Prediction

During language comprehension, the processing of each incoming word is facilitated in proportion to its predictability. Here, we asked whether anticipated upcoming linguistic information is actually pre-activated before new bottom-up input becomes available, and if so, whether this pre-activation is limited to the level of semantic features, or whether extends to representations of individual word-forms (orthography/phonology). We carried out Representational Similarity Analysis on EEG data while participants read highly constraining sentences. Prior to the onset of the expected target words, sentence pairs predicting semantically-related words (financial “bank” – “loan”) and form-related words (financial “bank” – river “bank”) produced more similar neural patterns than pairs predicting unrelated words (“bank” – “lesson”). This provides direct neural evidence for item-specific semantic and form predictive pre-activation. Moreover, the semantic pre-activation effect preceded the form pre-activation effect, suggesting that top-down pre-activation is propagated from higher to lower levels of the linguistic hierarchy over time.

To comprehend language, we continually use prior context to pre-activate expected upcoming information, resulting in facilitated processing of incoming words that confirm these predictions. But what are the consequences of disconfirming prior predictions? To address this question, most previous studies have examined unpredictable words appearing in contexts that constrain strongly for a single continuation. However, during natural language processing, it is far more common to encounter contexts that constrain for multiple potential continuations, each with some probability. Here, we ask whether and how pre-activating both higher and lower probability alternatives influences the processing of the lower probability incoming word. One possibility is that, similar to language production, there is continuous pressure to select the higher-probability pre-activated alternative through competitive inhibition. During comprehension, this would result in relative costs in processing the lower probability target. A second possibility is that if the two pre-activated alternatives share semantic features, they mutually enhance each other’s pre-activation. This would result in greater facilitation in processing the lower probability target. To distinguish between these accounts, we recorded ERPs as participants read three-sentence scenarios that constrained either for a single word or for two potential continuations – a higher probability expected candidate and a lower probability second-best candidate. We found no evidence that competitive pre-activation between the expected and second-best candidates resulted in costs in processing the second-best target, either during lexico-semantic processing (indexed by the N400) or at later stages of processing (indexed by a later frontal positivity). Instead, we found only benefits of pre-activating multiple alternatives, with evidence of enhanced graded facilitation on lower-probability targets that were semantically related to a higher-probability pre-activated alternative. These findings are consistent with a previous eye-tracking study by Luke and Christianson (2016, Cogn Psychol) using corpus-based materials. They have significant theoretical implications for models of predictive language processing, indicating that routine graded prediction in language comprehension does not operate through the same competitive mechanisms that are engaged in language production. Instead, our results align more closely with hierarchical probabilistic accounts of language comprehension, such as predictive coding.

We used magnetoencephalography (MEG) and event-related potentials (ERPs) to track the time-course and localization of evoked activity produced by expected, unexpected plausible, and implausible words during incremental language comprehension. We suggest that the full pattern of results can be explained within a hierarchical predictive coding framework in which increased evoked activity reflects the activation of residual information that was not already represented at a given level of the fronto-temporal hierarchy (“error” activity). Between 300 and 500 ms, the three conditions produced progressively larger responses within left temporal cortex (lexico-semantic prediction error), whereas implausible inputs produced a selectively enhanced response within inferior frontal cortex (prediction error at the level of the event model). Between 600 and 1,000 ms, unexpected plausible words activated left inferior frontal and middle temporal cortices (feedback activity that produced top-down error), whereas highly implausible inputs activated left inferior frontal cortex, posterior fusiform (unsuppressed orthographic prediction error/reprocessing), and medial temporal cortex (possibly supporting new learning). Therefore, predictive coding may provide a unifying theory that links language comprehension to other domains of cognition.

The N400 event-related brain potential is elicited by each word in a sentence and offers an important window into the mechanisms of real-time language comprehension. Since the 1980s, studies investigating the N400 have expanded our understanding of how bottom-up linguistic inputs interact with top-down contextual constraints. More recently, a growing body of computational modeling research has aimed to formalize theoretical accounts of the N400 to better understand the neural and functional basis of this component. Here, we provide a comprehensive review of this literature. We discuss “word-level” models that focus on the N400’s sensitivity to lexical factors and simple priming manipulations, as well as more recent sentence-level models that explain its sensitivity to broader context. We discuss each model’s insights and limitations in relation to a set of cognitive and biological constraints that have informed our understanding of language comprehension and the N400 over the past few decades. We then review a novel computational model of the N400 that is based on the principles of predictive coding, which can accurately simulate both word-level and sentence-level phenomena. In this predictive coding account, the N400 is conceptualized as the magnitude of lexico-semantic prediction error produced by incoming words during the process of inferring their meaning. Finally, we highlight important directions for future research, including a discussion of how these computational models can be expanded to explain language-related ERP effects outside the N400 time window, and variation in N400 modulation across different populations.

