... true in the FLAT model mayonly be true for, say, Indo-European, and the remain-ing languages are considered noise. While this maybe interesting in its own right, we are more interested in discovering ... complicated. Instead of perform-ing a simple Gibbs sample for m in Step (4), wefirst sample the m values for the internal nodes us-ing simple Gibbs updates. For the leaf nodes, weuse rejection sampling. ... nor sufficient for the model to be good. For instance, finding 1000 implications of the form A1⊃ X, A2⊃ X, . . . , A1000⊃ X iscompletely uninteresting if X is true in 99% of thecases. Similarly,...