The State of Cognitive Control in Language Processing.

Understanding language requires readers and listeners to cull meaning from fast-unfolding messages that often contain conflicting cues pointing to incompatible ways of interpreting the input (e.g., "The cat was chased by the mouse"). This article reviews mounting evidence from multiple methods demonstrating that cognitive control plays an essential role in resolving conflict during language comprehension. How does cognitive control accomplish this task? Psycholinguistic proposals have conspicuously failed to address this question. We introduce an account in which cognitive control aids language processing when cues conflict by sending top-down biasing signals that strengthen the interpretation supported by the most reliable evidence available. We also provide a computationally plausible model that solves the critical problem of how cognitive control "knows" which way to direct its biasing signal by allowing linguistic knowledge itself to issue crucial guidance. Such a mental architecture can explain a range of experimental findings, including how moment-to-moment shifts in cognitive-control state-its level of activity within a person-directly impact how quickly and successfully language comprehension is achieved.

Rational Adaptation in Lexical Prediction: The Influence of Prediction Strength.

Recent studies indicate that the processing of an unexpected word is costly when the initial, disconfirmed prediction was strong. This penalty was suggested to stem from commitment to the strongly predicted word, requiring its inhibition when disconfirmed. Additional studies show that comprehenders rationally adapt their predictions in different situations. In the current study, we hypothesized that since the disconfirmation of strong predictions incurs costs, it would also trigger adaptation mechanisms influencing the processing of subsequent (potentially) strong predictions. In two experiments (in Hebrew and English), participants made speeded congruency judgments on two-word phrases in which the first word was either highly constraining (e.g., "climate," which strongly predicts "change") or not (e.g., "vegetable," which does not have any highly probable completion). We manipulated the proportion of disconfirmed predictions in highly constraining contexts between participants. The results provide additional evidence of the costs associated with the disconfirmation of strong predictions. Moreover, they show a reduction in these costs when participants experience a high proportion of disconfirmed strong predictions throughout the experiment, indicating that participants adjust the strength of their predictions when strong prediction is discouraged. We formulate a Bayesian adaptation model whereby prediction failure cost is weighted by the participant's belief (updated on each trial) about the likelihood of encountering the expected word, and show that it accounts for the trial-by-trial data.

From pre-activation to pre-updating: A threshold mechanism for commitment to strong predictions.

Ample evidence suggests that during sentence processing comprehenders can "pre-activate" lexical/semantic knowledge stored in long-term memory. A relatively recent development suggests that in some cases a stronger form of prediction is employed, involving "pre-updating" the predicted content into the sentence's representation being built in working memory. The current study argues for an activation threshold mechanism by which pre-updating is initiated, within the routine processing stages of a word in a context. By combining a speeded cloze task with event-related potentials, we were able to analyze electrophysiological data measured prior to when participants were prompted to produce a completion, based on the participant's cloze response, reflecting their strongest prediction at that specific moment in time. A P600 effect reflecting pre-updating was observed in high (relative to low) constraint sentences, even in trials where the participant predicted a low cloze word. The results support a mechanism in which multiple predictions accumulate activations, "racing" toward a retrieval threshold. Once the activation level of a certain word passes the threshold, the word is integrated into the sentence representation in working memory. Pre-updating occurs if a certain prediction passes the retrieval threshold prior to its realization in the input.

Love thy neighbor: Facilitation and inhibition in the competition between parallel predictions.

Ample evidence suggests that during word recognition and production, simultaneously activated lexical and sublexical representations interact, demonstrating varied patterns of facilitation and inhibition in various tasks and measures. A separate line of research has led to a growing consensus that prediction during sentence processing involves activating multiple possible predictions. However, very little is known about the nature of the interactions between parallel predictions. The current study employed a speeded cloze task to probe competition between simultaneously activated predictions. We focused on the modal response (the most probable completion for a sentence) and its strongest competitor (the second most probable completion). Examining production latencies of the modal response, the results showed an interaction between competitor strength and the semantic relatedness between the competitor and the modal: when the two were related, the stronger the competitor was, the more it facilitated production; however, when the two were unrelated, the stronger the competitor was, the more inhibition it caused. These results contrast with the pattern observed for the influences of near and distant semantic neighbors on word recognition and production. However, we show that when the different nature of the tasks is taken into consideration, these patterns of interaction between parallel predictions can be accounted for by the interactive activation and competition (IAC) model used to account for previous neighborhood effects (Chen & Mirman, 2012).