To make sense of the world around us, we must be able to segment a continual stream of sensory inputs into discrete events. In this review, I propose that in order to comprehend events, we engage hierarchical generative models that “reverse engineer” the intentions of other agents as they produce sequential action in real time. By generating probabilistic predictions for upcoming events, generative models ensure that we are able to keep up with the rapid pace at which perceptual inputs unfold. By tracking our certainty about other agents’ goals and the magnitude of prediction errors at multiple temporal scales, generative models enable us to detect event boundaries by inferring when a goal has changed. Moreover, by adapting flexibly to the broader dynamics of the environment and our own comprehension goals, generative models allow us to optimally allocate limited resources. Finally, I argue that we use generative models not only to comprehend events but also to produce events (carry out goal-relevant sequential action) and to continually learn about new events from our surroundings. Taken together, this hierarchical generative framework provides new insights into how the human brain processes events so effortlessly while highlighting the fundamental links between event comprehension, production, and learning.

During language comprehension, online neural processing is strongly influenced by the constraints of the prior context. While the N400 ERP response (300-500ms) is known to be sensitive to a word’s semantic predictability, less is known about a set of late positive-going ERP responses (600-1000ms) that can be elicited when an incoming word violates strong predictions about upcoming content (late frontal positivity) or about what is possible given the prior context (late posterior positivity/P600). Across three experiments, we systematically manipulated the length of the prior context and the source of lexical constraint to determine their influence on comprehenders’ online neural responses to these two types of prediction violations. In Experiment 1, within minimal contexts, both lexical prediction violations and semantically anomalous words produced a larger N400 than expected continuations (James unlocked the door/laptop/gardener), but no late positive effects were observed. Critically, the late posterior positivity/P600 to semantic anomalies appeared when these same sentences were embedded within longer discourse contexts (Experiment 2a), and the late frontal positivity appeared to lexical prediction violations when the preceding context was rich and globally constraining (Experiment 2b). We interpret these findings within a hierarchical generative framework of language comprehension. This framework highlights the role of comprehension goals and broader linguistic context, and how these factors influence both top-down prediction and the decision to update or reanalyze the prior context when these predictions are violated.

It has been proposed that hierarchical prediction is a fundamental computational principle underlying neurocognitive processing. Here we ask whether the brain engages distinct neurocognitive mechanisms in response to inputs that fulfill versus violate strong predictions at different levels of representation during language comprehension. Participants read three-sentence scenarios in which the third sentence constrained for a broad event structure, e.g. Agent caution animate-Patient. High constraint contexts additionally constrained for a specific event/lexical item, e.g. a two-sentence context about a beach, lifeguards and sharks constrained for the event, Lifeguards cautioned Swimmers and the specific lexical item, “swimmers”. Low constraint contexts did not constrain for any specific event/lexical item. We measured ERPs on critical nouns that fulfilled and/or violated each of these constraints. We found clear, dissociable effects to fulfilled semantic predictions (a reduced N400), to event/lexical prediction violations (an increased late frontal positivity), and to event structure/animacy prediction violations (an increased late posterior positivity/P600). We argue that the late frontal positivity reflects a large change in activity associated with successfully updating the comprehender’s current situation model with new unpredicted information. We suggest that the late posterior positivity/P600 is triggered when the comprehender detects a conflict between the input and her model of the communicator and communicative environment. This leads to an initial failure to incorporate the unpredicted input into the situation model, which may be followed by second-pass attempts to make sense of the discourse through reanalysis, repair, or reinterpretation. Together, these findings provide strong evidence that confirmed and violated predictions at different levels of representation manifest as distinct spatiotemporal neural signatures.

It has been proposed that people can generate probabilistic predictions at multiple levels of representation during language comprehension. We used magnetoencephalography (MEG) and electroencephalography (EEG), in combination with representational similarity analysis, to seek neural evidence for the prediction of animacy features. In two studies, MEG and EEG activity was measured as human participants (both sexes) read three-sentence scenarios. Verbs in the final sentences constrained for either animate or inanimate semantic features of upcoming nouns, and the broader discourse context constrained for either a specific noun or for multiple nouns belonging to the same animacy category. We quantified the similarity between spatial patterns of brain activity following the verbs until just before the presentation of the nouns. The MEG and EEG datasets revealed converging evidence that the similarity between spatial patterns of neural activity following animate-constraining verbs was greater than following inanimate-constraining verbs. This effect could not be explained by lexical-semantic processing of the verbs themselves. We therefore suggest that it reflected the inherent difference in the semantic similarity structure of the predicted animate and inanimate nouns. Moreover, the effect was present regardless of whether a specific word could be predicted, providing strong evidence for the prediction of coarse-grained semantic features that goes beyond the prediction of individual words.

When semantic information is activated by a context prior to new bottom-up input (i.e. when a word is predicted), semantic processing of that incoming word is typically facilitated, attenuating the amplitude of the N400 event related potential (ERP) – a direct neural measure of semantic processing. N400 modulation is observed even when the context is a single semantically related “prime” word. This so-called “N400 semantic priming effect” is sensitive to the probability of encountering a related prime-target pair within an experimental block, suggesting that participants may be adapting the strength of their predictions to the predictive validity of their broader experimental environment. We formalize this adaptation using a Bayesian learning model that estimates and updates the probability of encountering a related versus an unrelated prime-target pair on each successive trial. We found that our model’s trial-by-trial estimates of target word probability accounted for significant variance in the amplitude of the N400 evoked by target words. These findings suggest that Bayesian principles contribute to how comprehenders adapt their semantic predictions to the statistical structure of their broader environment, with implications for the functional significance of the N400 component and the predictive nature of language processing.