FOA Home | UP: Spreading activation search


Basic representation

In AIR, each new document first causes a corresponding document node to be generated. An author node is then generated (if it doesn't already exist) for each author of the document. Two links are then created between the document and each of its keywords (one in each direction), and two more between the document and each of its authors. Weights are assigned to these links according to an inverse frequency weighting scheme: the sum of the weights on all links going out of a node is forced to be a constant; in our system that constant is one. Figure (figure) shows the subnet corresponding to the book Parallel Models of Associative Memory, by J. A. Anderson and G. E. Hinton [REF35] .

The initial network is constructed from the super-position of many such documents' representations. Most of the experiments to be described in this report used a network constructed from 1600 documents, forming a network of approximately 5,000 nodes. This is a trivial corpus, and used relatively crude lexical analysis and keyword weighting ideas. However, AIR requires only that the initial automatic indexing assign some weighted set of tentative keywords to each document.

There is one property of the inverse weighting scheme on which AIR does depend, however. A network built using this keyword weighting scheme, together with similar constraints on the weights assigned author links, has the satisfying property of {conserving activity}. That is, if a unit of activity is put into a node and the total outgoing associativity from that node is one, the amount of activity in the system will neither increase nor diminish. This is helpful in controlling the spreading activation dynamics of the network during querying.


Top of Page | UP: Spreading activation search | ,FOA Home


FOA © R. K. Belew - 00-09-21