Symbol Grounding

Patrick Logan suggests that the problems being discussed by Jon Udell are related to Harnad's Symbol Grounding Problem. I'm not sure this is quite right, but the Problem keeps cropping up, mostly in the arguments about what use the Semantic Web is.

The "Problem" is only a Problem if we are trying to make computers "think", or "know". AI people crashed up against it because they were trying to do just that. But in failing to achieve intelligence, they produced a raft of useful tools and techniques, all of which stop short of grounding symbols. And it will stay that way: symbols must be grounded by humans. If we develop intelligent machines, or if we develop machines capable of evolving which then evolve intelligence, then it's a bit of a leap to suppose that this machine intelligence will have remotely the same mechanism of grounding symbols as we do. Vocabularies represent negotiated sets of symbols, grounded by means of natural language definitions: they may be good or bad, ambiguous or unambiguous, controversial or widely accepted. There will always be overlaps, gaps, clashes, and drift.

This problem is one which results from, but is not itself, the Symbol Grounding Problem. Since all the symbols we use in these manipulation systems we are building must be grounded, and they must, as just mentioned, be grounded in human understanding of the meaning of those symbols, we suffer all the fuzziness and ambiguity of which semiotics is the study (not that I claim to be an expert on semiotics, I've been dragged into thinking about it by reading Mike Abbott's work and desperately want to find time to explore further).

Another way: because the symbols we use in our thought and communication are grounded at some other, non-symbolic, level, they squirm to be free of the straight jacket of the Symbol System (Alan Newell, 1980).

Asking for a "universal aggregator", able to do something with any embedded data -- meta or otherwise, and in RDF form or more basic XML -- is asking for precisely what the AI people were hoping for. A machine which is intelligent in the same way as us. We don't want this, however much we might think we do. We use computers precisely because they are good at different things.

But this is no reason for everyone to give up and go home. Current efforts with RDF and Topic Maps are trying to expand what computers are good at, to make them better at manipulating information which we can then make use of. The "Semantic" of the Semantic web is one of named associations between nodes. Higher level semantics are still understood only by humans, who write software based on that understanding which does something with the data, something which is of use to other humans. If data is flowing around in a structured form, even just plain XML, then it makes it easier to add plugins to tools such as aggregators and PIM software to handle new forms of data, it doesn't mean that the aggregator software can handle all well formed data which is thrown at it.

The complaint that agreeing on the meaning of terms is too hard is a bit of a red herring: tools will simply work better, more transparently, where that agreement exists.