Shifting expectations: when knowledge-based predictions and linguistic context collide

To cope with the demands of language comprehension, young adults often actively engage in prediction of upcoming information—which may be more or less successful depending on each individual's specific knowledge. However, limited research has directly investigated the link between existing knowledge and real-time mechanisms of prediction. Here, we focus on a specific knowledge domain, the fictional world of Harry Potter (HP). Participants with varying degrees of HP knowledge read sentences about general topics and then about HP, each containing a predictable, unexpected-but-plausible, or implausible critical word, while we recorded event-related brain potentials. As expected, HP knowledge modulated N400 amplitudes (an ERP known to index availability of word meaning) to predictable words in HP sentences. HP knowledge also modulated late frontal positivities (LFPs; associated with shifting meaning interpretation upon encountering prediction violations) to unexpected-but-plausible words. The extent to which domain knowledge modulated both N400s and LFPs to unexpected-but-plausible continuations depended on how generally well-known the content in the sentence was. High-knowledge individuals showed reduced initial facilitation (i.e., larger N400 amplitudes) for unexpected-but-plausible words when the sentence contents were generally well-known (compared to less well-known), suggesting that they used their domain knowledge to “override” a more generic interpretation. They additionally showed a greater frontal positivity when sentence contents were less (compared to more) well known, suggesting a willingness to consider alternate interpretations when knowledge is weaker and/or more uncertain—but less so when knowledge is strong. We conclude that possessing relevant knowledge may shape predictive processes during language comprehension, suggesting people may shift their “mode” of language processing depending on existing knowledge and comprehension demands.

Navigating the world requires flexible, rapid responses to shifting environmental demands. Prominent theories in cognitive neuroscience suggest that the brain has adapted to these demands by forming forward, predictive models of the world, which are continuously compared with incoming input (reviewed in Clark, 2013). Similarly, many current theories of language comprehension propose that readers and listeners routinely construct and maintain high-level mental models of linguistic and non-linguistic context, enabling predictions about what linguistic input will come next, with comparisons between expected and received information occurring at multiple levels and grains of linguistic representations (Kuperberg, 2020; Ryskin and Nieuwland, 2023). “Prediction” is thus a widely used term within the field of psycholinguistics and has been described in many ways (see discussions in DeLong et al., 2014b; Kutas et al., 2014; Kuperberg and Jaeger, 2016; Ryskin and Nieuwland, 2023). One inclusive definition was put forth by Kutas et al. (2014): “…prediction involves the activation of information about likely upcoming stimuli, prior to their receipt, that plays a causal role in stimulus processing” (p. 649).

Some psycholinguistic theories seem to assume that, like the broader neurocognitive theories of predictive processing/predictive coding (e.g., Clark, 2013), engaging in prediction is a default, automatically engaged component of language comprehension (e.g., Michaelov et al., 2024). Indeed, there is a great deal of empirical evidence that comprehenders—especially young adult college students—routinely engage in prediction of this sort during real-time language comprehension (reviewed in DeLong et al., 2014b; Kutas et al., 2014; Kuperberg and Jaeger, 2016). However, mounting evidence also suggests that prediction is not an obligatory component of language comprehension (Federmeier et al., invited review)1. Engagement in prediction varies with task demands and the statistics of linguistic materials; for example, when linguistic materials rarely contain predictable continuations, participants are less likely to engage in prediction (Lau et al., 2013; Brothers et al., 2017). Prediction during language comprehension also varies with language experience and across the lifespan, with older adults and less literate young and middle aged adults generally predicting less than younger adults (Federmeier et al., 2002; Ng et al., 2017; reviewed in Payne and Silcox, 2019), and shows moment-to-moment fluctuations (Jongman et al., 2023).

