Friday, June 15, 2007

Symbol grounding and relations

I've been reading a number of Peter Turney's papers and I've lately been catching up on his blog, which has a number of really interesting posts related to his work - and remarkably related to the things I've been pondering about over the last few months.

For instance: machine learning spent years learning how to recognize classes of objects by their attributes; a popular topic now is collective classification, or recognizing the class of an object (in part) by considering how that object is related to other objects. Are attributes only a "convenient fiction" - a useful abstraction that ultimately must be defined in terms of relations? Is an apple intrinsically red, or is "redness" something that describes the interaction of the apple and the sensory system of the observer? likewise is every attribute properly a description of an object and some sensory system or measurement instrument?

This seems like a rather strange and abstract question, but it's intimately connected with the "symbol grounding problem", the subject of another of Peter's posts, which in turn is connected with my long-standing interest in data integration - a very practical real-world problem. To combine data from two heterogeneous knowledge bases is, properly speaking, impossible to do automatically: if they are different formal systems, and there is no surefire way to translate between them. There is no common ground. By the same token, communication between two people is also impossible. How do we know that what I mean by "red" is the same as what you mean?

The solution to the problem, for human communication, appears to be that language is grounded - in part by common perceptual systems. (Goleman's book Social Intelligence is a nice description of some of the ingenious mechanisms that have evolved for establishing this common grounding.) The effect is that I don't really know that what I mean by "fear" is the same as what you mean; an in fact, it may not be the same. But if we're both neurologically typical it is almost certainly highly similar to what you mean.

Back to attributes and relations - I'm not sure, but I think the apparent primacy of relations starts to emerge when you start thinking about these issues. Everything is ultimately defined in terms of relationships with other things, which are finally grounded in our own perceptions--those few things that we don't need words to explain or understand.