Thanks for the vote of confidence. I should say that while I think my paper presents things well, I cannot take credit for the statistics or experimental design. Tim van Gelder already had the machinery in place to measure pre-post gains, and had done so for several semesters. The results were published in Donohue et al. 2002. The difference here was that I took over teaching, and we continued the pre-post tests.

Although argument maps are usually used to map existing natural language arguments, one could start with the map. I like to think that the more people use these maps, the more their thinking naturally follows such a structure. I'm sure I could use more practice myself.

Just a note on terminology: the tree does have two kinds of nodes, but by virtue of being a tree, it is not a <a href="http://en.wikipedia.org/wiki/Bipartite_graph">bipartite graph</a>.

I think arguments in argument maps can be made probabilistic and converted to Bayesian networks. But as it is, it takes long enough just to make an argument map. I've recently discovered <a href="http://lac.gmu.edu/">Gheorghe Tecuci's work.</a> He's just down the hall from me, but I didn't know his work until I heard him give a talk recently. He has an elaborate system that helps analysts create structures very much like argument maps by filling in schemas, and then reasons quantitatively with them. The tree structure and the simplicity of the combination rules (min, max, average, etc.) are more limited than a full Bayesian network, but it seems to be a very nice extension of argument maps.