We (@ecconrad @olzama and I) have discovered a case where grammars created by the most recent version of the Grammar Matrix are significantly slower to process with (using ace) than grammars produced from the same choices file but the pre-append list version of the Grammar Matrix. On inspection of sample sentences, the global ambiguity is the same and the parse charts look the same (so same number of passive edges; the profiles produced by art aren’t recording this info for us so we can’t verify that way).
The grammars have the same set of rules, and though they are not identical, we suspect that the problem is more likely to be in the cost of unification rather than in say extra active edges. The switch over to append lists isn’t the only change between the two versions of the Matrix, but it is a big one. I’m wondering whether there is some optimization for unification with diff-lists which doesn’t automatically roll over to append-lists or (worse) if there is something inherently less efficient about the append-lists.
@sweaglesw and @guyemerson would you be interested in investigating whether the append-lists are in fact the source of the slow down? @ecconrad has sample grammars & profiles she can share.