ASSOCIATIVE NETWORKS

When we consider links with strengths continuously varying between 0 and 1 (or even between -1 and 1, if we want to express negative or inhibitory links), the interpretation of a link must shift from a generalized "if...then..." relation, to the weaker "is associated with". This brings us from "semantic" to "associative" networks. Associative networks are in principle more general and more flexible, allowing the expression of different "fuzzy", "intuitive" or even "subconscious" relations between concepts. Such networks have been regularly suggested as models of how the brain works. They are similar to the presently popular "neural" networks, except that the latter are typically used as directed, information processing systems, which are given a certain pattern of stimuli as input and are supposed to produce the correct response to that pattern as output. In the present "bootstrapping" perspective, there is no overall direction or sequence leading from inputs to outputs; there are only nodes linked to each other by associations, in such a way that they are coherent with each other and with the user's understanding of the knowledge domain.

In principle, associative networks could be created by the same type of knowledge elicitation techniques underlying THOUGHTSTICKER or CONCEPTORGANIZER, where a user enters a number of concepts and links and is prompted by the system to add further links and concepts under the main constraint of avoiding ambiguity. These links must then be attributed some variable degree of strength. However, the very weak requirement of "associativity" allows virtually any pair of concepts to be linked, if only with a very small link strength. Moreover, there is no obvious generalization of the bootstrapping axiom, which is based on discrete linking patterns, to continuously varying linkages. If everything is linked to everything, then all nodes become uniformly indistinguishable. Finally, it is in practice impossible to let users realistically estimate a strength value for each of the huge number of possible links. Rocha (1991) has suggested a method to "fuzzify" conversation theory, by calculating continuously varying conceptual distances between nodes in an entailment mesh, on the basis of the number of linked nodes they share, but this approach has never been fully worked out.

In psychology, rudimentary associative networks have been created through experiments in which subjects were given a word (say cat), and were asked which other word first came to mind (e.g. dog, mouse, or milk). The more often a certain word b is given in response to the cue word a, the stronger the association from a to b. Since this approach usually only finds a small number of associations for any given word, association strengths for links between other words are calculated by taking into account indirect associations (e.g. knowing the strengths of dog -> cat and cat -> mouse would allow one to calculate the strength of dog -> mouse). Note that such associations are in general asymmetric. For example, when cued with penguin the probability that you would say bird is not so small, whereas the probability to respond with penguin, when cued with bird is virtually zero. This methodology, however, requires a lot of work from designers and users, and is only useful for simple, well-known items like common words.