In the present study, we investigate an underexplored source of variability that seems likely to affect the extent to which people predict during language comprehension. Specifically, we ask whether an individual's degree of (domain) knowledge specific to the topic at hand may shape the extent to which individuals actively engage in predictive processing. Domain knowledge could either promote or deter prediction. On the one hand, generating accurate predictive models would seem to require the availability of relevant knowledge. Thus, knowing more about the contents of any given utterance seems like a useful precursor to successful prediction. Indeed, as reviewed below, there is evidence that strong contextual constraint, and other factors influencing predictive validity, may encourage greater (i.e., more frequent engagement in) prediction. On the other hand, prediction is also theorized to underlie error-based learning mechanisms, both in children and adults (Reuter et al., 2019; Hodapp and Rabovsky, 2021; reviewed in Ryskin and Nieuwland, 2023). If prediction serves learning, its utility may be minimal for individuals with extensive knowledge who have less to learn. Furthermore, highly knowledgeable individuals may be more likely to rely on top-down processing (i.e., their “priors”), down-weighting novel bottom-up input that contradicts their existing knowledge (e.g., Kuperberg, 2020), especially if they view their interlocutor (or other linguistic source) as lacking expertise or credibility, or being otherwise unreliable (cf. Lau et al., 2013; Brothers et al., 2017, 2019).2 Thus, if prediction is deeply linked to learning mechanisms, it might be reduced when comprehenders recognize that they do not wish to update their knowledge based on new input.

Before presenting our study, we overview two event-related brain potential (ERP) components central to research on comprehension and prediction—the N400 and a later, post-N400 frontal positivity. We then briefly review recent research on the role of prediction in language comprehension as well as its potential relationship to an individual's content-specific (domain) knowledge.

Studying prediction using event-related brain potentials: N400s and late frontal positivities

ERPs measure neural activity time-locked to stimuli of interest, such as a word, image, or sound, providing insight into the mechanisms involved in real-time comprehension (reviewed in Kutas and Federmeier, 2000; Kutas et al., 2014). Here we review two ERP components, N400s and post-N400, late frontal positivities, which have been particularly useful in studies of the dynamics of prediction during language comprehension.

The N400 (Kutas and Hillyard, 1980) is a negative-going potential peaking approximately 400 ms after receipt of a potentially meaningful stimulus. Although theoretical interpretations differ (reviewed in Kutas and Federmeier, 2011; see Eddine et al., 2022, for a review of recent computational models), perhaps all agree that N400s reflect processes involved in making sense of word meaning. Our perspective (reviewed in Federmeier, 2022; see also Kutas and Federmeier, 2000) is that N400s reflect the moment when the current linguistic input makes contact with (lexico-)semantic memory, a process of “connecting” that is automatic and obligatory upon receipt of any potentially meaningful input.

N400 amplitudes, which are large to isolated, initial, or contextually unsupported words, become smaller when a word is preceded by a facilitative context (e.g., Van Petten and Kutas, 1990). A reduction in N400 amplitude is thus generally considered to reflect facilitated processing, likely because information relevant for making sense of a word's meaning has already been made available by the context. This is true when the context sets up an expectation for a particular word, but also when contextual factors facilitate a word's processing—even if it could not have been predicted. For example, when participants were asked to read sentences about everyday scenarios like hiking up a hill, N400 amplitudes are most reduced (compared to an anomalous word) for expected words, as in “ [The hikers] were awed by the beautiful view,” but are also reduced (albeit to a lesser degree) for linguistically inappropriate but contextually related words; for example “ [The hikers] were awed by the beautiful boots ” (Metusalem et al., 2012).

Because N400 amplitudes are highly sensitive to contextual support and predictability, as indexed by cloze probability (the proportion of participants who provide a given continuation to a sentence stem; Taylor, 1953; see DeLong et al., 2005) or word predictability estimates gleaned from large language models (Szewczyk and Federmeier, 2022), some computational and theoretical models of N400s posit that N400 amplitudes directly index prediction error—i.e., that they reflect some aspect(s) of an (ostensibly) obligatory process that tracks linguistic differences between prior expectations (i.e., predictions) and the current input (Rabovsky and McRae, 2014; Kuperberg and Jaeger, 2016; Bornkessel-Schlesewsky and Schlesewsky, 2019; Fitz and Chang, 2019; Eddine et al., 2022; Michaelov et al., 2024; reviewed in Ryskin and Nieuwland, 2023). On such a view, N400 amplitudes are larger when prediction error is higher, as when words are unexpected in context.

