT is able to “tidy up” the structure of rather messy source text sentences.
  • MT often does a good job rearranging sentences and moving elements to a more natural position in the target language.

Drawbacks:

  • As DeepL often uses different translations for the same word, the translator has to be especially careful to ensure consistency.
  • MT can’t recognize gender or verb forms with left-out subject pronouns, which can lead to mistranslations.
  • MT frequently delivers words that are slightly – or completely – off.

It is crucial to watch out for these types of errors in MT output when working on the second draft. However, due to the significant time savings mentioned, there is plenty of time to find the right terms, work on word order, etc.

This means that MT allows a wordsmith to do more wordsmithing and more extensive research, which, Simon noted, results in better texts and increased satisfaction with his work.

Before concluding, Simon briefly addressed the question of whether what he does is translation or editing (it’s certainly not post-editing). He still classifies his work as translation – with an element of editing in it – and wonders if translation and editing are effectively converging.

The presentation was followed by an interactive exercise: the Spanish-into-English translators in the audience were given a short paragraph to translate, while the non-Spanish speakers received a machine translation of the same text to edit. We then compared the results and discussed them briefly. Afterwards, Simon answered questions, and a few audience members shared how they use different MT engines.

Thank you, Simon, for an excellent presentation on a seldom-discussed topic! You have demonstrated that machine translation actually can be used to produce higher-quality texts. Let’s hope that more translators openly discuss MT and their use of this tool, which would benefit all of us!

This METM22 presentation was chronicled by Sabine Holz.

Featured photo by METM22 photographer Jone Karres.

One thought on “METM22 Chronicles: Simon Berrill

  1. A lovely, balanced account, Sabine!
    As to whether using DeepL in the manner Simon describes is translation or editing, I would argue for the former. The *process* is essentially the same as the one I (and many others) followed in the MIddle Ages: Mentally arranging a whole sentence in my head and then writing down the translation, examining it for correctness, and then typing it when satisfied. One point Simon touched on but did not state explicitly in his presentation (aside from the warning to watch out for MT-type errors) is that the key to using DeepL effectively is to *know better* than the machine does and, obviously, not to rely on it as the only path towards producing a decent, accurate target text.

Leave a Reply