Correcting the Incorrect: Local Coherence Effects Modeled with Prior Belief Update

0. Introduction In the past four decades, the field of sentence processing research has generated a number of models of the incremental operation of the human sentence processor. One assumption common to most of these theories is that the difficulty of a word is determined by the possible syntactic structures of the preceding words, and thus, the difficulty of a word should be unaffected by ungrammatical analyses of the preceding words. Put formally, the difficulty of word wn of a sentence is determined only by the grammatical syntactic structures covering the preceding words w1 · · ·wn−1(≡ wn−1 1 ). Recent results by Tabor, Galantucci, and Richardson (2004), however, appear to show evidence of a case where a syntactic structure which is not possible given wn−1 1 nevertheless influences the difficulty of wn. They attribute such effects to ‘merely locally coherent’ syntactic structures and term such impossible structures local coherences. Follow-up studies by Konieczny (2005) and Konieczny and Müller (2006, 2007) provide further evidence that these impossible structures are being constructed and even semantically evaluated. These results have been taken to support a small class of dynamical systems models of sentence processing (e.g., Tabor and Hutchins 2004), in which, crucially, structures which are not possible given the current input are nevertheless constructed and compete with other, tenable structures. Unfortunately, the existing theories in this class have a large number of interacting free parameters, making interpretation somewhat difficult and leaving unspecified how to scale up such a system to make broad-coverage reading time predictions. This paper fills two gaps in the literature on local coherences. First, it demonstrates from two experiments with an eye-tracking corpus that effects of local coherences are evident in the reading of naturalistic text. Second, it describes a new computational model of local coherences that is motivated by a view of sentence processing as updating prior beliefs over syntactic structures.