A bright future awaits computer system Watson now that it has shown it can compete with previous winners of the complicated US quiz
Read more: Our technology blog will report on Watson鈥檚 final Jeopardy! appearance on Wednesday
A medical robot; a Google-killer; a financial adviser; a tool for trawling legal documents; an aide for the intelligence services.
Advertisement
These are just some of the careers that could be in store for Watson, a supercomputer created by IBM which this week
The show demands an encyclopaedic knowledge and an ability to understand complex verbal clues. Watson will face the two most successful players ever to appear on Jeopardy!. By proving that it can compete with elite human players, Watson gives us a glimpse of a future in which a new breed of smart software permeates our lives.
When IBM first suggested a Jeopardy! machine four years ago, many artificial intelligence researchers were doubtful the company would succeed. Computers are great at following clearly-defined rules, but Jeopardy! questions can involve puns and clues-within-clues, with topics ranging from pop culture to technology.
Jeopardy! clues are also unusual in that they are phrased as answers and contestants are required to supply the right question. For example, if the host says that 鈥淭his cigar-smoking prime minister led Britain during the second world war鈥�, a contestant might reply 鈥淲ho is Winston Churchill?鈥�.
Creating a set of rules that a machine could use to understand and answer the game鈥檚 questions would be impossible, as the questions are too complex. Instead, IBM鈥檚 team, lead by David Ferrucci based in Yorktown Heights, New York, gave their computer a few rules and a huge amount of memory and processing power.
Huge databases
The team loaded 200 million pages of text from encyclopaedias, newspapers and other sources into Watson鈥檚 memory. They also added custom-built stores of data on geographic facts and other types of information.
Watson starts by guessing the subject of a question. It knows, for example, that the word 鈥渢his鈥� often precedes the subject. So, if a question begins 鈥淭his 19th-century novelist鈥︹��, Watson primes itself to search for writers.
Once it has worked out what it is looking for, Watson searches its huge databases for the answer. Thanks to a processor array that is around 2000 times more powerful than a desktop computer, Watson is able to search the databases and come up thousands of possible answers in less than 3 seconds.
To decide which answer is most likely to be the right one, Watson uses 100 different tests to rate its confidence in each answer.
For example, if Watson thinks that 鈥�1961鈥� is the answer to 鈥淧resident Obama was born in this year鈥�, it would search for 鈥淧resident Obama was born in 1961鈥�. If it sees matches in sources that it trusts, like encyclopaedias, Watson ups its confidence in that answer.
The tests also check for word-play, such as puns. Faced with a question about an 鈥渁rresting landscape painter鈥�, Watson looks up meanings of 鈥渁rresting鈥� and checks for connections with the names of famous landscape painters. Linked answers 鈥� in this case, 鈥淐onstable鈥� 鈥� get a higher confidence score.
Full of confidence
Once it has carried out these tests, Watson takes the highest-rated answer and, if the confidence level attached to it is high enough, tries to buzz in and answer the question.
None of the tests represents a major breakthrough in language processing, but by combining them all Watson has leaped ahead of rival question-and-answer systems. The outcome of the pre-recorded final, which will begin airing in the US tonight, is still a carefully guarded secret. Based on practice games in January, Watson is expected to win.
Whatever the outcome, the fact that a machine is able to compete against the most successful players ever to appear on Jeopardy! 鈥渋s a remarkable achievement鈥�, says , an artificial intelligence researcher at the Massachusetts Institute of Technology.
Watson wasn鈥檛 built just to play in game shows, however. The software could help with any task that involves a large number of text documents. For instance, if patient records and a database of medical literature are uploaded to its 鈥渂rain鈥�, Watson would become 鈥淒r Watson鈥�.
, an IBM researcher, imagines a situation in which voice-recognition technology is used to relay a patient-doctor conversation to Watson. Watson鈥檚 remarkable speed means that by the end of the conversation, it would have searched its databases of medical literature and would be able to suggest possible diagnoses, each with its own confidence level.
Intelligence services might also have a use for Watson. Members of the Watson team are part of a machine-reading research programme funded by the (DARPA), a US government body based in Arlington, Virginia.
Report overload
DARPA is interested in machine reading in part because intelligence analysts are faced with far more reports than they can synthesise, says , a computer scientist at Carnegie Mellon University in Pittsburgh who worked with the Watson team. Instead of having to study every report, analysts could use Watson to extract information on, say, how individuals discussed in the documents are linked.
Watson鈥檚 potential also extends to everyday tasks. It is not currently connected to the internet, but if it were given an index of the web, it could take a question expressed in natural language and return a one-line answer. 鈥淲atson could replace Google for some kinds of searches,鈥� say Nyberg.
Like a human brain, a Watson search engine would continually be adding to its knowledge. The software could 鈥渓earn鈥� by searching for new documents to add to its database, so it would automatically update its database of search results as new facts emerge. Prem Natarajan of Raytheon BBN Technologies in Boston, Massachusetts 鈥� one of the companies involved in the DARPA project 鈥� compares this application to an oracle.
Thankfully, perhaps, Watson isn鈥檛 fail-proof: it can find many simple questions baffling. This is in part because the software relies heavily on finding text that looks like the right answer to a question. As a result, it misses out on information that is too obvious to have been written down. Not only that, but Watson sometimes thinks that fictional characters are real, says Gondek. It once named the first woman in space as 鈥淲onder Woman鈥�.
To get a grasp on such common sense issues, Watson will need to improve its understanding of the questions it is asked, rather than just searching for answers that look plausible. And that鈥檚 a much more difficult challenge.
鈥淲atson is an awesome machine,鈥� says Katz. 鈥淏ut sometimes it makes amazingly silly mistakes. That should tell researchers that we鈥檙e not done. We鈥檝e not even scratched the surface.鈥�
