Artificial intelligence (AI) has taken front stage in many circles. AI is that mysterious knowledge and skill that enables robotic automation. Like all technologies, the methods to impart robots with AI continuously evolve. Today’s state-of-the-art method is called machine learning. As the name implies, the robot learns its knowledge instead of engineers bestowing AI.
We tend to think of robots as modern creations. On closer look, we find that Philo Mechanicus amazed his house guests in ancient Alexandria with “history’s first robot” that, on demand, automatically poured a precise mixture of wine and water.
What is a robot?
You may never have heard of translation robots, but they’ve been around for decades and now they’re changing our translation industry.
Before delving into our translation industry’s automatons, let’s look at robots in general. A robot is a technology that executes complex tasks that humans need done. I think it’s safe to say that today’s tasks are more complex than those in Philo’s 3rd century BC.
Furthermore, today’s robots come in all shapes and sizes – from monstrous arms that boom across a factory’s assembly line, to personal assistants that sit on your kitchen counter. Some robots even have no size at all, like the automated software “bots” that catalog the Internet and spam filters that protect our email.
Robots and artificial intelligence
Artificial intelligence is what drives a robot to preform its tasks. A series of levers and inclined planes were the AI that Philo Mechanicus “programmed” into his robot. Two millennia later, engineers programmed the operating rules into early computers that controlled factory automation robots.
Many of today’s robotic tasks have grown so complex that engineers had to develop new programming techniques, like machine learning, to create AI. With today’s state-of-the-art machine learning, a robot studies complex tasks from the past to learn its AI. Then, the resulting AI drives the robot to perform similar new tasks.
Machine learning differs from other AI because it divides the AI creation process into two distinct tasks. For one task, engineers create algorithms that study, learn and do tasks based on the learned knowledge. The second task has subject experts collect, organize and select the tasks from the past that the algorithms will study.
For whatever reason, the way they designed the world, data is only available about the past.Clayton Christensen
Human Learning Metaphor
In human learning terms, engineers are the school’s administration. They create the learning environment much like a classroom with electricity, blackboard, lights, desks and chairs. Subject experts are teachers who create and present the textbook to the robots for study. Robots are students that study the textbook, learn and perform tasks within the scope of their learning.
If the school adjusts the classroom lights and rearranges the desks, there is an impact on the learning experience. Once the classroom is complete, adjusting the room has fewer benefits, if any.
The textbooks are a different story. Changing the textbook causes a significant and immediate change in learning. A teacher uses a math textbook to teach mathematics and students learn mathematics. If the teacher changes to a humanities textbook, students learn humanities not mathematics.
In human learning, teachers don’t use the entire encyclopedia as their textbook. They focus their students to study textbooks by subject matter. Therefore, if a student takes a mathematics test after studying math, we can expect a high score. If a student takes a humanities test after he studied mathematics, we expect a low score.
Rise of the Machines
Machine translation (MT) software executes the complex tasks of converting human language. It sounds very much like a robot and I propose we could rightfully call MT a translation robot. After decades languishing in quality disappointments, the quality from translation robots significantly improved after the introduction of a machine learning technology called statistical machine translation (SMT). In effect, the engineers didn’t adjust the classroom, they changed from a kindergarten classroom to a new elementary school classroom.
In the early 2000’s, speculation grew that SMT had finally elevated translation robots to the potential of replacing human translators. I wonder if Philo’s human servants ever feared the wine-pouring robot would replace them in their daily chores.
SMT quickly found its way into cloud-based services like Google Translate. To satisfy the unlimited demands from millions of people, these services hired data scientists who created encyclopedias and teach their robots.
Over time, however, they learned that the machine learning process emulated human learning all too well. Like human students, when the robots study an encyclopedia it becomes a jack of all trades but master of none.
SMT’s quality peaked early in the game but it took 10 years before the experts acknowledged it had matured. In a recent blog interview, the Microsoft’s MT Group Program Manager, Chris Wendt, said:
(SMT) systems are more mature. They have gone through more scrutiny which has addressed many shortcomings. We see that (SMT) systems are flattening out. Adding more data doesn’t help anymore. Tuning it here and there doesn’t really make too much of a difference.Chris Wendt, Microsoft MT Group Program Manager
With Chris’ confirmation, we now know that these SMT translation robots have learned all they can learn from the encyclopedia. Their quality has peaked — but just what is the quality level of this peak?
In July 2016, with more than 10 years of improvements to SMT, Memsource published their analysis of 38 million words in this blog. When professionals used with Google and Microsoft, less than 10% of the SMT’s predictions were correct across 23 of 25 language pairs.
Salvation With a Mature Technology
As Chris points out, SMT is a mature technology. The quality from these translation robots has peaked in the use case of serving millions of people, but this does not mean SMT is a the end of its life cycle. There are other use cases where our SMT-based translation robots perform significantly better.
|Who does the robot serve?||Millions of laymen||One translator|
|Who is the teacher?||Data scientist||Same translator|
|Textbook or encyclopedia?||Big data||Personal TMs|
|Personalized performance?||0% to 10% correct||30% to 50% correct|
The translator is the ultimate subject expert who collects, catalogs and selects the translation memories that become textbook’s contents. When shifting to the one translator use case, there’s a natural change and acceptance. The translator teaches the robot and in turn the robot takes on the translator’s personality to do the translator’s bidding.
Translators as teachers of translation robots simply need the tools to convert translation memories and interact with their robot on their desktop. That’s when the robot transforms from the Terminator to a personalized translation robot, programmed by its creator to serve its creator.