by Jeff Stibel
Harvester ants generally don’t fight, but “there seem to be seasonal bursts of fighting, often just after the summer rains,” says Dr. Gordon. “Maybe the rain washes away chemical signals on the ground, such as colony-specific cuticular hydrocarbons, and the absence of those signals stimulates fighting.” So it seems that while ants avoid foreign-smelling ants, the absence of scent creates all sorts of mayhem.
Ant networks require robust and precise communication, and evolution has outfitted ants with the necessary organs both to receive communication—their 400 olfactory receptors—and to “speak” their own language—through an amalgamation of glands on various body parts including the rectum, sternum, and hind tibia. Thus equipped, ants within a colony speak fluently to each other. Outside of the colony, however, communication breaks down.
So it is with humans. We tend to understand best those who speak the same language. Of course, even that general statement isn’t true without certain caveats. Flawless communication is a rarity because language is subject to many complex factors, including age, education level, and geography. Many an American tourist in London has found that George Bernard Shaw was correct in his assessment that America and Britain are two nations divided by a common language. Closer to home, parents lament that their teenage sons and daughters speak a completely different—and often completely unintelligible—language.
We forge stronger connections with those with whom we communicate best. Our most important personal networks are made up of people who “get” us, and a large portion of that designation is determined by how well they speak our unique language.
To be sure, the internet has broken down countless barriers between people who speak different languages. It’s a truly universal platform, enabling us to understand people a world away in terms of both distance and experience. The underlying process is remarkable. To exchange thoughts with a colleague in Japan, an American professor must first input his thoughts in English into the computer through his fingers. These thoughts then go through the web (different language), then across the internet (different language), only to be translated and displayed in Japanese (different language). Currently the internet attempts to make sense of 8,512 computer languages, dozens of HTML-based web languages, and nearly 6,500 active spoken human languages. But we’re not there yet.
No network—ant, human, or technological—can achieve maximum success without efficient communication. The internet will not reach its full potential until we find a way to overcome the language barrier. Though the network is already solid, it will become infinitely stronger once we figure out how to coat ourselves—all of us—in the same cuticular hydrocarbons. In order for the internet to evolve, it has to learn to communicate, just as ants did millions of years ago and humans did thousands of years ago.
I
Most people have never met a linguist in real life and assume they must be something on the order of Noam Chomsky, a quixotic character from MIT in a tweed jacket who has the nasty habit of correcting elocution in public. You wouldn’t want to sit next to one at a dinner party. But at the risk of overstating the case, linguists today have become the Indiana Joneses of brain science—great explorers of the unknown.
The previously stodgy world of linguistics has become a hotbed of innovation. In fact, almost every large internet company now employs linguists, and there is an army of them across the web. “Until recently, linguistics was mainly an academic pursuit and those jobs were hard to get and not that well-paying,” said Ed Stabler, associate professor of computational linguistics at UCLA. “Now—in spite of university offers—almost all of my PhDs are in the dot-com industry. In the past five years, the change has been phenomenal.”
Language underlies many of the technologies that make the internet useful to all of us. Take search engines, for example. Google deploys intelligent software to “read” a website, and that’s how it matches the best webpages to a search keyword. At this very moment, Google and the other search engines are actively reading thousands of websites to determine their relevance to various user searches. Words, after all, are the fundamental unit of human intelligence, and language is the foundation of civilization. Sigmund Freud had it right when he said, “The first human who hurled an insult instead of a stone was the founder of civilization.”
Speaking in words is a uniquely human accomplishment, something that no other animal or computer has mastered. For linguists, the study of language, grammar, and the logic and reason implicit within is the key that unlocks the secrets of the mind. Language is enlightening not only because it allows us insight into how we perceive the world, but also because it indicates how we retrieve information. The ways in which we structure words into sentences and sentences into paragraphs reveal the logic behind our thoughts. And, of course, they reveal the continuous illogic that makes human beings human. Google and the other Silicon Valley companies hiring Professor Stabler’s students are attempting to leverage the power of language to turn the internet into a thinking, reasoning brain. But there is a catch: language acquisition happens in the brain before the breakpoint, and it is likely that the same will be true of the internet. This is the primary reason that so many companies are rushing to hire linguists—they see the opportunity but they also know it may slip away.
II
Linguists have learned that there is a critical developmental period when it is easiest for a child to acquire a first language. This concept was first raised theoretically in 1959 by McGill University neurologist Wilder Penfield, and subsequent research has provided confirmation. While scholars disagree on some of the specifics, it is generally accepted that the critical period for learning a primary language is from the age of 4 months until roughly age 5. After that time, it is very difficult to learn a first language. (It is believed that another critical period exists for the learning of second languages. This stage ends roughly at puberty, at which time language acquisition in general becomes far more difficult.)
