It was only last week that we discussed the flaws of machine translation. In that case, the focus of the article was on the WeChat messaging app’s racial slurs. Now, a new story has once more highlighted the dangers posed by the machines that we rely on to translate for us.
Arrest in Israel
The case in question concerns a Palestinian gentleman working in Israel. The man works in the construction industry near Jerusalem. He recently posted a photograph of himself at work, leaning against a bulldozer. He added the caption “يصبحهم”, or “yusbihuhum,” which means “good morning.”
Unfortunately, this seemingly innocent act triggered a great deal of trouble for this individual. Facebook’s artificial intelligence (AI) powered automatic translation system, which handles some 4.5 billion translations per day, made a mistake. Instead of translating the man’s innocent greeting correctly, the system translated it as “attack them” in Hebrew and “hurt them” in English. Local police put the translated comment together with the bulldozer in the picture and promptly arrested the poor construction worker.
Human translation lacking
Sadly, no Arabic translation was undertaken before the arrest. Nor were any Arabic-speaking officers consulted prior to the gentleman being hauled in for questioning. The suspicion that the gentleman was planning some form of bulldozer attack caused police to react first and ask questions later. Of course, once they did begin asking questions, the truth quickly became apparent.
The incident serves to highlight once more the importance of human translation rather than reliance on machines. In this case, the issue was resolved within a few hours when the erroneous translation was discovered. However, one can’t help but wonder how many other instances of fault Facebook translation are taking place each and every day, as users rely on artificial intelligence to communicate in other languages.
Facebook has, quite rightly, apologised for making such a mistake. Necip Fazil Ayan, an engineering manager in Facebook’s language technologies group, commented,
“Unfortunately, our translation systems made an error last week that misinterpreted what this individual posted.
“Even though our translations are getting better each day, mistakes like these might happen from time to time and we’ve taken steps to address this particular issue. We apologise to him and his family for the mistake and the disruption this caused.”
The translation error occurred in part due to the difficulties that machines face when it comes to translating Arabic. Facebook’s system is far from alone in finding Arabic a hard language to translate. As well as the internationally used Modern Standard Arabic, the language has a large number of different dialects. This provides machines with a level of complexity that they don’t often face when working with other languages.
Even so, the error shows how far machine translation still has to go before we can rely on it. If a simple phrase such as “good morning” can serve to flummox Facebook’s translation system so utterly, despite the enormous funds available to Facebook to invest in that system, one shudders to think of how poorly more complex chunks of text are being translated.
Over the past few years, we’ve heard about numerous breakthroughs in the field of machine translation, from Google’s neural networks translating without transcribing courtesy of deep learning techniques, to Facebook’s much publicised shift to a system of entirely neural machine translation. So why is it that machine translation is still so flawed compared to human translation?
Google recently announced that its Deep Mind project had made a leap forward in AI research. The AlphaGo Zero project saw Deep Mind achieve superhuman abilities at the game Go after being given a board and a set of instructions with which to teach itself. Whereas the previous project – AlphaGo – required thousands of hours of learning from human players before it was able to beat a human champion, AlphaGo Zero taught itself to do so in just a few days, defeating the original AlphaGo by 100 games to 0. The machine even taught itself Go techniques and strategies that humans have never thought of, despite the game being over a thousand years old.
Perhaps this new approach to machine learning will finally be the masterstroke that leads to machines being able to translate as proficiently as humans – but, then again, we’ve heard such proclamations many times before and humans still have the edge over machines when it comes to professional translation!
Will the latest advances in AI lead to machine translation being perfected? Or will we continue to see machine translation mistakes leading to confusion and, in extreme cases, arrests? Share your thoughts by leaving a comment below.