Semantics and CAs

A big question is why bother with CAs? We can do most of this processing quite well with statistical and symbolic models.
The first answer is semantics, in particular the symbol grounding problem.
When traditional symbolic parsers combine the and dog, they are just symbols.
Statistical correlations can be drawn (e.g. word pair frequencies), but there really is no semantic basis.
CAs can categorise patterns based on their occurence in the enviornment.
A CA is a group of features that tends to travel around together.
Theoretically, a CA is a symbol that is grounded. (I have no immediate plans to introduce an interface to the real world, but a primitive one could be started immediately.)