However, several findings challenge a direct prediction-error interpretation (see Federmeier, 2022, for a discussion). N400 amplitudes are insensitive to sentential constraint once disentangled from word expectancy (Federmeier et al., 2007; DeLong et al., 2011, 2014a). Constraint and expectancy covary at high levels of predictability (as highly-expected words can only complete strongly constraining sentences) but differ at lower levels (as low-probability words can complete both strongly and weakly constraining sentences)—and N400s to low predictability words are equivalent across high- and low-constraint contexts. N400 amplitudes to incongruent words are also unaffected by individual differences in domain knowledge and the differences in expectation that differences in domain knowledge should entail (Troyer and Kutas, 2018, 2020). These findings align better with accounts in which N400s index the amount of new semantic information that must be activated, not the difference between predicted and actual input (Chia and Willits, 2025). Our view is that, while N400 amplitudes might indeed reflect a process that is sensitive to whether a word has been predicted (as in, e.g., Federmeier and Kutas, 1999a), they do not necessarily index an aspect of the predictive process itself (reviewed in Kutas et al., 2014).

The present study is not designed to adjudicate between differing viewpoints on the N400; however, the interpretation of our study (and what constitutes evidence for prediction) will hinge on which perspective the reader takes. In framing our study, we adopt the view that, although N400 amplitudes can be modulated by prediction (e.g., Federmeier and Kutas, 1999a), they do not directly index the predictive process itself (reviewed in Kutas et al., 2014; see discussion in DeLong et al., 2011). Instead, we take the N400 to reflect multiple mechanisms that together determine the extent to which a context facilitated a word's semantic processing upon receipt. For stronger evidence of the engagement of predictive processing, we instead examine a late positive component, which we term the late frontal positivity (henceforth, LFP), which reliably appears when strong predictions are disconfirmed by plausible input (Federmeier et al., 2007; Lai et al., 2024). Importantly, because LFPs are linked to prediction disconfirmation, they can serve as evidence that prediction had to have occurred in the first place.

LFPs—sometimes described as post-N400 positivities (PNPs), late anterior positivities, or frontal late positive components (LPCs; although this typically instead refers to posteriorly-distributed effects)—emerge ~500 ms after word onset and generally last for several hundred milliseconds. Unlike N400s, LFPs are non-obligatory and generally occur when individuals are likely to be more deeply “considering” (Federmeier, 2022) linguistic input. They are typically elicited by unexpected but plausible continuations of high-constraint sentences—cases in which a predicted word is disconfirmed by plausible but different input (Federmeier et al., 2007; see review in Van Petten and Luka, 2012). For example, “ story ” is a strongly expected continuation in the sentence, “ The little girl wanted her dad to read her a bedtime … ” but this sentence might plausibly continue with a different (unexpected) word, e.g., “ poem ”—which would then likely elicit an LFP. Importantly, late frontal positivities have a functional profile different from post-N400 late positivities that are more posteriorly distributed (frequently referred to as late positive components or complexes; LPCs). In sentence contexts, the amplitude of LPCs is enhanced for a number of reasons, most notably to words which are attention-capturing because they are anomalous or are otherwise highly informative (see reviews in Van Petten and Luka, 2012; Brouwer et al., 2012). In studies assessing memory, larger-amplitude LPCs are also predictive of subsequent memory (e.g., Paller et al., 1987; Contier et al., 2025).

Researchers disagree about the functional significance of the LFP. Some interpretations focus on lexical-level processes such as inhibition of a predicted word or switching from a predicted to an unpredicted alternative (Kutas, 1993; Moreno et al., 2002; Thornhill and Van Petten, 2012). However, these accounts are challenged by findings that anomalous words do not elicit LFPs but instead produce posterior positivities linked to episodic encoding and retrieval (reviewed in Van Petten and Luka, 2012). Thus, the LFP reflects more than a simple “switch” from a predictable toward an unpredictable word.

Some work links the LFP to revision of one's mental model (i.e., situation model; Zwaan and Radvansky, 1998). Support for this view comes from studies examining relatively more conventional vs. novel types of language (e.g., Coulson and Van Petten, 2002; Coulson and Wu, 2005). For example, Davenport and Coulson (2011) observed larger LFPs for low-cloze words used in unconventional compared to conventional ways and suggested that the LFP might reflect processes involved in suppressing a prepotent/expected continuation as well as changing one's interpretation of the meaning of the sentence. Additional evidence comes from multi-sentence discourses in which predictive constraints emerge across several sentences rather than locally (Kuperberg et al., 2020; Brothers et al., 2020). Brothers et al. (2020) found LFPs when plausible unexpected words required readers to update rich situation models constructed over multiple sentences, but not when the predictive context was confined to the final sentence.

