EXPERIMENTAL RESULTS

In order to test these ideas in practice we have conducted two experiments. We built a network consisting of 150 nodes, corresponding to the 150 most frequent nouns of the English language. Every node was assigned 10 links to other nodes. These links were randomly selected from the 149 remaining nodes to initialize the web, but would then evolve according to the above learning rules (with t = 0.5 and s = 0.3). We made the web available on the Internet, and invited volunteers to browse through it, selecting those links from a given node which seemed somehow most related to it.

For example, if the start node represented the noun dog, a user would choose a link to an associated word, such as cat, animal, or fur, but not to a totally unrelated word, such as mathematics. Of course, in the beginning of the experiment, there would be very few good associations available in the lists of 10 random words, and users might have to be satisfied with a rather weak association, such as meat. However, when reaching the node meat, they might be able to select there another association, such as carnivore. Through transitivity, a new link to carnivore might then appear in the node dog, displacing the weakest link in the list, while providing a much better association than the previously best one, meat.

Although there were some unexpected side effects, such as the development of approximate `attractors', the development of the associative network was surprisingly quick and efficient (Bollen & Heylighen, 1996). After only 2500 link selections (out of 22500 potential links) both experimental networks had achieved a fairly well-organised structure in which most nodes had been connected to large clusters of related words. This may be illustrated by a typical example of how connections are gradually introduced and rewarded until their strength reaches an equilibrium value (Table 2). The position of these associated words shifted upwards in the list until they reached a position that best seemed to reflect their relative strength.

Table 1: self-organization of the list of 10 strongest links from the word "knowledge", in different stages: initial random linking pattern, after 200 steps, after 800 steps, and after 4000 steps. A step corresponds to a user selecting a link on one of the 150 nodes, in a web that evolves according to the direct, transitive and symmetric learning rules.

The net result of the experiment was a 150 x 150 matrix of association strengths, which reflected fairly well the most important intuitive associations existing among the concepts. For example, a cluster analysis performed on the matrix produced 9 general categories (Time, Space, Movement, Control, Cognition, Intimacy, Vitality, Society, Office), grouping most of the related words in a single class. Such statistical clustering techniques may provide the equivalent for associative webs of the discrete conceptual clustering we discussed earlier.

We are now trying to determine to what degree these results from our learning web correlate with different word associations derived by other means (e.g. free association experiments, or letting people judge the degree of synonymy). We also plan to test the usefulness of the self-organization, by checking in how far users find knowledge more effectively in a self-organized network, as compared to a network that did not undergo learning. This can be done by measuring the average number of steps needed to find a node, and the average time needed to choose a link. We are further considering additional learning rules, such as similarity (nodes sharing several links would get stronger cross-connections), that may make the learning more effective. Although this research is still in its initial stage, and will need much empirical testing to confirm its usefulness, it seems like a very promising approach to quickly and easily develop complex associative networks that are more adequate than hypertexts built manually.