Artificial intelligence may seem like science fiction, but it’s technically existed for decades. In 1951, students at the University of Manchester created a program for the Ferranti Mark I computer that allowed it to defeat amateurs in checkers and chess. That may not seem too impressive anymore, but it spurred a period of major innovation that continues to this day. Now, that technology can be applied to how we conduct website translations.
Google recently upgraded the AI of Google Translate, making it potentially much more effective than past web translation services. To understand how this new breakthrough works, it’s necessary to get some background on the basic classifications of AI.
While a computer that can play chess may have been impressive in 1951, now there are plenty of similar (and more sophisticated) programs you can download straight to your phone. These early developments were examples of narrow AI, in which a programmer “teaches” a computer to perform basic, rule-based functions and tasks.
This type of AI can learn how to play checkers, but it can never learn to research the history of checkers. Basically, it can’t develop its own natural curiosity, and wouldn’t know how to apply said knowledge if it could.
Machine learning AI became more prominent in the 1990s. Rather than playing a game with constant rules, machine learning AI represents a shift towards programs that can actually “learn” on their own.
Essentially, the machines leverage specialized algorithms and refer to substantial amounts of data to acquire knowledge. For example, when you ask Siri a question, it sorts through data and breaks it down into subsets, arriving at what is most likely the correct answer. Siri doesn’t technically learn how to research on its own, nor does it retain knowledge or act independently in the same way a human does. What it can do, however, is adapt to learning situations that exist outside of the basic rules it was programmed to follow. Compared to narrow learning AI, which can only do the one particular thing it’s assigned, machine learning AI isn’t nearly as restricted.
Deep learning AI has been on the rise in the past decade. Structurally, the algorithms are based on the human brain. Although tech visionaries understood the value of this approach, until the right hardware and technologies were available, it was impossible to design such a complex system.
Unlike machine learning AI, which can mimic some form of actual thought, deduction, or reasoning, deep learning is the first type of AI which can use knowledge of past behavior and apply it to new problems outside of its programming. For example, a deep learning program from Google was able, after being exposed to 10 million images, to recognize specific objects (like cats) twice as accurately as previous image recognition programs.
In all likelihood, general AI — the kind featured in sci-fi movies with independent, thinking robots — will be a part of the everyday reality in the near future. This phenomenon will likely have a major impact on translating language.[bctt tweet=”AI will likely have a major impact on translating language.”]
Most online translation services work by dividing a sentence or phrase into smaller parts, referring to dictionaries to identify the equivalent words, and relying on post-processing to adjust the sentence structure according to the language’s specific grammatical rules. Anyone who has used one of these tools before knows the results are far from perfect.
That’s why Google Translate is shifting to a new method. Previously, the service worked by using hundreds of narrow AI programs to translate text. Now that Google has begun implementing deep learning AI in its translation service, the program will actually be able to learn from past experience. This allows it, in theory, to avoid the kinds of errors that are commonplace in online, machine translations.
The upgrade also allows it to bridge the gap between language problems it may not have encountered before. Perhaps it translates a Japanese text into English, then a Korean text to English. Early results indicate it will then translate Japanese to Korean with decent-to-remarkable accuracy.
Researchers call this breakthrough the “zero-shot translation,” and it represents the breakthrough fact that Google neural machine translation — the current moniker for the new translation program — basically “learned” how to achieve it independently. In fact, Google experts only have theories regarding how it accomplished this feat; they’re not entirely sure.
While this absolutely marks a major leap forward in online translation services, it doesn’t mean human translators will be replaced anytime soon. Effective translation requires not only an understanding of the nuances of language, but also an understanding of a given culture and how cultural attitudes impact the effectiveness of language. Until a machine can do that, it’s not quite human just yet.