These findings have been taken to suggest that LFPs may signal moments when comprehenders must substantively revise mental models after a prediction violation. However, LFPs are also attested in much simpler contexts, where rich mental model formation is unlikely. For example, Federmeier et al. (2010) observed LFPs to atypical (but again plausible) category exemplars following a category cue (e.g., “ a type of bird … goose ”). Chiu et al. (under revision)3 observed LFPs to an unlikely (but appropriate) single word completion of a constraining word stem cue (three initial letters) when the task demanded that comprehenders engage with the unexpected word while also maintaining their original prediction (to report it later). Thus, our working hypothesis is that LFPs occur when individuals have made a strong commitment to some construal of a word, sentence, or other message-level meaning and must hold aspects of that construal in mind while revising elements of it to cope with unexpected but sensible input.

Regardless of what the precise functional specifics of LFPs turn out to be, the eliciting contexts for LFPs have been fairly well-characterized. What all of these situations have in common is that there is a context that permits a comprehender to make a strong prediction, the prediction is disconfirmed, and the comprehender must then deal with multiple, competing representations—the unexpected input as well as aspects of the original context/prediction—in order to integrate the new information or complete their task. Importantly for our purposes, although the exact mechanisms underlying the LFP remain underspecified, their presence is a strong indication that individuals were actively engaging in prediction. We aim to elicit LFPs in the present study as a way to probe whether individuals engage in prediction differentially as a function of their domain knowledge.

Possible relationships between (domain) knowledge and prediction

Before further motivating our study on the relationship between knowledge and prediction, it is useful to describe in more detail how we conceptualize prediction in language comprehension. As detailed in DeLong et al. (2014b), linguistic prediction may include semantic, morphosyntactic, and/or form information about words, or information at a level higher than a word (e.g., syntactic and thematic structures, event knowledge, prosodic contours). Indeed, predictive processing has been evidenced at all of these levels (reviewed in DeLong et al., 2014b). On a broad view, prediction might include “multiple (even incompatible) predictions with varying strengths, all subject to dynamic revision” (Kutas et al., 2014, p. 649).

Federmeier (2007, 2022) emphasizes a specific form of prediction: preactivation of a (restricted set of) potential upcoming words. Evidence for this kind of prediction comes from studies in which contextually anomalous words that share features with (are from the same category as) the most expected word elicit reduced N400s compared to equally anomalous unrelated words (Federmeier and Kutas, 1999a,b; Federmeier et al., 2002), especially when sentential constraint is higher. Federmeier (2007) proposes that this type of prediction is likely a function of left-hemisphere-driven production processes, relying on tightly timed access to producing words in sequence, deployed during comprehension. Supporting this claim, visual hemi-field studies yield evidence that the categorically related anomaly effects described above seemed to be driven by left-hemisphere-biased semantic networks (Federmeier and Kutas, 1999b).

Other types of (pre-)activated linguistic information seem to rely more heavily on right-hemisphere biased networks. Metusalem et al. (2012) found that N400 facilitation for the event-related anomalies described in the previous section was apparent only with presentation to the right hemisphere. Taken together, these results suggest that, when it is possible to do so, the left hemisphere prepares quite specific lexico-semantic information about likely upcoming words using top-down mechanisms (Federmeier, 2007) whereas the right hemisphere maintains a wider scope of information including event knowledge (Metusalem et al., 2016) and contextually-related information that may be used for drawing inferences in both literal and non-literal language like jokes, metaphor, sarcasm, and so forth (Coulson and Williams, 2005; Coulson and Van Petten, 2007).

Implicit in the above description is that prediction depends on relevant knowledge. This has been shown many times at the group level, such that when the context constrains toward a particular word or event (based on norming data at the group level), participants show evidence of prediction, on average. However, to our knowledge, few studies have set out to examine how individual differences in content knowledge specific to the discourse at hand might shape the extent to which individuals actively engage in predictive processing. A recent line of work from our laboratories using the Harry Potter (HP) fictional universe to investigate the impact of individual differences in domain knowledge has laid the groundwork to do this. The HP world offers a controlled yet rich domain in which participants (primarily young adults) naturally vary widely in knowledge. It affords a wealth of narrative and linguistic content, including many people, places, objects, and events occurring in a fictionalized (but often realistic) fantasy world. A benefit of using this approach is that the world is (relatively) constrained, with the “canon” comprising seven books and eight parallel movies. As such, “fictional facts” describing true events in the books can be contrasted with sentences containing false information that may nonetheless seem plausible (given no prior knowledge of the series) or information that is not correct/not expected but somehow relates to the series. Moreover, it is possible to obtain measures of knowledge at the group level (e.g., in sentence-completion studies akin to cloze norming studies), individual level (e.g., by testing participants using Harry Potter trivia), and specific sentence/trial level (e.g., by testing individuals on specific facts described by linguistic materials used in ERP studies).

