Machine Translation Fails Because of the Context Conundrum

Is Star Trek’s “Universal Translator” just around the corner? Recent advancements in machine translation from companies like Google, Microsoft, and Twitter seem to indicate as much, according to Wired—if Tweets, video conversations, and smartphone images can be translated in real time, is there any room left for the human touch? Absolutely! Context is complicated.

Lessons from History

One of the most famous mistranslations in history comes from Saint Jerome, who studied Hebrew so he could translate the Old Testament into Latin directly from the source. But when he got to the scene where Moses comes down from the mountain, the saint misinterpreted the word “karan,” which means “radiance,” as “keren,” which means “horned.” Because written Hebrew doesn’t use vowels, the mistake isn’t completely without merit, but given the context, this should have seemed suspicious to Saint Jerome. As a result of his error, images of Moses had horns for hundreds of years. In a more recent example from Russia Beyond the Headlines, President Vladimir Putin was mistranslated when he used the Russian word “gosudarstvennost” in reference to the conflict in Ukraine. The word has two meanings and North American sources interpreted the word as “statehood,” while the Kremlin insists the President meant “governance.” Again, context is crucial.

Machine Learning

If humans do so poorly with translations, shouldn’t high-performance machines have the upper hand? Consider Google’s Word Lens app. Sounds like a great idea, but in practice the app can butcher simple phrases: “parking service vehicles” in Russian becomes “stand official motor transport” in English, and an instruction in Spanish turns into “take to end the following procedure.” And machines really fall down when it comes to humor or word-play, prompting the Technical University of Berlin to suggest developing a humor switch that “analyzes the full syntactic chart of the source sentence” in machine translation. It’s easy to see how this could be applied to context in a broader sense—if the environment of source data is better understood, the result should be an improved translation.

Machines may be fast and accurate, but can’t handle context. Native speakers have the ability to recognize context problems and find solutions. Maximizing the speed and quality of translations—especially for technical or specialized content—is often best-served by a combination of professional translation and review and a centralized translation management system. Consider cloud-based translation, where the content is collected and then sent to the ideal human endpoint, such as transcreation experts who specialize in marketing taglines, where direct translation won’t have the desired effect. Human translation, when supported by cloud technology, keeps context errors to a minimum.

Bottom line? Machine translation comes close, but humans help close the language gap.

Image source: Wikimedia Commons

About Doug Bonderud

Doug Bonderud is a freelance technology writer with a passion for telling great stories about unique brands. For the past five years, he's covered everything from cloud computing to home automation and IT security. He speaks some French, is fluent in Ancient Greek and a master of Canadian English — and yes, colour needs a 'u'.