Google Crosses a New Frontier in AI with Machine Translation

Special guest post from Rachel Wheeler at Morningside Translations

Artificial Intelligence is nothing new, but the 21^st century has seen some of the most significant advances in the technology. In the past, AI systems were very simple. There were only able to perform tasks that they were programmed to do, and only in the ways in which they had been programmed to do them.

With recent developments in artificial neural networks and deep learning, AI systems are now capable of performing tasks that are more complex. They can even learn how to perform tasks without having to be specifically programmed to do them. Recently, Google has made some significant breakthroughs through applying deep learning to machine translation.

Why Deep Learning is Important

Deep learning is a type of machine learning that allows a computer to process information in a way that is much more similar to that of a human. To achieve this, programmers use artificial neural networks designed to work similarly to our brains.

The human brain has billions of neurons that form connections to create neural pathways. As a person learns how to perform a task, a neural pathway might become stronger or weaker based on experience and the quality of the result provided by the pathway. In an artificial neural network, the system will have artificial neurons that are arranged in layers. As the system is exposed to information, it will form connections within the artificial neural network. These connections can become stronger or weaker based on the quality of the result as the system learns from experience.

With artificial neural networks and deep learning, a machine can ‘learn through doing’. This is different from older systems, which could only do what they were told. To make any change or improvement, a human programmer had to develop new commands and update the code. This newer technology means that it does not require this level of human intervention for the machine improve at its job.

Narrow AI Translation Systems

Machine translation systems have been around for many decades. The older systems operate on what is called narrow AI. With narrow AI, a system can perform a specific task based on pre-programmed rules. In the case of machine translations, the rules would tell the system how to break a sentence up into fragments and translate those fragments to a different language. It would then apply a set of post-processing rules to construct a meaningful sentence from the translated fragments.

Unfortunately, this often results in mismatched, inorganic translations. Narrow AI is not capable of translating changes in cultural nuances or idioms. While narrow AI has certainly made doing large amounts of translations easier, it is still necessary to employ content translation services which can catch any missed detail.

Google’s Major Breakthrough with Deep Learning

Google Translate used to operate on a variety of narrow AI systems. Then, in 2016, the tech giant announced that they were switching to a single multilingual system that uses an artificial neural network. This is what we now know as Google Neural Machine Translation (GNMT). The system was trained using millions of sample translations, and it continuously learns and improves from experience.

The researchers at Google expected the GNMT to get better at its job; that was part of the plan. However, they were surprised to find out that the system could use its existing knowledge to learn how to perform translations for which it has yet to be trained.

As an experiment, the researchers wanted to see if it could use its understanding of Korean to English and Japanese to English translations to teach itself how to perform Japanese to Korean translations.

The GNMT was able to successfully complete this task by creating an artificial language that could be used between the three source languages. The system would translate the source language into the artificial language that it created, and then translate the artificial language into the target language.

This is what researchers have termed zero-shot translation – the ability to translate between a language pair for which it has not been trained. Instead, the system transferred what it had learned about the two original language pairs and used that knowledge to create a system for performing translations in the new pair. Although the system is not yet perfect, this new technology has the potential to completely revolutionize the translation industry.