Sequential vs. Hierarchical Syntactic Models of Human Incremental Sentence Processing

Experimental evidence demonstrates that syntactic structure influences human online sentence processing behavior. Despite this evidence, open questions remain: which type of syntactic structure best explains observed behavior–hierarchical or sequential, and lexicalized or unlexicalized? Recently, Frank and Bod (2011) find that unlexicalized sequential models predict reading times better than unlexicalized hierarchical models, relative to a baseline prediction model that takes wordlevel factors into account. They conclude that the human parser is insensitive to hierarchical syntactic structure. We investigate these claims and find a picture more complicated than the one they present. First, we show that incorporating additional lexical n-gram probabilities estimated from several different corpora into the baseline model of Frank and Bod (2011) eliminates all differences in accuracy between those unlexicalized sequential and hierarchical models. Second, we show that lexicalizing the hierarchical models used in Frank and Bod (2011) significantly improves prediction accuracy relative to the unlexicalized versions. Third, we show that using stateof-the-art lexicalized hierarchical models further improves prediction accuracy. Our results demonstrate that the claim of Frank and Bod (2011) that sequential models predict reading times better than hierarchical models is premature, and also that lexicalization matters for prediction accuracy.