在语言理解过程中，预期起着重要作用。已有研究主要集中考察预期加工的整合阶段，对预期的形成过程尚无清晰的认识。本研究借助脑电技术探索了语言理解过程中预期形成的认知机制及其与工作记忆的关系。我们一方面，操纵句子的语境限制性（高语义预期、中语义预期、低语义预期），并把预期加工的“形成阶段”和“整合阶段”分离开。另一方面，我们把被试分为高言语工作记忆广度组和低言语工作记忆广度组；并对被试的工作记忆的三种“中央执行功能”（更新、抑制、转换）进行测验。
结果发现，1）对于高言语工作记忆广度者：在预期的“整合阶段”，“高语义预期”和“中语义预期”条件都比“低语义预期”条件诱发了更小的N400；同时“高语义预期”比“中语义预期”条件诱发了更小的N400。在预期的“形成阶段”，“高语义预期”和“中语义预期”条件都比“低语义预期”条件诱发了更小的前部负波；同时“高语义预期”比“中语义预期”条件诱发了更小的前部负波。即，无论是在预期的“形成阶段”还是“整合阶段”，较强的语境限制性都会促进语义加工，而且此加工效益随着语境限制性的梯度变化而呈现出梯度变化。2）对于低工作记忆广度者，在“整合阶段”，“高语义预期”比“低语义预期”诱发了更小的N400，而“中语义预期”和“低语义预期”诱发的N400没有显著差异。在预期的“形成阶段”，则出现相反的模式，即，“高语义预期”比“低语义预期”条件诱发了更大（而不是更小）的前部负波；同时“中语义预期”条件也比“低语义预期”条件诱发了更大的前部负波；而“高语义预期”和“中语义预期”无显著性差异。即，较强的语境限制性虽然会促进“整合阶段”的语义加工、但却会给预期“形成阶段”的加工带来更大的认知负荷；中等限制性语境被试图用来形成预期、但最终没有起到促进语义加工的作用；同时语境限制性的加工效益/负荷都不会表现出梯度变化。3）进一步的相关分析发现，预期“形成阶段”的加工难易程度与“更新”能力呈负相关，更新的反应时越长、在预期阶段的语境限制性的加工获益越小。以上结果说明，语言理解过程中，预期形成过程与言语工作记忆广度和更新能力密切相关；高、低言语工作记忆广度者的预期形成过程是基于不同的认知机制，低工作记忆者需要更多的自上而下的控制过程参与；只有高言语工作记忆广度者的预期机制符合贝叶斯模型的假设。

Predictive processing plays an important role during language comprehension. Although numerous studies have examined the integration stage of predictive processing (namely, integration between top-down predictions and bottom-up new inputs), there is still no clear picture of the anticipatory stage of predictive processing (namely, before the predicted/unpredicted information appears in the bottom-up signal). This electroencephalograph study examined the cognitive mechanisms underlying anticipatory processing during language comprehension, and the relationship between anticipatory language processing and working memory. Participants read Mandarin Chinese sentences that were either strongly, moderately, or weakly constraining, and we examined the effects of semantic predictability on anticipatory processing prior to the onset of the critical nouns and on integration of the critical nouns. Meanwhile, two groups of participants were selected (namely, participants with high working memory capacity and those with low working memory capacity) and their executive-control abilities (inhibiting, updating, and shifting) were measured.
The result showed that, first, for the high-working-memory-capacity group, at the integration stage, both strong-constraint and moderate-constraint conditions elicited a smaller N400 than weak-constraint condition, and strong-constraint condition further evoked a smaller N400 than moderate-constraint condition. At the anticipation stage, both strong-constraint and moderate-constraint condition elicited a  reduced anterior negativity compared to weak-constraint condition, and strong-constraint condition further elicited a reduced anterior negativity compared to moderate-constraint condition. That is, the magnitude of N400 or anterior negativity decreases gradually from strong-constraint to moderate-constraint and from moderate-constraint to weak-constraint conditions. Second, for the low-working-memory-capacity group, at the integration stage, strong-constraint condition elicited a smaller N400 than weak-constraint condition, whereas there was no significant difference between moderate- and weak-constraint conditions. At the anticipatory stage, a reversed patter of results was observed: strong-constraint condition evoked an enhanced (rather than a reduced) anterior negativity compared to weak-constraint conditions; moderate-constraint condition also elicited an enhanced anterior negativity compared to weak-constraint condition; no significant difference was observed between strong- and moderate-constraint conditions. That is, relative to weakly constraining context, strongly constraining context facilitated semantic processing at the integration, but put a higher cognitive cost at the anticipatory stage of processing; the moderately constraining context might have been used to form predictions, but didn’t lead to processing benefit at a later stage; the ease/difficulty of processing didn’t show gradual changes from strong-constraint to moderate-constraint and then to weak-constraint conditions. Third, there is a significant negative correlation between the ease of anticipatory processing and updating ability: the slower the process of updating, the more difficulty the anticipatory language processing is.
There results indicated that, during language comprehension, anticipatory language processing is closely correlated with working memory capacity and the speed of updating; the high-working-memory-capacity group and the low-working-memory-capacity group recruit different cognitive mechanisms to perform anticipatory language processing, with the later relying more on top-down control processes; only the high-working-memory-capacity group’s anticipatory
language processing works as the Bayesian model hypothesizes.