After adaptive hypertext networks have been trained by the Adaptive Web system, all links have also been assigned weights: their weight expresses the degree with which items in the network are associatively related. These values can be used by automatic retrieval mechanism to more accurately and flexibly locate and retrieve information from the network.
Spreading Activation one of the well-know retrieval mechanism in associative memory networks as described and implemented in cognitive psychology and AI. Spreading activation is a proces whereby an associative memory network is brought to retrieve certain items from memory by the initial activation of several cue-concepts (partial 'descriptions' of what is being looked for). Activation is propagated parallelly through the network from these concepts on. Activation can accumulate in certain nodes in the network that can then be selected for retrieval.
Spreading Activation in associative network can easily be simulated by an interative proces in which an initial activation vector is repeatedly fed back into a multiplication with the networks weight-matrix. Each consequent activation vector indicates to where activation in the network has spread. We have made available an on-line demonstration in which a network that has been trained by our Adaptive Hypertext system, can be queried by the process of spreading activation. (try activating "building" and "work", or "woman" and "father")
The properties of Spreading Activation are such that even very partial, associatively described information can be retrieved from the network. The parallel spreading of activation over several chains of associatively related items in the network and the condition that activation must accumulate in certain network nodes, makes that the nodes that are truly at the associative "center" of the cued items have highest probability of being retrieved. Because Spreading Activation acts through the activatiion of multiple and often quite long chains of activation it can deliver highly original and unexpected results.
Spreading Activation can as such not only relieve the users of formulating lexical/boolean search-keys, but is also capable of more intelligent and accurate searches than systems based on the lexical comparison of search strings and items stored in databases.