Changing Expectations: Syntactic Dependency Formation Weakens Surprisal Effects

Our capacity to understand language relies on building rapid structural representations of sentences. This process can be less or more taxing, depending on the type of word or phrase we encounter at any given moment. A debated question is whether certain momentary difficulties that arise during sentence comprehension are caused by the expectations we have already formed about upcoming material or by memory-based mechanisms used to construct the requisite structural representation. Here, we examine relative clauses (RCs) in Mandarin, a language with prenominal RCs. We present a self-paced reading experiment which uses nominal classifiers to provide a syntactic cue during parsing, allowing for disambiguating a sentence-initial RC from a matrix clause. To evaluate predictions of memory-based accounts against expectation-based ones, we compared models predicting reading times from surprisal estimates with models using categorical structural predictors. Results show that surprisal better predicts reading times before the dependency is encountered (i.e., during the RC), whereas the categorical model better predicts reading times from the dependency region onward (i.e., after active dependency formation). Here, we find an object RC advantage which is immediately surpassed by a sustained subject RC advantage. Overall, our findings indicate that expectation and memory-based dependency processing contribute to a lesser or greater extent at different temporal stages in a sentence, depending on their compositional properties.