As far a linking SPs into larger patterns I have been promoting the work of William Calvin where he proposes organizing individual columns in a minor variation of the stock SP where the columns organize into a grid configuation.
The strong point of this mechanism is that local recognition of some bit of a larger pattern reinforces other columns if they are also recognizing some other part of the same larger pattern.
I would like to think that this links the individual columns if they are semantically related.
One of the by-products of this method is that large numbers of patterns can be intermingled and yet not interfere with each other. Cells that learn many patterns can also be re-used in many larger learned patterns.