It is largely believed that the reason for the critical period is that the brain is far more plastic during the growth phase than it is beyond the breakpoint. When neural connections are growing, the brain remains highly adaptive, which encourages learning. As those connections dissipate, the brain becomes more highly tuned but also more rigid. As a result, it is far easier for a child to acquire language, which requires flexible neurons that can learn and adapt. Later in life, language acquisition is more difficult because the neurons are more fixed.
An alternative theory is equally plausible: after the brain’s breakpoint, we may actually lose many of the neural connections necessary to encode language. Once the brain has acquired a primary language, there is no need to carry those costly networks with us, as there is little evolutionary benefit to having multiple languages. Harvard linguist Steven Pinker provides an eloquent argument for this theory in The Language Instinct: “Language acquisition circuitry is not needed once it has been used. It should be dismantled if keeping it around incurs any cost . . . Greedy neural tissue lying around beyond its point of usefulness is a good candidate for the recycling bin.”
Regardless of which theory is correct, it is now clear that language learning has its own breakpoint, which is distinguished from the network stages of the brain. As University of Texas language professor David Birdsong makes clear of language acquisition, “Typically, there is an abrupt onset or increase of sensitivity, a plateau of peak sensitivity, followed by a gradual offset or decline, with subsequent flattening of the degree of sensitivity.” In other words, there is a growth phase, a breakpoint, and then a period of equilibrium, as Birdsong illustrated.
Image 9.1: The Critical Period in Language Acquisition
III
Learning language is not as easy as merely learning words. As Steven Pinker notes, “If there is a bag in your car, and a gallon of milk in the bag, there is a gallon of milk in your car. But if there’s a person in your car, an
d a gallon of blood in a person, it would be strange to conclude that there is a gallon of blood in your car.” It might take the hundred billion neurons residing in your cerebral cortex a second or two to figure that out. But chances are good that no computer today, regardless of its vaulted silicon IQ, would understand. These are the kind of language and thinking cues that the internet must pick up on. Otherwise, for all of its ability to make calculations a million times faster than the neurons in a brain, the internet will be horribly ineffectual.
Linguists weren’t always interested in the internet, and until recently, internet companies similarly ignored the field of linguistics. Despite the internet’s dependence on words, the general consensus was that it was just too difficult to tie words to meanings. But that all changed with a single innovation developed at Princeton by the late renowned psychologist George Miller. Miller’s innovation was something called WordNet, created in 1985 and perfected over the following 15 years. By the turn of the century, the principles of WordNet were in use across the internet, spurring the need for legions of linguists. For his part, Miller won dozens of awards, including the National Medal of Science from the White House.
WordNet was a bold attempt to categorize and store human language in computers in a way similar to the way the brain stores language. Consider the following:
vehicle ➞ motor vehicle ➞ automobile = auto = car ➞ sports car ➞ Porsche ➞ 911, 944, Boxster, Cayenne, Cayman, Panamera
Almost every word in natural language has both generalizations and specializations of this kind. These relationships form a network structure that sits on top of our neurons in our memory systems. The power of this type of network representation is that it puts specific information in a more general framework that can be used to compute answers to queries.
This is great news for search engines. Without WordNet, when a user types in “Boxster,” all that can be searched is “Boxster.” But with a network structure like WordNet, search engines can also activate the “sports car” and “Porsche” nodes, which elicit even more robust information. The user can quickly discover that Boxsters have high-powered engines, usually seat two people, and aren’t exactly cheap. Word networks are also powerful tools for spell checkers and thesauruses, allowing—among other things—the autocorrect function on your emails to act smarter based on the context of what is being written. It also solves the problem of context for technologies like Apple’s Siri, which tries to understand natural language.
Of course, there is an even bigger problem with language, namely, that words are ambiguous. If words used in natural language had single, well-defined meanings, life would be simple. Unfortunately, that is not the case. 911, as an example, has many meanings. Outside of being a type of sports car, it also conjures up images of September 11 specifically and emergencies in general. Language is a complex, ever-evolving instrument: a quick look at a dictionary will show that essentially all common words have multiple meanings. In fact, the more frequently a word is used, the more meanings it is likely to have. To complicate things further, it is a truism in linguistics that each word means something at least slightly different, so even synonyms are never really true equals.
Consider the two words “board” and “plank.” Both can refer to pieces of wood. The two statements, “He went to Home Depot and purchased a knotty pine board” and “He went to Home Depot and purchased a knotty pine plank,” mean about the same thing. However, both “board” and “plank” have multiple meanings, and these other meanings are completely different. For example, “board” and “plank” are not interchangeable in the following sentence: “The venture capitalist will throw the CEO off a plank if he is not elected to the board of the corporation.”
We humans deal with this very well. As an example, when someone hears the three words “bat, ball, diamond,” she knows the topic is almost surely related to baseball, even though all the words in the string are ambiguous. “Bat” could refer to a flying mammal or a wooden club, “diamond” to a gem or a shape, and “ball” to a dance or a sphere. The common association of the words in the string, “baseball,” is obvious given the context of our language network.
