Corpus-based Learning in Stochastic OT-LFG: Experiments with a Bidirectional Bootstrapping Approach

This paper reports on experiments exploring the application of a Stochastic Optimality-Theoretic approach in the corpus-based learning of some aspects of syntax. Using the Gradual Learning Algorithm, the clausal syntax of German has to be learned from learning instances of clauses extracted from a corpus. The particular focus in the experiments was placed on the usability of a bidirectional approach, where parsing-directed, interpretive optimization is applied to determine the target candidate for a subsequent application of generation-directed, expressive optimization. The results show that a bidirectional bootstrapping approach is only slightly less effective than a fully supervised approach.