Stochastic Realisation Ranking for a Free Word Order Language

We present a log-linear model that is used for ranking the string realisations produced for given corpus f-structures by a reversible broadcoverage LFG for German and compare its results with the ones achieved by the application of a language model (LM). Like other authors that have developed log-linear models for realisation ranking, we use a hybrid model that uses linguistically motivated learning features and a LM (whose score is simply integrated into the log-linear model as an additional feature) for the task of realisation ranking. We carry out a large evaluation of the model, training on over 8,600 structures and testing on 323. We observe that the contribution that the structural features make to the quality of the output is slightly greater in the case of a free word order language like German than it is in the case of English. The exact match metric improves from 27% to 37% when going from the LM-based realisation ranking to the hybrid model, BLEU score improves from 0.7306 to 0.7939.