The dream of making a machine understand human language is younger than the human dream of being able to fly. As of today, talking and reading machines, while no longer a dream, are yet to become commonplace. But even in its infancy the history of natural language processing shares many similarities with the history of aviation.
Both aviation and natural language processing looked to borrow ideas from Mother Nature. Many of the early flying machines were modeled after birds, like the proto-planes with wings they never got to flap in the air. In the less tangible world of natural language processing, arguments went on about how the brain processes language. Is there a universal grammar at work? Or is it a complex learning process with neural networks? The common perception was that in order to replicate the functionality, we have to mimic the original.
There is always more than one way to solve an engineering problem. After a while, with more eyeballs on the problem, several trends emerge, each with its own advantages and disadvantages. Hot-air balloons and zeppelins vs. heavier-than-air contraptions with propellers and wings. This is not unlike the schism in the natural language processing, frequencies and co-occurrences vs. grammar and rules.
And, of course, there are the skeptics. In 1895, Lord Kelvin himself said, “heavier-than-air flying machines are impossible”, only to be proven wrong in 8 years. The number of skeptics saying that the machines will never figure out the intricacies of the human language has reduced significantly, but they still exist.
The skeptics largely go away after convincing proofs of concept. In aviation, it was December 17, 1903, when the Wright brothers flew their powered, controlled aircraft near Kitty Hawk, North Carolina. Apparently, the definitive first is not as clear-cut, as there were several claims with different degrees of success. But this is when the aviation became a part of the public psyche.
In natural language processing, the last years were marked by several “Kitty Hawk moments”. In 2011, IBM demonstrated software that defeated top human players in Jeopardy! by accurately answering trivia questions. In the same year, Apple acquired a startup that created a revolutionary app named Siri that allows the user to control their smartphone by talking to it.
What happens after a daring stunt is less spectacular but not less interesting.
After years of patent lawsuits and bickering, the pioneers of aviation had to solve two problems: 1. Build machines that work longer than a few seconds, and 2. Find the first mainstream use.
The flying machines needed better engines; and so in our natural language world, the next step after an impressive proof of concept is to build a reusable component, an engine which can become a part of other software, and find a mainstream use.
When I co-founded LinguaSys, a natural language processing startup later acquired by Aspect Software, that was my motivation to focus on a framework while many others were building apps. What is now called Aspect NLU was planned as a comprehensive answer for all (or most) natural language tasks.
With the proliferation of bots and bot-friendly messaging conduits, I believe we found our first commercial use. Just like the creation of the world’s first airline, DELAG, in Germany in 1909, the tipping point for the natural language world was several big enterprises, including Aspect Software, betting big on the machines that understand human language. Aspect Software is the first of the large customer service software providers to go beyond limited experiments and release natural language solutions in production. Being a newcomer to the customer service industry, it took me some time to appreciate the audacity of this move.
But I know it will fly.