- From: Drew McDermott <drew.mcdermott@yale.edu>
- Date: Sun, 23 Nov 2003 21:27:18 -0500 (EST)
- To: www-rdf-rules@w3.org, clay@shirky.com
Bijan Parsia pointed me in the direction of this anti-Semantic-Web article by Clay Shirky: http://www.shirky.com/writings/semantic_syllogism.html It's worth reading. (Full disclosure: He praises an old paper of mine; this is not the only reason to read it.) The paper argues that deduction (which he calls "syllogism" for no reason I can see) is hopelessly inadequate for realistic applications. I half-agree with him. I think he underestimates the need for deductive rules in tasks such as datatype transformations; but it's equally true that many of the people involved in the SW are overoptimistic about how much mileage can be gotten from deduction in supporting things like reasoning about contractual obligations. The perennial debate (recently revived on www-rdf-rules@w3.org) about the need for negation-as-failure illustrates the point. Those who deny the need for NAF believe that somehow deductive methods will arise that can draw conclusions of equivalent use monotonically. (Yes, I know that one can view NAF as a simple abbreviation convention for inferences that are really deductive, but in practice nonmonotonic inference is a device for _escaping_ deduction.) Consider the planning algorithms that are an important application of OWL-S. Are they deductive? Some are, some aren't. For others it's hard to say. The fact is that computation is a more important category for the Semantic Web than deduction -- just as it is everywhere else. It is usually much easier to think about algorithms as producing outputs than as producing conclusions. These outputs often achieve status as "conclusions" as a pragmatic postprocessing phase. E.g., a planner's output is taken as a recipe for guiding behavior. The agent using the planner concludes that this is the best course of action for it to take. It may be, but for a self-justifying reason: the only planner the agent has couldn't come up with something it thought was better. Another example: Turbotax concludes that you owe a certain amount of tax. Is that a deductive conclusion? Possibly. But it doesn't produce a proof, and it would be rather difficult to produce one. Another: A vision program might conclude that you are in the room. This is clearly not a deductive conclusion. Etc., etc. It's annoying that Shirky indulges in the usual practice of blaming AI for every attempt by someone to tackle a very hard problem. The image, I suppose, is of AI gnomes huddled in Zurich plotting the next attempt to --- what? inflict hype on the world? AI tantalizes people all by itself; no gnomes are required. Researchers in the field try as hard as they can to work on narrow problems, with technical definitions. Reading papers by AI people can be a pretty boring experience. Nonetheless, journalists, military funding agencies, and recently the World-Wide Web Consortium, are routinely gripped by visions of what computers should be able to do with just a tiny advance beyond today's technology, and off we go again. Perhaps Mr. Shirky has a proposal for stopping such visions from sweeping through the population. -- -- Drew McDermott Yale Computer Science Department
Received on Monday, 24 November 2003 11:49:51 UTC