In the early 1970’s, while an Assistant Professor at Stanford, Roger Schank achieved worldwide fame when he was the first to get computers to be able to process typewritten everyday English language sentences.
In order to do this, Schank developed a model for representing knowledge and the relationships between concepts that enabled his programs to predict what concepts might be coming next in a sentence. This spawned an entire field in psychology devoted to determining how people make inferences from what they hear.
After moving to Yale in 1974, Schank worked on getting computers to read newspaper stories. His work was heavily funded by the U.S. Department of Defense, which was interested in trying to get computers to predict world trouble spots by reading the news and analyzing it. He built the first newspaper story-reading program in 1976. Five years later, Schank was made Chairman of Computer Science at Yale and ran their Artificial Intelligence lab.
In order to get computers to know enough about the world to tie sentences together, Schank came up with the notion of a script. Scripts were needed to keep the inferences that computers made from exploding exponentially. For example, a computer could understand that what you order is what you eat in a restaurant if it had a set of expectations about what happened in a restaurant (the script.) Scripts were a powerful idea that enabled Schank’s machines to read about any subject that was well structured. Psychologists began testing people to see if they operated with scripts as Schank had suggested and the evidence was overwhelming that Schank had discovered something important about people even though he working in computer science. His book with Robert Abelson on the subject, Scripts, Plans, Goals and Understanding: An Inquiry Into Human Knowledge Structures became a classic, overwhelmingly cited by social scientists for years to come.
While in the process of getting the computers in his lab to understand news stories, conduct conversations, answer questions, tell stories, and imitate other human cognitive capabilities, Schank began to realize that something important was missing from his computer models. They did not have the memory capabilities that humans have. This seemed odd at first, because humans would appear to be less capable than computers when it comes to sheer capacity. Entire volumes can be “remembered’ by a computer after all. No human can match this feat.
What was missing was the ability to generalize. Schank’s programs would read stories that were coming in but they would fail to notice if the same story came in twice or if one story was an update on a situation found in a previous story. In short, while the programs could answer questions and summarize what they had read, they really could not remember what they had read. They had no real understanding because they could not see events as being similar, so they didn’t get smarter as a result of what they had read. You could keep saying the same thing to them and they wouldn’t notice. Schank began to realize that understanding and memory and the ability to generalize were all really the same thing.
Schank began to turn his attention to learning. He believed that if we could understand how people learn then we could apply that knowledge to getting computers to understand more deeply. The story-understanding problem was, after all, really a learning problem. The computers hadn’t learned anything from what they had read. To get smarter over time, the computers needed to match new information to old information—in other words, they needed a clear model of what they already knew in the first place.
So, Schank began to build knowledge of real world events into the computer so that new events could be matched to that information. At this point Schank noticed two phenomena about people’s memory processes that were critical to learning: reminding and expectation failure.
People get reminded all the time. A person reminds you of another person. A place reminds you of another place. And, an experience you have reminds you of another experience you have had. Schank began to study how reminding works. He observed that remindings revolve around expectation failure. You expect something to happen as similar situations have happened before-- and if it doesn’t, you wonder why. Schank realized that people have expectations about everything, what word will come next in a sentence, what a person is likely to do next, what things will look like, what will happen as a result of actions taken, and so on. When these expectations fail, people must make sense of what happened. They can’t continue to be surprised by the same things. At some point, they need to modify their expectations to include things that they hadn’t been originally able to predict. When a restaurant serves bad coffee, at first you are surprised. Eventually you predict the bad coffee and stop ordering it. This is learning in its most rudimentary form.
In order to get computers to learn, Schank realized, they would have to have expectations and they would have to know when new events failed to meet those expectations. They would then have to explain the expectation failures and modify their expectations.
At the same time that Schank was doing this work, psychologists at Stanford who were still following what Schank was doing from the days when he was on their faculty found some interesting experimental results. They discovered that people confuse very different events in their memories when those events have certain similarities. They argued that Schank’s scripts didn’t explain this. Schank agreed and he modified his theory in a way that accounted for the Stanford data and his new concerns about learning. He suggested that people stored memories in packages that were concerned with event groupings smaller than “restaurant.” If a person left their wallet somewhere they might not remember where but they would know that they had used the wallet in a “paying” event and they would try to reconstruct where various paying events may have taken place by seeing if they could connect them to the script of which they were a part. (Maybe it was in a restaurant they had eaten in.) This meant that people were learning and storing memories inside small packages of expectations (Schank called them Memory Organization Packages.) This accounted for the psychologist’s data and enabled Schank to begin to build knowledge into the computer in such way that learning could take place. This was the basis of his most famous book, Dynamic Memory: A theory of reminding and learning in computers and people.
Around this time something happened that changed Schank’s focus permanently-- his children went to school.
Part 1 of 2