Predictive Pre-updating and Working Memory Capacity: Evidence from Event-related Potentials.

It was recently proposed that lexical prediction in sentence context encompasses two qualitatively distinct prediction mechanisms: "pre-activation," namely, activating representations stored in long-term memory, and "pre-updating," namely, updating the sentence's representation, built online in working memory (WM), to include the predicted content [Lau, E. F., Holcomb, P. J., & Kuperberg, G. R. Dissociating N400 effects of prediction from association in single-word contexts. Journal of Cognitive Neuroscience, 25, 484-502, 2013]. The current study sought to find evidence for pre-updating and test the influence of individual differences in WM capacity on the tendency to engage in this process. Participants read strongly and weakly constraining sentences. ERPs were measured on the predictable noun as well as on the preceding verb, where the prediction is generated. Increased P600 amplitude was observed at the verb in the strongly constraining sentences, reflecting integration of the predicted upcoming argument, thus providing evidence for pre-updating. This effect was greater for participants with higher WM capacity, indicating that the tendency to engage in pre-updating is highly affected by WM capacity. The opposite effect was observed at the noun, that is, for participants with higher WM span, a greater decrease in P600 amplitude in the strongly constraining sentences was observed, indicating that the integration of a pre-updated word was easier. We discuss these results in light of previous literature and propose a plausible architecture to account for the interplay between pre-activation and pre-updating, mediating the influence of factors such as WM capacity.

Lexical inhibition due to failed prediction: Behavioral evidence and ERP correlates.

During sentence processing, comprehenders form expectations regarding upcoming material, and may even predict a specific word. Previous event-related potential (ERP) studies have shown that disconfirmed predictions elicit a post-N400-positivity (PNP) with two distinct distributions. A frontal-PNP (f-PNP) is elicited when an unexpected but congruent word appears instead of a highly predictable word, whereas an anomalous word elicits a posterior-PNP. The current study tested the hypothesis that during the processing of a sentence, the predicted word is inhibited to enable the integration of unexpected but congruent material, and that this inhibitory process is reflected in the f-PNP component. In contrast, anomalous continuations, which are not compatible with the preceding context, do not induce inhibition. Experiment 1 used cross-modal lexical priming to test inhibition patterns of predicted words, demonstrating inhibition when integration of a congruent-unexpected word was needed, but not when an anomaly was encountered. Experiment 2 showed that the inhibition observed in Experiment 1 is specific to the predicted word and does not stem from competition between two congruent continuations. In Experiment 3 we recorded ERPs using the same materials, and found that the f-PNP component is elicited under the same conditions giving rise to behavioral inhibition, and that the two are correlated, thus providing preliminary support for the hypothesis that this component reflects an inhibitory process. (PsycINFO Database Record

Working Memory in the Processing of Long-Distance Dependencies: Interference and Filler Maintenance.

During the temporal delay between the filler and gap sites in long-distance dependencies, the "active filler" strategy can be implemented in two ways: the filler phrase can be actively maintained in working memory ("maintenance account"), or it can be retrieved only when the parser posits a gap ("retrieval account"). The current study tested whether filler content is maintained during the processing of dependencies. Using a self-paced reading paradigm, we compared reading times on a noun phrase (NP) between the filler and gap sites in object relative clauses, to reading times on an NP between the antecedent and ellipsis sites in ellipsis sentences. While in the former type of dependency a filler by hypothesis can be maintained, in the latter there is no indication for the existence of a dependency prior to the ellipsis site, and hence no maintenance. By varying the amount of similarity-based interference between the antecedent and integration sites, we tested the influence of holding an unresolved dependency on reading times. Significantly increased reading times due to interference were found only in the object relative condition, and not in the ellipsis condition, demonstrating filler maintenance costs. The fact that these costs were measured as an effect on similarity-based interference indicates that the maintained representation of the filler must include at least some of the features shared by the interfering NP.