Compiling an HPSG-based grammar into more than one CFG

Recently, the performance of HPSG parsing has been improved so that the parsers can be applied to real-world texts. CFG filtering is one of the techniques which contributed to this progress. It improved parsing speed by filtering impossible parse trees by using the CFG compiled from a given HPSGbased grammar. However, there is a limit in the speed-up. This is because the compiled CFG grows into an impractical size when the original grammar is complex. This caused serious memory problems in actual parsing. This paper describes a technique to avoid this obstacle. We compile a given HPSG-based grammar to more than one CFG, and utilize all the CFGs in a filtering process. We show that, by this technique, we can achieve almost the same parsing speed with far less memory through a series of experiments.