He is a lawyer. She is a housewife. He is smart. She is beautiful. Simple enough to translate. So, what could possibly go wrong when machine translation learns from us?

Neural machine translation (or NMT), like the new Google Neural Machine Translation system, copies what it learns from its training data and produces what it believes we humans want to read. Even if the results are patently sexist (or racist, or biased in another way—I’ll have to save that for another blog post).

Google is currently trying to reduce the gender bias in its NMT, but will it be enough?

Neural machine translation is differentiated from its predecessor, phrase-based statistical machine translation, by its use of neural networks and deep learning to predict one word at a time. NMT is recognized for generating fewer errors than statistical machine translation. However, it still can’t replace the quality of a qualified human translator, especially when that translation is combined with a round of editing performed by a second linguist and a set of quality assurance checks from a third linguist, like we do at Responsive Translation.

NMT systems require training data—lots and lots of native texts and their translations—to learn from sentence and word patterns (not all of it prize-winning material either). Some patterns are harmless, while others not so much. The original writers of these texts may not be aware of their biases. Neither are the engineers when they feed the system those huge data sets, but neural machine translation automatically replicates those same biases in aggregate.

When translating, many a qualified human linguist will naturally stay clear of loaded language. However, machine translation algorithms can’t identify (or neutralize) biases on their own. That’s why Google is now trying to devote extra problem-solving power to teaching its NMT not to be so sexist all the time. Or at least for major language pairs.

Google is starting to train its NMT to detect gender-specific and gender-neutral requests. An example of a gender-specific request might be: “How many languages does he speak?” And an example of a gender-neutral request could be the Turkish sentence: “O bir mühendis.” Previously, this was translated as “He is an engineer.” but now the gender-ambiguous phrase has two options—“He is an engineer. (male)” and “She is an engineer. (female)”. (Before, “O bir hemşire.” in Turkish was translated as “She is a nurse.” but that phrase has been replaced by two options as well.)

Nevertheless, there is still a lot of bias out there in the wild and it is affecting what Google and others can get out of machine translation. Is Google committed to tracking down and neutralizing all the biases its neural machine translation learns so quickly and so regularly? I guess we shall have to wait and see if the results can speak for themselves.

When You Need Translation That Won’t Offend

Certified for ISO 9001, Responsive Translation provides a range of foreign-language services in more than 200 languages and dialects, including translation and human-edited machine translation.

Responsive Translation offers the highest-quality translation services currently available with our three-step human translation process: translation, editing and quality assurance. Each step is performed by an experienced, highly-qualified linguist and is backed by our satisfaction guarantee.

We also offer a cost-efficient alternative (not recommended for legal or medical texts, or many customer-facing applications), combining the speed and low cost of machine translation with the editing expertise of our human linguists.

For more information on how we can help you, please contact us at 646-847-3309 or [email protected].

Image credit: Tim Mossholder