A wonderful interview produces a promising tangent.
Last September I mentioned in this column that interviews with legends such as the Caspian Network’s Lawrence Roberts keep me in this business. Well, my interview with Prof. Michael Dertouzos, head of MIT’s Laboratory of Computer Science (LCS), will keep me going for quite some time at my job, which can be both an arduous task and an awesome responsibility.
Having set aside just 30 minutes for our talk, Dertouzos and I overstepped his tight schedule by nearly an hour. He spoke in a deep voice that reminded me of Henry Kissinger with a Greek accent. And even over the phone, I pictured him smiling at me as he has done in each issue of Technology Review since I took this job.
We talked at length about the subject of this month’s cover story: Dertouzos’ Human Centric Computing philosophy, especially as it applies to this month’s topic–Web developments. But we also spent a good deal of time talking off topic. We were chatting about one of LCS’s projects–Tim Berners-Lee’s Semantic Web–when suddenly the conversation headed to deeper waters. It became an exhilarating tangent that I knew at the time would not make the final cover cut. But it seems to fit the mission of this column perfectly.
I mentioned at one point in the interview that I think the Semantic Web is a bit overoptimistic. I studied human language in graduate school for eight years, and in that time I came to appreciate its complexity. The more I came to grips with linguistic complexity, the more my suspicions of conventional semantics grew. Not that it’s wrong, any more than Newtonian physics is wrong. But just as Newtonian mechanics is an oversimplification of physical reality, contemporary semantics is an oversimplification of linguistic reality. That’s my view, anyway.
The upshot is that because Berners-Lee’s Semantic Web is based on contemporary semantics, it will not get at meaning on the Web. To be sure, it will provide an improvement over what we have now. Take search engines. Right now, we are happy to get 25 percent relevant results using Boolean strings in Google (the search engine that bills itself as the most relevant). Through what Dertouzos calls synonym-based algorithms with specialized versions of XML, the Semantic Web could improve those results to perhaps 50 percent. And it will improve other aspects of Web-based communication as well. But it will never approach Berners-Lee’s goals of machine understanding (except in predefined contexts) because of the limitations of semantics.
So far, we were not that far afield from our cover conversation. That soon changed. I was surprised by how willing Professor Dertouzos was to accept my argument. It is not a very popular view. Modern semantics and its basis in analytic philosophy is a cornerstone of academia in the English-speaking world. Still, quoting philosophers as far-flung as Lao Tzu and Aristotle, Dertouzos revealed that he agreed with me. In his view, this problem is at the heart of the failure of Artificial Intelligence in all its forms, and no one has yet been able to solve it. But he thinks the Semantic Web will do a lot to advance the cause of a human-centric Web nonetheless.
Because we were already over our time, I expected him to make a graceful exit at this point. I had everything I needed for the cover story and was prepared to sign off when it seems his academic instincts kicked in and he asked me the quintessential thesis-committee question: “What do you think could do a better job?”
I told him that no linguistic model can capture human language precisely–some problems would take our fastest supercomputers hundreds of years to solve. But the closest model to what we really want–more relevance to our day-to-day affairs–is fittingly called Relevance Theory (RT). Developed by linguists Diedra Sperber and Dan Wilson, RT does to modern semantics what Relativity Theory did to Newtonian physics.
First of all, rather than basing the theory on meaning, Sperber and Wilson base it on relevance. RT states that relevance is relative to lots of conditions between communicators–what they know, their families and cultures of origin, the circumstances of the exchanges, etc. RT builds all these factors into its model. Anyone in a committed relationship knows that even though you know your spouse better than anyone on the planet, you still spend lots of time explaining what you mean by simple, face-to-face statements. Add a layer of technology in between, like e-mail, and your odds of successful communication go down. Get on the Web and you’re often at sea.
Rather than focusing just on the meaning of successful communication in idealized environments, RT can tell us why our communication is more or less relevant to our needs in everyday cases. By building in these complexities and allowing for degrees of success in communication, RT does a much better job of modeling human language. And I think it would be a better choice as the basis for a project like the Semantic Web. Dertouzos listened carefully, and after I was done with my little spiel, he said it might work, though he was concerned about the computing power needed to deal with all the extra variables. I guess that’s the beauty of the Semantic Web; it may not be perfect, but it’s elegant, and it works fairly well with the tools we have.
Dertouzos thanked me for the “thoroughly enjoyable” discussion and we finally said so long. I sat at my desk pondering what had just happened. I’ve often thought that when the time comes for me to leave this job, I would take up the project of programming a search engine based on RT. But I’ve always said in the same breath I would finish the great American novel as well. In other words, it’s more of a pipe dream than a goal. Still, the talk made me think it’s worth pursuing at some point. It’s not that I’m in any hurry to leave, but it would be nice to actually improve people’s computing lives rather than merely serving as an advocate for computer users and living for interviews with computing legends.
James Mathewson [email protected] is editorial director of ComputerUser.