Exponential Decay Pruning for Bottom-Up Beam-Search Parsing

We describe and motivate bottom-up beamsearch parsing, a probabilistic constituent parsing architecture that combines the advantages of best-first agenda parsing and bottom-up chart parsing into a unified framework. We also present Exponential Decay Pruning (EDP), a novel beam-width pruning technique that is simple and effective, increasing parsing speed up to 43% with no loss in accuracy. Using these techniques together, our system requires no knowledge of the underlying grammar, has a small memory footprint, and parses at stateof-the-art speed and accuracy levels. We demonstrate the flexibility of our approach by parsing with five different grammars and validate its effectiveness by comparing it to both the best-first and coarse-to-fine parsing algo-