Grammatical Competence and Parsing Performance

Much psycholinguistic experimentation in recent years has been concerned with the nature of the human processes that give rise to the difficulties associated with sentences such as The author read the review smiled and Whilst he was writing the review flowed onto the page. The problem, of course, is that language is ambiguous, and these examples illustrate that in the face of local ambiguity, the human processor chooses to pursue just one analysis, even if that analysis subsequently turns out to be false. The fact that there exist such consistent preferences (consistent across examples and across different people) suggests that the decision about which analysis to pursue is not arbitrary. Studying such examples may thus tell us something about the nature of human sentence processing, and even about the relationship between grammar and process, a topic of interest to psycholinguists, linguists, and computational linguists. Bradley Pritchett's book takes the reader through a wide range of examples of socalled "garden path" sentences, and reminds us that there are two questions that any theory of sentence processing must address: What is the cause of the various parsing preferences? And what determines ILhe ease of re-analysis? The second question is posed in the face of materials such as those above that require conscious effort, and others that require no conscious re-analysis but do nonetheless require some syntactic restructuring (compare He knew the book well and He knew the book was interesting, which differ in terms of the phrasal position to which the book is assigned). In the brief summary that follows, I shall gloss some of the main claims purely for the sake of brevity. Two principles are proposed, one to explain why certain analyses are chosen over others, and another to explain why certain forms of syntactic reanalysis are problematic. Theta attachment proposes, in effect, that at any moment the processor attempts to map as many thematic roles onto as many arguments as possible. Pritchett is clear, however, that the semantic content of the roles is largely irrelevant: what matters is the number of roles and their configurational position. He then proposes the theta reanalysis constraint, which states, again in effect, that the processor cannot reassign a role to an argument if the new role comes from outside the theta domain of the original role. Over the course of a very rigorous examination of how these principles account for a large variety of parsing preferences and associated reanalysis phenomena, the two principles evolve into generalized theta attachment ("every principle of the syntax attempts to be maximally satisfied at every point during processing") and the on-line locality constraint ("the target positiort .. . assumed by a constituent must be governed or dominated by its source position . . . . otherwise attachment is impossible"). Both principles are cast in the framework of Government-Binding theory.