A Computational Model of Prediction in Human Parsing: Unifying Locality and Surprisal Effects

There is strong evidence that human sentence processing is incremental, i.e., that structures are built word by word. Recent experiments show that the processor also predicts upcoming linguistic material on the basis of previous input. We present a computational model of human parsing that is based on a variant of tree-adjoining grammar and includes an explicit mechanism for generating and verifying predictions, while respecting incrementality and connectedness. An algorithm for deriving a lexicon from a treebank, a fully implemented parser, and a probability model for this formalism are also presented. We devise a linking function that explains processing difficulty as a combination of prefix probability (surprisal) and verification cost. The resulting model captures locality effects such as the subject/object relative clause asymmetry, as well as surprisal effects such as prediction in "either ... or" constructions.