Building and Exploiting Constraint-Based Treebanks

We present in this paper a technique aiming at enriching existing treebanks with constraint-based information. The technique consists in acquiring a constraint grammar from the original constituency treebank and applying it with a constraint parser to the source treebank. The result is an hybrid treebank, with both constituency and constraint-based representations. After a presentation of the general technique in the framework of \emphProperty Grammars, we propose an application for the hybridization of the \emphFrench Treebank. This kind of resource offers several interests, including the possibility of identifying the relative importance of constraints. We present a method assigning weights to individual constraints as well as constraint types. An experiment by means of human acceptability judgments gives a first validation of the approach