Limits on semantic prediction in the processing of extraction from adjunct clauses

Processing filler-gap dependencies (‘extraction’) depends on complex top-down predictions. This is observed in comprehenders’ ability to avoid resolving filler-gap dependencies in syntactic island contexts, and in the immediate sensitivity to the plausibility of the resulting interpretation. How complex can these predictions be? In this paper, we examine the processing of extraction from adjunct clauses. Adjunct clauses are argued to be syntactic islands, however, extraction is permitted if the adjunct clause and main clause satisfy specific compositional and conceptual semantic criteria. In an acceptability judgment task, we found that this generalization is robust. Additionally, our results show that this is a property specific to adjunct clauses by comparing adjunct clauses to conjunct VPs, which are similarly argued to permit extraction depending on semantic factors. However, in an A-Maze task, we found no evidence that this knowledge is deployed in real-time sentence processing. Instead, we found that comprehenders attempted to resolve a filler-gap dependency in an adjunct clause regardless of its island status. We propose that this is because deploying this linguistic constraint depends on a second-order serial search over event schemata, which is likely costly and time-consuming. Thus, comprehenders opt for a riskier strategy and attempt resolution into adjunct clauses categorically.