Strong evidence for expectation adaptation during language understanding, not a replication failure. A reply to Harrington Stack, James, and Watson (2018)

Recent input affects subsequent language processing. One explanation for this holds that comprehenders adapt their implicit linguistic expectations based on the input, so as to facilitate efficient processing. A number of studies have identified support for this hypothesis. Harrington Stack, James, and Watson (2018) report a failure to replicate one of these studies (Fine, Jaeger, Farmer, and Qian, 2013). We show that, to the contrary, the data from Harrington Stack and colleagues constitute strong support for the hypothesis of expectation adaptation. Several factors contribute to the difference in conclusions. For example, Harrington Stack and colleagues argue based on differences in p-values, which is known to be problematic. We instead employ well-formed Bayesian measures of evidentiary support to assess replication success. Most critically though, the new experiments by Harrington Stack and colleagues differ in design from the experiments they aim to replicate. We correct for these differences by means of the same single-parameter belief-updating model previously employed by Fine and colleagues. The model provides trial-level predictions for the surprisal that comprehenders experience, based on previous input within and outside of the experiment. Trial-level analyses find that surprisal based on adapted expectations strongly predicts reading times in both the original and the replication data. In fact, once the differences in design are corrected for, the two data are highly similar; replication tests estimate the posterior probability of a replication success to be ≫ .9999. We show how the same belief-updating model also predicts trial-to-trial priming, cumulative priming, and the inverse preference effect in priming.

近期输入对后续语言处理具有显著影响。一种解释观点认为，理解者在根据输入调整其隐含的语言预期，以此促进高效的加工处理。众多研究已证实此假设的支持。Harrington Stack、James 和 Watson（2018）报告称，未能复制其中一项研究（Fine, Jaeger, Farmer, 和 Qian，2013）。我们则展示，相反，Harrington Stack 及其同事的数据为预期适应性假设提供了强有力的支持。数个因素导致了结论间的差异。例如，Harrington Stack 及其同事基于p值差异进行论证，而p值差异已知存在缺陷。我们则采用结构良好的贝叶斯证据支持度量来评估复现的成功率。然而，最关键的是，Harrington Stack 及其同事的新实验在设计上与他们试图复现的实验存在差异。我们通过采用Fine 及其同事先前使用的同一参数信念更新模型来纠正这些差异。该模型基于实验内外先前输入，对理解者所经历的意外程度进行试验级预测。试验级分析发现，基于适应性预期的意外程度强烈预测了原始数据和复现数据中的阅读时间。事实上，一旦纠正了设计上的差异，两种数据的高度相似性便显现出来；复现测试估计复现成功的后验概率远大于0.9999。我们展示了相同的信念更新模型如何预测试验间的启动效应、累积启动效应以及启动效应的逆向偏好效应。