Efficient implementation of beam-search incremental parsers

Beam search incremental parsers are accurate, but not as fast as they could be. We demonstrate that, contrary to popular belief, most current implementations of beam parsers in fact run in O(n²), rather than linear time, because each statetransition is actually implemented as an O(n) operation. We present an improved implementation, based on Tree Structured Stack (TSS), in which a transition is performed in O(1), resulting in a real lineartime algorithm, which is verified empirically. We further improve parsing speed by sharing feature-extraction and dotproduct across beam items. Practically, our methods combined offer a speedup of ∼2× over strong baselines on Penn Tree-bank sentences, and are orders of magnitude faster on much longer sentences.