Evaluation.

Evaluation of coverage and accuracy of a computational grammar usually is based on some metric which compares tree structures (such as recall and precision of (labelled) brackets or bracketing inconsistencies (crossing brackets) between test item and parser output). As is well-known, such metrics have a number of drawbacks. Therefore, [7] propose to annotate sentences with triples of the form $\langle$ head-word, dependency relation, dependent head-word$\rangle$. For instance, for the example in (1) we might obtain:

$\langle$zou, su, mercedes$\rangle$	$\langle$aangekondigd, obj1, model$\rangle$
$\langle$hebben, su, mercedes$\rangle$	$\langle$model, det, haar$\rangle$
$\langle$aangekondigd, su, mercedes$\rangle$	$\langle$model, mod, nieuwe$\rangle$
$\langle$aangekondigd, mod, gisteren$\rangle$

Dependency relations between head-words can be extracted easily from the dependency structures in our treebank, as well as from the dependency structures constructed by the parser. It is thus straightforward to compute precision, recall, and f-score on the set of dependency triples.