It is not an overstatement to say that Machine Translation (MT) revolutionized multilingual communication in many fields. But while the quality of MT is high and always improving, it’s imperfect enough that the US Department of Health and Human Services (HHS) proposed the addition of a rule to the Affordable Care Act. That ACA rule would require a “qualified human translator” to review MT output for content that is “critical to the rights, benefits or meaningful access,” to good healthcare outcomes for patients when accuracy is essential.
Review of MToutput by a “qualified human translator” is called post-editing. When text is entered into MT* software and a translation is returned, a post-editor then edits the ‘raw’ output to prepare it for use in any number of products — brochures, manuals, websites. A project manager will choose from three degrees of processing, depending on what the text is for:
- Leaving the MT “raw,” unaltered
- Light post-editing to ensure that the ‘gist’ of the message is conveyed
- Full post-editing, which attends to formatting, style, semantics, and precise terminology, resulting in an error-free, publishable final product
There are numerous situations in which raw MT is not helpful on its own because the text it produces isn’t clear and precise enough. Related to the HHS example above, the biggest problem for MT with medical translations is not difficulty with complex medical jargon, because with repeated use all words can become part of the MT software memory base. No, in fact the larger obstacle is that, “The context-dependent nature of common words in specialized health and medical domains… is causing subtle yet clinically significant errors and confusion.”** This MT weakness reflects that unique ability that only human translators possess, which is to understand the elements that are not embedded in the words: context, nuances, and the potential misunderstandings that make a slightly different word choice a good idea.
Another perspective that helps explain exactly what post-editing is and how it partners with Machine Translation is to look at the way it is actually used with MT content. For instance, in order to get the best possible MT outcome, a post-editor often does pre-editing, too. In the pre-editing process post-editors work through text to make it more likely that the MT software will handle it without mistakes: they manage terminology, shorten sentences, apply style guides for consistency, and reduce ambiguities (such as double-negatives or a confusing tone like sarcasm). All of this pre-editing reduces the number of potential stumbling blocks and helps the MT software do a good job producing quick translations.
Still there are instances when MT may not offer enough benefit to play a key role in a project. The development of localized marketing slogans and translation of poetry are examples of translations to which the creative capacity of the human mind offers value that MT can’t replicate at any level.
The result of the MT/Post-editing partnership is that for many types of documents more translation can get done faster, with greater consistency, and at the level of accuracy appropriate for the content. Skrivanek’s post-editors most commonly work with documents processed through our in-house Neural Machine Technology software (NMT). Neural Machine Technology has been available since 2016-2017 and its introduction to the world of translation brought a significant increase in efficiency and accuracy to LSP work. NMT functions more like the human mind does, relying on “deep learning” that is based on the text it is fed.
Skrivanek NMT is a sophisticated resource that reduces translation cost, especially as compared with 100% human translation that is done from scratch. The investment we have made into its development and the training of translators to use it has paid off a hundred fold in terms of error reduction, money saved, and our customers’ satisfaction with final content. For documents containing repeated phrases, industry terminology, long passages of technical or legal narrative, Machine Translation is essential to best outcomes. It is an amazing tool that slashes time, tedium, and costs, and our experienced linguists and post-editors are trained to maximize its benefits.
J. V. McShulskis
* There are three types of Machine Translation: Rules Based, Statistical, and Neural, Neural Machine Translation (NMT) being the most advanced.
** Wenxiu Xie, Meng Ji, et al. in the International Journal of Environmental Research and Public Health