How Machine Translation is Transforming With AI DevelopmentsReading Time: 5 minutes
When we think of artificial intelligence, we usually think of the types of things that we see in movies or read about in science fiction stories. But advanced AI is no longer limited to the imagination of science fiction writers. Artificial intelligence has become a part of our everyday lives, and its use is only going to expand as time goes by. Recently, one of the most interesting developments in the world of AI has occurred in the field of machine translation.
The consequences of this technological advancement can benefit all forms of translation, from legal document translation to multinational clinical research. This is what you should know about the latest AI developments contributing to machine translation improvements.
A look at AI developments through the years
As early as the 1950s, we started to see developments in narrow AI. Narrow AI is a form of AI that focuses on a specific (or narrow) function. It runs on a set of rules that help it to perform its intended task, but it can’t learn or expand beyond what it is programmed to do. For example, this would include something like the AI in a videogame software.
Flash ahead to the 1990s, and you start to see developments in machine learning. Instead of following a strict set of rules, this type of program can learn from its experience. A machine learning program uses algorithms to process large data sets, and separates relevant information to make predictions and get better at its task. Examples of this include personal assistant programs, like Apple’s Siri and Cortana from Microsoft.
In the 21st century, the most advanced forms of machine learning use what is known as “deep learning”. This is a form of machine learning that resembles how the human brain works. With deep learning, the program has artificial neural networks and it can learn from past experience to perform new tasks without being programmed to do so.
In the future, researchers are looking to develop general AI. This would be a form of AI that is more akin to human intelligence. In essence, this would be a form of artificial intelligence that could perform general operations in the same way that a human could. This would be robots like C3PO from Star Wars, or if you are thinking of something a little darker, HAL 9000 from 2001: A Space Odyssey.
Hi, I’m Alex Knight. What’s your name?
What makes deep learning a major technological leap?
With traditional forms of machine learning, the AI can only learn about the tasks and functions for which it was previously programmed. Deep learning allows a machine to learn independently. In other words, it can learn in a way that is similar to that of a human.
The human brain has billions of neurons, with each neuron being connected to others to form a network. When one neuron shares an electrical signal with another, it forms a neural pathway. If this pathway is helpful to learning or performing a new task, it gets stronger with time and experience.
Borrowing this idea from the human brain, researchers have built neural networks for machines. This allows the machine to learn from their experience of the world around them, instead of having to learn based on a set of preprogrammed rules.
By exposing these machines to significant volumes of data, they can use trial and error to learn from their experience. This is what researchers call “training”. When the AI is properly trained, it should be able to provide reliable answers every time. However, to perform this type of training, you need to expose the machine to vast amounts of real-world data.
The idea of building an artificial neural network is not new. This concept for deep learning has been around for decades. The reason it is just now coming to practical use is that technology has finally caught up to the idea.
In the past, a task like this would have required computing resources and volumes of data that would have been impossible or impractical to obtain. In the 21st century, computing power and data are much more accessible, making it possible for researchers to experiment with deep learning.
Google’s big breakthrough in machine translation
Most machine translations use a form of narrow AI. These programs work on a preprogrammed set of rules that perform the task step-by-step. The steps are as follows:
- First, the program breaks the sentence up into fragments
- It translates the fragments using a dictionary of statistically-derived terms
- Then it uses a set of rules to rearrange the translated terms into a sentence in the target language.
If you have ever used a machine translation, you know what these systems can do well, but you have probably also noticed that they have some serious flaws. These flaws are the result of limitations that are placed upon the program by its rules.
It can do word-for-word translations and it can rearrange words based on structuring rules, but languages are more complex than this. The machine cannot pick up on things like tone of voice, and it cannot account for situations where the same word has different meanings that depend on context.
Google Translate was one of these services. It could work with dozens of different languages, and it had an array of different programs that worked to provide the translations. Then Google announced that it was switching to a new system that would be based on an artificial neural network.
This new system is the Google Neural Machine Translation. The AI program applies deep learning, and it can work with a multitude of languages while continuously learning from the language examples that it experiences.
Considering it is a machine that applies deep learning, researchers already expected the machine to get better at its intended task. Surprisingly, the machine did more than just get better at translations. Researchers found that it could also teach itself how to do new types of translations that it had never seen before. As an example, if the machine could do translations between English and Korean and English and Japanese, could it also do translations from Korean to Japanese without being taught? The answer was yes.
This is what researchers are calling a “zero-shot translation”. The GMNT basically used its experience with the past translations of both English/Korean and English/Japanese to create a system for conducting translations between Korean and Japanese. It required no training from a human or new programming. Instead, it used the information it already had to teach itself how to perform a new task.
This represents a significant development for the world of machine translations, and it is also a huge leap for AI and deep learning. The GMNT came up with an original idea that allowed it to perform a new task.
As these technologies improve, we are going to see AI that does more than provide improved machine translations. We could be moving toward a future where we see machines that have the intelligence to think and act like humans. Perhaps this will give rise to an opportunity for CSB Tech Emporium to enter the AI and deep learning field.