Filling Statistics with Linguistics -- Property Design for the Disambiguation of German LFG Parses

We present a log-linear model for the disambiguation of the analyses produced by a German broad-coverage LFG, focussing on the properties (or features) this model is based on. We compare this model to an initial model based only on a part of the properties provided to the final model and observe that the performance of a log-linear model for parse selection depends heavily on the types of properties that it is based on. In our case, the error reduction achieved with the log-linear model based on the extended set of properties is 51.0% and thus compares very favorably to the error reduction of 34.5% achieved with the initial model.