It’s generally effortless for a reasonably intelligent person to choose the right meaning based on context, even when a word has many possible meanings. However, this problem is so difficult for computers that it stopped early attempts at artificial intelligence dead in their tracks. A word is surrounded by an invisible cloud of context and world knowledge that is tapped easily by humans but largely unavailable to computers.
WordNet deals with multiple word meanings by the formation of what Miller called “synsets,” that is, sets of synonyms. Each synset consists of groups of words that share a particular meaning; they act as synonyms, but only for a single meaning. To follow up on the previous example, “board” and “plank” form a synset when referring to the meaning “pieces of wood.” However, both “board” and “plank” have other meanings that are not shared.
WordNet uses a brain science algorithm called “spreading activation” to solve the problem of ambiguity. Spreading activation is a process by which closely connected words wire and fire together, just like neurons. One word activates others nearby in the network. Consider again a bat, a ball, and a diamond. Each word has multiple meanings. Yet each word contains a meaning linked to baseball among its sets of meanings. If we simply excite links connected to each meaning, the “baseball” node will get three times as much excitation as the meanings that are not common to all the words. Spreading activation thus solves the meaning problem: when someone says bat, ball, and diamond, WordNet can tell that baseball is involved (rather than flying mammals or weddings). In this way, WordNet approximates semantic maps with synsets that allow the internet to build context into language through spreading activation.
For a more practical example, consider someone who needs to buy a new outfit. He may search for shirts, pants, shoes, jackets, and socks. Clothing is the unifying category. If each of these specific terms “spreads” activation to parts of the network connected to it, “clothing” will be activated multiple times, once from each subordinate. Such a computation is immeasurably valuable in e-commerce because we know that shirts are rarely sold in “shirt stores” or pants in “pants stores,” but both are found in clothing stores. Knowing the right level of generality needed is a matter of great importance and is often difficult for computers to determine without help. WordNet provides the key.
Without this intelligence, companies across the web are prone to make the same comical errors we expect from children. It brings to mind the first version of Google’s advertising system, which inadvertently placed a luggage ad on a news article describing how a woman was murdered and stuffed into a suitcase. Google went on to buy a company leveraging George Miller’s WordNet and incorporated it into their advertising system. Virtually all search engines, and many advertising systems, now use some of these techniques to make their engines and algorithms “smart.”
IV
Until the early 1990s, it was thought that all neurons worked the same way. They acted as simple processing units that fired as a result of specific stimuli. But in 1991, an amazing discovery happened that has fundamentally changed how we think about the brain.
Giacomo Rizzollati, a physiologist working at the University of Parma in Italy, was studying how neurons reacted when monkeys reached for and grasped objects. Rizzollati implanted electrodes into a monkey’s brain and watched as the neurons predictably fired each time the monkey grasped a handful of nuts. This was nothing new or exciting: Rizzollati was studying processes that neuroscientists had long since confirmed. The body moves in response to the neurons in the motor cortex.
But then something unexpected happened: a graduate student walked into the lab with an ice cream cone. The monkey shifted his attention and watched as the student began to eat the ice cream . . . and the monkey’s motor
neurons became active.
Based on what we knew at the time, this result was just not possible. Individual neurons are supposed to perform single, simple functions. They are insular; motor neurons fire with our own motor movements, not the actions of someone else. And most importantly, they can’t track our actions and the actions of others—it just doesn’t work that way. It would be as if a car responded not just to the pressure of your foot on the gas pedal but also to that of the driver across the street in a different car. The result, in both cases, would be a crash.
As a neuroscientist, Rizzollati understood this as much as anyone. His first conclusion was simply that his gear was broken or that the neurons were connected improperly. So he tested other machines, other monkeys, other brain regions. In each case, he found the same result: neurons firing in response to the actions of others. Rizzollati had found a new type of neuron, now referred to as mirror neurons. (Interestingly, for his efforts, Rizzollati went on to win a prestigious award from the Cognitive Neuroscience Society named after George Miller, the creator of WordNet.)
Over the past ten years, other neuroscientists have duplicated Rizzollati’s results and also demonstrated that these neurons exist in humans. Mirror neurons are truly remarkable and are redefining how we think about the brain. They are especially pronounced in humans; it is argued that we have far more mirror neurons—and that they are more complex—than do any other animals. Neuroscientists now believe that mirror neurons are responsible for much of our cognitive ability, especially in areas of empathy, culture, and language. The always-bold neuroscientist V. S. Ramachandran has gone so far as to predict that “mirror neurons will do for psychology what DNA did for biology.”
Mirror neurons are not homunculi, nor are they intelligent. But they do something other neurons cannot do: make predictions. Mirror neurons in the motor cortex are able to predict the intention of an action (holding an ice cream cone versus eating it). Somehow, they fire only with the true intent; they don’t respond to meaningless actions or random gestures. University of Southern California neuroscientist Michael Arbib places the mirror neuron into its full context: “The neurons, located in the premotor cortex just in front of the motor cortex, are a mechanism for recognizing the meaning of actions made by others.”