Dramatic innovations in neural machine translation are making high-level translation tools accessible to the masses. But what impact do they have on professional translation services? Sure, translation agencies can use them to get a leg up on translating a report from Spanish to English. But can machine translation tools really compete with the quality of a human professional translator?
Advances in AI technology are transforming professional translation services. We’ll consider the revolution that AI has brought to Machine Translation, focusing on scholarly and corporate developments, and their impact on consumer tools and resources. We’ll consider the latest trends and innovations in neural machine translation, deep learning, conversation learning and more. While AI continues to trail humans in some translation genres, it has caught up in others. We’ll consider an emerging synergy where humans cooperate with machines to augment productivity.
The AI Transformation that Changed Machine Translation
Deep learning and AI-driven applications have been applied to natural language translation for decades. But the revolution in machine translation happened just five years ago. In 2014, the first scientific paper appeared describing the application of neural networks to linguistic translation. In 2015, a neural machine translation application appeared in the Open MT public competition for machine translation. WMT, in the same year featured a neural contender. In subsequent years, 90% of all competition winners were neural network translators of one sort or another.
The variations on NMT began to proliferate: Large-vocabulary NMT, Subword-NMT, Multi-Source NMT, Zero-Resource NMT, Zero-Shot NMT, Fully Character-NMT – you could start an NMT ice cream store. In 2015 an NMT system first appeared in Open MT a public machine translation competition. WMT'15 also for the first time had a NMT contender; by the following year, fully 90% of NMT systems were among its winners. NMT’s pre-eminence was also helped along by the introduction of a NMT section to the annual WMT and Google’s own NMT workshop, also now an annual event. Indeed, NMT has effectively annexed machine translation as smartphones supplanted phones. All significant machine translation now is based on neural nets.
What innovation did neural networks bring? To make a complex story relatively simple, neural machine translation replaced the linear, phrase-based translation. NMT applied a more contextual approach which referenced a vast knowledge base of linguistic examples and applying the heuristics of neural networks to grasp the meaning of the whole passage or document and then working down to the details of individual sentences and worlds. In other words, it worked from the whole to the part rather than snaking along from part to part.
Neural Machine Translation, Meet the Internet
It was a match made in heaven, or at least in the cloud. All of the big corporate players who today dominate the internet and media, our economies and societies – Facebook, Amazon, Microsoft, Google and Apple – were intensely interest in translation. The growth of their businesses depended on it.
Each of the FAMGA mega-corporations has since invested heavily in translation technology. Among the noteworthy developments has been Google Translate, Microsoft Translator, and Facebook’s current NMT project which trolls its billions of user comments to understand how to translation conversational language – or at least what passes for translation online (how do you translate “lol”?). Compare that to the running start Google gained when it translated hundreds of millions of European Union documents to jump start its translation knowledge base.
Today the flagship products of the FAMGA companies are offered for free in both desktop and mobile versions. While Google Translate and Microsoft Translator differ in some respects, they have largely overlapping functionality and feature sets. In addition to straight translation from texts in one language to texts in another, both offer camera translation, which applies optical character recognition to interpret signs and menu in a foreign language and overlays them with characters in a more familiar language of your choice using a form of Augmented Reality. And both offer two-way voice translation, which enables two people who speak differently language to converse fluently using their smartphones as interpreters.
How Does Neural Machine Translation Compare with Human Translators?
Google has run tests to test its translation algorithms. it found is that, for many straightforward translation tasks, algorithms are producing translations that are getting high scores (4.6 out of 5 in one trial) assigned by human linguists. Not perfect but good. Online translation services are flesh-and-blood translators a run for their money.
Take a Step Back: Who Needs Professional Translation Services?
It’s important to keep in mind what drives demand for professional translation services. Because professional language translation services are more expensive than freelance translators, and of course more expensive than the largely free machine translation services available on demand.
Professionals who publish or communicate in foreign language are concerned with their reputations for quality and for the impact of their messaging. Therefore, their quality bar is set very high. 80% is not a passing grade for professional translations. To not tarnish the reputation of a corporate or government client, to not expose a law firm or medical practitioners to a malpractice lawsuit, perfection or at least excellence is expected. And that is not something that machines alone can consistently deliver. Google and Microsoft admit as much.
But the knowledge that machine translations are flawed does not deter massive usage. Marketing and advertising firms have become dependent on using machine translation tools as the only scalable and cost-effective way to keep up with foreign language conversation and research. An article about the marketing team as ASICS, responding to the question of “MT or not MT?” make a stark admission of the group’s dependency on machine translation to keep up with the multilingual workload. Machine-translated tools give a good “rough” idea about what foreign posts or articles may be about.
Gone with the Translation: Of Beatings and Big Winds
That “rough” idea, however, can have major human consequences. An internal manual produced by U.S. Citizenship and Immigration Services, guides officials who sift through non-English social media posts of refugees that “the most efficient approach to translate foreign language contents is to utilize one of the many free online language translation services provided by Google, Yahoo, Bing, and other search engines.” The article quotes Douglas Hofstadter, famed author of Gödel, Escher, Bach: An Eternal Golden Braid and a professor of cognitive science and comparative literature at Indiana University as calling the edict “naïve”, “deeply disheartening” “stupid and shortsighted.”
An article in Propublica cites a tweet from Pakistan which Mustafa Menai, who teaches Urdu at the University of Pennsylvania, translated as “I have been spanked a lot and have also gathered a lot of love (from my parents).” Google provided this translation: “The beating is too big and the love is too windy.” In 2017, Facebook apologized after its machine-translation rendered a post by a Palestinian man that meant “good morning” in Arabic as “hurt them” in English and “attack them” in Hebrew.
An Iranian caption satirically depicting elites raising their hands when asked if they had a child in the United States was human-translated into Farsi as: “Whose child lives in America?” Google rendered that as “When will you taste America?” Microsoft’s rendition was: “Who is the American?” Sheida Dayani, who teaches Persian at Harvard, forbids her students to use machine translation tools for their assignments. “The thing about Persian and the Iranian culture is that people love to make jokes about anything,” she said. Humor and irony are lost in translation.
Ceasefire and Modus Vivendi in the Human vs. Machine Translation War
The ongoing competition between humans and machines for professional translation is not a zero-sum game or a winner-takes-all war. It is safe to say that humans will for the foreseeable future enjoy advantages over machines in translating originals that are complex, creative or funny. Human linguistic, cognizant of cultural nuances, will still understand the complexities of culture and language better than algorithms.
On the other hand, computers can increasingly perform the “grunt work” of translating routine repetitive and highly structured language, which represents the lion’s share of professional documents out there. This does not mean that there is no role for human linguists but their responsibilities may evolve into something more akin to an editor or a proofreader who serve to check the work of the software rather than replicate it.
As for students and educators in the fields of Artificial Intelligence, this new balance opens new horizons for study and development. AI Education, after all, is still in diapers relative to many other fields of endeavor. Future computer scientists may well seek to improve the translatability of humor and nuance and poetry. But those of us who live and breathe may be well-advised not to hold our breaths to wait for the zillionth software monkey or mechanical Turk to churn out a Shakespearean masterpiece.
