Using neural machine translation software in CAT tools: installation, cost, performance

As a long-time user of both computer-assisted translation (CAT) tools and machine translation (MT), I was very interested to attend Jennifer Gray’s session on neural machine translation (NMT) in CAT tools at METM24. And I was not disappointed; the session was highly informative and practical.

The session opened with a review of the current offering, focusing on Google Translate, DeepL, ModernMT and Microsoft Translator, followed by a step-by-step description of the integration process for two CAT tools, Trados and memoQ. This was complemented by a handout with links to guides on how to subscribe to the different MT services and install them in your CAT tool.

The simple explanation of what an application programming interface (API) is and how to obtain one was apposite. I did once, after a great deal of difficulty, manage to obtain an API from Google, but I had never understood what it really was or how to manage it. Now I think I understand the process a little better.

We then heard a detailed explanation of how to subscribe to each NMT service, the cost structure for each one and advice on the factors that might influence our choice. For me, the interesting new piece of information was the existence of a service called “Custom MT”, which allows the user to access multiple MT options for a single subscription fee.

Jennifer went on to discuss the factors that might influence our decision to use these tools: confidentiality, cost and performance.

Confidentiality. All four companies assert that the texts submitted to them for translation are not stored and are never used for training purposes. Despite these reassurances, the general feeling seemed to be that most people using them do not place enormous faith in these assertions. In other words, if your text is highly sensitive or confidential, the best way to preserve that confidentiality is to not submit it to any online service.

Cost. The cost of a subscription ranges from €20 to €25 per month. However, Google and Microsoft are essentially free for individual users, because the first 100,000 and 400,000 words, respectively, are free. The advantage of Custom MT is that it costs €25 per month but gives access to over ten services.

Performance. Are MT services worth the time and investment involved? The answer is complex and depends on the languages being translated, the subject matter, the type of text and the preferences of the translator. These factors were discussed and we saw the results of one study in which MT performed well for translations on education and healthcare, and poorly for colloquial writing and entertainment.

Finally, we heard about the results of an interesting study of the four tools undertaken with the help of four MET members, who evaluated translations from four source languages (Spanish, Finnish, French and German) in four domains (medicine, literature, engineering and law). In the opinion of the German-to-English legal translator, none of the MT services provided any target language text that was fit for purpose or useful. In the other three languages and domains, DeepL scored highest in overall quality and medicine was the field in which all four provided the best result, albeit with plenty of mistranslations. All the translators indicated that “post-editing” would require careful comparison with the source text to check for errors and mistranslations as well as editing to improve style and register.

The slides and handout for this excellent presentation will be very helpful to anyone considering using NMT in a CAT tool.

This METM24 presentation was chronicled by Ann King.

Featured photo courtesy of MET. Presentation slide reproduced with presenter’s permission.