On February 14, 2011, Alex Trebek will play host to a battle of man versus machine.
In a special edition of Jeopardy, IBM’s deep analytics question-answer computer “Watson,” named after IBM president Thomas J. Watson, will play against top Jeopardy winners Ken Jennings and Brad Rutter. IBM researchers hope to answer the question of how a computer will fare in a competition of knowledge retrieval and natural language processing when it’s up against human players.
The rest of us are hoping for an answer to the question of whether a computer can ever be smarter than a human. Although it is tempting to pit this as a fight between artificial intelligence and human intelligence, it is much more important to consider it as an event with profound implications for understanding the nature of intelligence and what it means to be an intelligent being.
Watson represents a big step in IBM’s long journey toward creating innovative intelligent machines. IBM’s previous landmark was a computer named “Blue Gene,” which scientists used to map the three billion base-pairs of the human genome. For Watson, the issue is instead to tackle a different but equally complex task: natural language.
“Language was an area that, even at the beginning of the computer era, people believed computers would be good at,” says Dr. Bill Murdock, a Watson algorithms researcher. “So far, computers have failed.”
From the beginning, the problem has always been “open question answering.” This problem is very different from a simple search task that everyday computers are built for. It also more closely resembles how humans actually communicate.
“People can understand language because we relate it to our own thinking and our cognition,” says Murdock. “Language is grounded in our experiences — not in a formal mathematical language that computers can only understand. Computers understand unambiguous things, not like human language.”
Indeed, one of the most curious and impressive challenges will be testing Watson’s ability to discern the many nuances, regionalisms, slang words, and short hand terms that run rampant in Jeopardy questions.
“Jeopardy is a playing field by which we can do some science,“ says Dr. Chris Welty. He explains that Jeopardy producers were at first hesitant about the idea, since they did not want the event to be a mere stunt or gimmick for their show. They reconsidered, however, when they realized the idea was not something to be passed off — a lucky break for IBM researchers, since Jeopardy provides the perfect conditions for testing a natural language processing machine.
“Jeopardy as the broad domain aspect asks all kinds of questions — something we really wanted to take on,” says Welty. “And you have to work quickly. Technology must be responsive. We needed to make a system that can extract unique information from a large amount of general information, and faster than a human can.”
Early tests showed that Watson’s processing speed would be a major problem to overcome. Researchers used a process of putting as many algorithms as possible into the system, and then seeing which ones they could trust to provide a correct answer. The system works on parallel processing. Once Watson is fed a question, it activates all relevant algorithms, and cross-references them with the question. If there is a high degree of overlap or similarity between certain algorithms, this increases Watson’s confidence that it has found the correct answer.
Watson can also input answers into its own system to see if the Jeopardy question appears. This further reinforces its confidence in its answer, highlighting the statistical basis of Watson’s processing.
As one might guess, to be a competitive player at Jeopardy, Watson would need as many algorithms as possible, representing all available knowledge. Luckily, this obstacle was overcome thanks to the Internet, where vast amounts of information are now available digitally.
Watson represents IBM’s most ambitious foray into deep analytics and natural language processing. However, Watson’s early test matches were difficult for IBM researchers to watch:
Watson: “I Love Lucy for 800.”
Question: “It is Ricky’s signature tune and later the name of his club.”
Watson: “What is song.”
According to Dr. David Gondek, “Watson didn’t have a good notion of what the answer type was, or what it was even being asked.”
It would take two years for Watson to play at a competitive level. This was accomplished by building a computer the size of a classroom.
Dr. Eduard Hovy of the University of Southern California says the implications for such a machine will become more apparent and astounding once we can imagine a Watson the size of a PDA.
A more practical-sized Watson could be used in the healthcare industry. IBM researchers hope to build a Watson that can store all the medical information in the world — from illnesses to tried treatments — in order to provide information to doctors when treating patients. Doctors would be able to ask Watson to name all the treatments that have been performed for a particular ailment in the past, and choose one accordingly.
Lawyers could also benefit from Watson when searching for precedent cases, by asking it to name all of the cases similar to the one at hand. Watson would essentially make information retrieval more efficient using its ability to understand human language.
“Can you imagine computers communicating more fluently in natural language?” asks Welty. On February 14, we won’t need to.
