There has been quite a bit of work on automatic generation of Bayesian networks.
Some time ago (92? --- I'm afraid I'm away from home and from my bib database), Mike Wellman, Jack Breese and I edited an IEEE SMC special issue on the subject, which we referred to as Knowledge-Based Model Construction. That contains a number of articles on the subject. We also wrote a survey article for Knowledge Engineering Review.
More recently, I can think of work by Kathy Laskey and Daphne Koller, Avi Pfeffer and their colleagues. Kathy and some co-workers had something in the most recent UAI proceedings.
David Poole's work on abductive provers/probabilistic abduction/independent choice logic is also very closely related to this problem.
Your particular problem, involving understanding computer programs, is closely related to the problem of plan recognition. Eugene Charniak and I have done some work on a Bayesian approach to that problem, using KBMC, and Chris Geib and I have done some more recent work in this area (as yet unpublished), working with David Poole's probabilistic abduction interpreter.
Cheers,
Robert
P.S. if you need better pointers, drop me a line in a few days.