Return of the Guru

AI in Insurance? Elementary, My Dear Watson

Ara Trembly
Insurance Experts' Forum, February 22, 2011

Perhaps like many of you, I watched with great interest the recent grand showdown between IBM’s Watson computer and two of the most successful Jeopardy champions, which turned out to be a romp for the aptly-named collection of servers that had been programmed with all kinds of factual information from almanacs, encyclopedias, and a plethora of other human sources.

While that result undoubtedly shocked few, IBM was quick to stress that the purpose of this event was to demonstrate the strides we have made in helping computers to understand natural human speech. If you think of the mythical computer that inhabits the Starship Enterprise, a machine that can answer all questions put to it in colloquial English, you’ll get an idea of the level of communication accuracy that is, or should be, the goal of such research.

Did Watson achieve that level of excellence? Alas, it did not. The television event was made more interesting by the answers (or, since this is Jeopardy, the questions) Watson got wrong, rather than those it gave correctly. The computer had no trouble with straight factual responses. Watson instantly knew the second largest city in New Zealand, while it seemed his human rivals could only ruminate and surmise. When it came to shades and subtlety of meaning, however, Watson was much less sure. The human contestants easily knew that “what is a crook?” was the correct response to “a thief or the bent part of an arm.” Watson’s top two guesses, however, were “knee” and “waist” (“crook was third”)—clearly wrong responses, since neither is related to “thief.”

Still, imperfect as they are at understanding our language, expert knowledge systems like Watson may indeed be useful tools for insurance enterprises in the near future, according to Jamie Bisker, global insurance industry leader, Institute for Business Value, IBM Business Consulting Services. Noting that AI (artificial intelligence) covers a number of disciplines from natural language processing to robotics, Bisker points out that “60 to 80% of underwriting” involves some form of AI.

He adds that expert systems—knowledge-based systems or rule-based systems—are already in use in insurance enterprises. In the future, says Bisker, question answering systems like Watson could contain information on actuarial science, state and federal regulations, insurance processes, and other critical facts in the insurance domain. If, for example, an insurer wanted to determine legality of its products across state lines, a dialog with the computer system would help the actuary to correct or create such a product, or understand why they are getting fined by another state.

“It’s like saying [to a knowledgeable colleague], ‘What do you think?’ about this problem? It can’t answer everything, but it would empower the people who use it to be more efficient,” says Bisker.

Sounds good as far as that goes, but just how far are we willing to go with expert systems, and what effect will that have on humans—and their jobs? I’ll delve into the theoretical side of this a bit more next time.

 

 

Comments (1)

IBM sets many goals for its research, and the computers in science fiction as envisioned in Star Trek were indeed part of the inspiration for Watson. However, what the Gentle Blogger thinks should or shouldn't be the goals for IBM Grand Challenge projects is not what drives such research. The BHAG (big, hairy, audacious goal) of competing against expert "Jeopardy!" players occurred to an IBM executive after seeing almost an entire bar full of people get up en mass to see if Ken Jennings could keep his winning streak alive. He wondered if a computer could ever compete at that level and thus began the idea for Watson.
And as you mentioned, a primary aspect of playing a game like "Jeopardy!" and one of IBM's interests in pursuing it is developing a better understanding of how to enable computers to deal with natural language. The grand challenge was to take on this game at a level that would allow it to not only play, but play against those that clearly had few, if any, rivals. And, while I hold the Library Computer System of Star Trek in high regard, it was rarely faced with queries that were intentionally vague, misleading, colloquial or rhyming in nature as is the case with Jeopardy! categories and clues. So the fact that Watson misunderstood a few is not really of any more concern than those that the other players missed.
In regard to categorizing the type of system that best describes Watson, it would be best to steer clear of the "expert system" moniker in general. Inside its many processes, algorithms, and sub-programs, Watson likely uses some rule-based mechanisms, but it clearly demonstrated that its internal workings do not suffer from the frailties of those type of systems. Working with unstructured data such as natural language and the text of the large amount of data reference material loaded into its memory would have broken any traditional rule-based system.
My goal is to explain how Watson works and to pursue why it would be a useful tool for insurers to consider when it is commercially available in a few short years. I look forward to reading your next installment and continuing the conversation.

Jamie Bisker

Posted by: jbisker | March 1, 2011 6:18 AM

Report this Comment

Add Your Comments...

Already Registered?

If you have already registered to Insurance Networking News, please use the form below to login. When completed you will immeditely be directed to post a comment.

Forgot your password?

Not Registered?

You must be registered to post a comment. Click here to register.

Blog Archive

What It Takes to Have a Tech-Savvy Workplace

The tools and technologies to build the next workplace are available, but not common yet in corporate settings.

Avoiding the Bermuda Triangle of Data

Handled poorly, questions around data ownership, data quality and data security can sidetrack big data conversations and alienate business stakeholders.

A Prototype of the Successful Innovation Leader

Celent research reveals the prototype for the successful senior innovation leader.

Global Supply Chain, Local Problem

As a technology provider, your client’s ability to deliver products and services to their customers, when and where they need them, is at the heart of their business success.

Legacy Systems Are Increasingly a Competitive Handicap

Legacy systems, while reliable, increasingly hold insurers back, a new study finds

From Her to Watson, and What’s Next?

Imagine a learning system that can replace the performance of your best employee to provide the same level of support across the organization.