Dependency structures represent the grammatical relations that hold in and between constituents. On the one hand they are more abstract than syntactic trees (word order for example is not expressed) and on the other hand they are more explicit about the dependency relations. Indices denote that constituents may have multiple (possibly different) dependency relations with different words. Fig. 2 shows the dependency tree for the sentence Kim wil weten of Anne komt. The dependency relations are the top labels in the boxes. In addition, the syntactic category, lexical entry and string position are added to each leaf. The index 1 indicates that Kim is the subject of both wil (wants) and weten (to know).
The main advantage of this format is that it is relatively theory independent, which is important in a grammar engineering context. A second advantage is that the format is similar to the format CGN uses (and that they in turn based on the Tiger Treebank), which allowed us to base our annotation guidelines on theirs (Moortgat, Schuurman and van der Wouden 2001). The third and last argument for using dependency structures is that it is relatively straightforward to perform evaluation of the parser on dependency structures: one can compare the automatically generated dependency structure with the one in the treebank and calculate statistical measures such as F-score based on the number of dependency relations that are identical in both trees [4].
