(1) We formalized the notion of prediction in human parsing by developing a probabilistic framework for prediction-based parsing, consisting of a generation component and a verification component.
(2) We designed and implement a broad-coverage computational model of prediction. This involved developing a grammatical formalism for representing predictions explicitly (using a variant of tree-adjoining grammar). Based on this, we then designed a parsing algorithm for this formalism, along with a lexicon induction scheme and a probabilistic model.
(3) We used the Dundee eye-tracking corpus to test the broad coverage aspects of the model, i.e., its performance on naturally occurring, unrestricted text. We also conducted a total of five eye-tracking experiments to test fine-grained predictions of the model.