Abstract

Hierarchical Hidden Markov Model (HHMM) parsers have been proposed as psycholinguistic models due to their broad coverage within human-like working memory limits (Schuler et al., 2008) and ability to model human reading time behavior according to various complexity metrics (Wu et al., 2010). But HHMMs have been evaluated previously only with very wide beams of several thousand parallel hypotheses, weakening claims to the model’s efficiency and psychological relevance. This paper examines the effects of varying beam width on parsing accuracy and speed in this model, showing that parsing accuracy degrades gracefully as beam width decreases dramatically (to 2% of the width used to achieve previous top results), without sacrificing gains over a baseline CKY parser. 1

Citations