Finds the best tree (relative to the gold tree) s.t. it's reacheable given the current anchoring.
Best is measured as number of correct labeled spans, as usual. If the given treebank symbol is correct,
bonus points can be awarded for getting the right refinement.
On the training set, the "best" reachable tree will not always (~5% of the time) be the correct tree,
because pruning will remove the right answer. We don't want to try to train towards an unreachable tree,
because the training algorithm will do bad things. Instead, we want the best possible tree that
our parser could conceivably produce. That is why this class exists.
If backupGrammar is provided, it will be used to find such a tree in the case that no tree can be
found with grammar (given the current constraints).
Typically, the first grammar will be a treebank grammar that has no horizontal markovization (i.e.
it is not forgetfully binarized) and it also remembers the functional tags like -TMP.
The backup grammar is usually the grammar
with which the pruning masks were produced; because of the way we prune, that parser will always
be able to find a tree (assuming that it was able to find a tree without pruning.)
Finds the best tree (relative to the gold tree) s.t. it's reacheable given the current anchoring. Best is measured as number of correct labeled spans, as usual. If the given treebank symbol is correct, bonus points can be awarded for getting the right refinement.
On the training set, the "best" reachable tree will not always (~5% of the time) be the correct tree, because pruning will remove the right answer. We don't want to try to train towards an unreachable tree, because the training algorithm will do bad things. Instead, we want the best possible tree that our parser could conceivably produce. That is why this class exists.
If backupGrammar is provided, it will be used to find such a tree in the case that no tree can be found with grammar (given the current constraints).
Typically, the first grammar will be a treebank grammar that has no horizontal markovization (i.e. it is not forgetfully binarized) and it also remembers the functional tags like -TMP. The backup grammar is usually the grammar with which the pruning masks were produced; because of the way we prune, that parser will always be able to find a tree (assuming that it was able to find a tree without pruning.)
TODO: should be a cascade of grammars