Are we finally on the brink of machine translation catching up to human translators? Recent coverage of neural machine translation (NMT) seems to suggest so. Beyond the justified skepticism about what machine translation (MT) can achieve, this attitude overlooks the choices we make in translations.
As Good As Human Translation?
The recently developed Google Neural Machine Translation has been hailed as producing superior results to (currently available) phrase-based machine translation and even approaching human translation quality. The translation company Systran has announced its own version of neural machine translation. However, as Kirti Vashee points out, Google’s method of scoring translations ends up overstating the actual improvements in the output. Experts interviewed by Slator also questioned the methodology used to assess the progress.
Many of the claims centered on whether neural machine translation was “nearly indistinguishable from human translation.” In fact, the basis for scoring Google’s translation was comparing machine-translated excerpts from a variety of texts with their human-translated counterparts. However, there is little discussion of what makes a good human translation or a good translation overall.
Automation and Decision-Making
At this stage of technological development, we would likely not brand a company with a computer-designed logo or publish computer-written fiction without having a human direct or edit the output (advances in computer-generated journalism notwithstanding). It is thought that when technology automates some of the menial jobs, humans will still be needed for the more creative tasks.
In my view, what makes these jobs hard to automate — at the current state of artificial intelligence — is the decision-making process involved. While a car being assembled has exact specifications of the final product, a logo or a marketing text is ostensibly a more open-ended task, where the final product isn’t obvious at the beginning of the process.
A “Perfect” Translation?
Yet translation is somehow treated differently. It is tempting to discount the infinite-possibility decision-making process involved — after all, the source text has already been written and it would seems that all decisions have been made. All that’s left to do is recast them in the other language. Indeed, Google limits its criteria of a “perfect” translation to “the meaning of the translation [being] completely consistent with the source, and the grammar [being] correct.”
This approach implies there is one correct translation, and the task of both human and machine translators is to arrive at it. However, there is arguably more than one acceptable output, depending on the purpose and target audience. A functional approach to translation postulates that
[It] is not the source text as such, or its effects on the source-text recipient, or the function assigned to it by the author, that determines the translation process, … but the prospective function or skopos of the target text as determined by the initiator’s, i.e. client’s, needs.
For example, the same public health brochure may justifiably have to be translated differently for a Russian-speaking population in the US as compared to a Russian-based target audience. The first group is more likely to be familiar with US-specific healthcare concepts, such as “co-pay” or “nurse practitioner,” whereas the second group will need an explanation or adaptation. The same is true for cases when an accurate translation evokes negative connotations.
Making the Choice
We see that most utterances in the source language allow for several adequate translations. Does that mean that machine translation that produces any of these potentially acceptable translations at random has fulfilled its purpose?
While I am not qualified to comment on the programming behind neural machine translation, according to published research, the probability of a certain translation occurring in the set the NMT system was “trained on” is taken into account when making the final choice. In other words, in the best case scenario, NMT will pick a reasonable, grammatically correct, most likely translation based on its training dataset.
For many text types, this may be quite satisfactory. But is the most likely or the most common choice always the most appropriate one? Even after machine translation has surmounted the challenges of grammar and syntax, which is no small feat, I believe many clients and authors who care about their message will still rely on the judgment of the human translator — if only to make sure the machine made the right choice.