ERP studies using HP materials, primarily focused on the N400, have shown that domain knowledge modulates immediate contextual facilitation. In particular, greater knowledge increases facilitation for contextually supported words (Troyer and Kutas, 2018, 2020), and knowledge also increases facilitation for contextually related but linguistically incorrect continuations (Troyer and Kutas, 2020). This facilitation of related content seems to be driven primarily by right-hemisphere-biased networks, whereas both hemispheres show basic contextual facilitation (Troyer et al., 2022). Moreover, high-knowledge individuals show contextual facilitation even for facts they do not explicitly know (Troyer et al., 2019, 2024). These findings show that domain knowledge quickly shapes semantic activation. However, none of our studies thus far has specifically examined whether domain knowledge impacts the extent to which individuals actively engage in prediction of specific words as they understand language about that domain in real time.

The current study

In the present study, we aim to directly investigate whether domain knowledge modulates the extent to which comprehenders actively engage in prediction as they read sentences describing expected vs. less-expected-but-plausible vs. implausible events set in the HP universe. We employ a self-paced reading design combined with ERPs (cf. Payne and Federmeier, 2017) to be able to examine both neuroelectric activity and behavior. To ensure that any between-subjects effects are due to domain knowledge, and not any other individual differences, we additionally include blocks of control sentences with expected, unexpected-but-plausible, and implausible sentence continuations (items from DeLong and Kutas, 2020), and we ask participants to complete multiple general language and knowledge measures.

The HP sentence pairs describe events from the world of HP that are somewhat open-ended.4 For example, in the book series, Harry (and other wizards) can conjure a spirit animal that wards off evil creatures. These are called Patronuses, and Harry's Patronus takes the form of a stag. Young wizards like Harry have to practice the Patronus charm (spell) before they can get it right. An example stimulus draws on this knowledge and describes an event where Harry is practicing conjuring his Patronus. Consider the following three versions of this item:

Harry was practicing his Patronus. Today he conjured a stag out of thin air. (Expected)

Harry was practicing his Patronus. Today he conjured a spirit out of thin air. (Unexpected-but-plausible)

Harry was practicing his Patronus. Today he conjured a statute out of thin air. (Unexpected-implausible)

The sentence is designed to set up a context in which it is possible that something atypical might occur. Someone with strong HP knowledge could reasonably imagine that since Harry is practicing, he might not perfectly conjure the final form of his spirit animal. Thus, “ spirit ” is a reasonable and plausible continuation, even though it is not the most expected word. By contrast, “ statute ” is designed to seem implausible.

In this study, we focus in particular on the LFP as a marker of active prediction. If high-knowledge individuals engage more in prediction and are better able to cope with prediction violations when they occur, then we expect to see more LFP activity to unexpected-but-plausible words (e.g., “ spirit ” in the above example), compared to implausible unexpected words (which the literature attests do not elicit LFP effects), as domain knowledge increases. By contrast, if knowledge discourages the use of prediction, we might instead see the reverse pattern, such that low-knowledge individuals are more likely to elicit LFPs—either to the unexpected-but-plausible words or, perhaps, even to the “expected” words (which, for low-knowledge individuals, might not in fact be so expected). In either case, we would take the presence of LFPs to indicate that individuals had attempted to predict a word which was not encountered, thus requiring them to revise their ongoing mental models of the situation being described.

We also examine item-level “pervasiveness” (i.e., how likely it is that a fact is known, in general; see discussion in Knoeferle, 2025).5 Our HP materials vary widely in terms of their pervasiveness. Some sentences include expected critical words that were frequently provided when a separate group of age-matched participants was asked to do a sentence-completion norming task (e.g., most people know that Quidditch is the sport played in the HP universe), whereas the best completions of other sentences were infrequently known. We thus also probe for LFPs as a function of this item-level “difficulty” variable (see also Troyer et al., 2019, in press). Because we expect such effects to differ among those with and without knowledge, for this comparison we focus on high and low knowledge subgroups (based on a median split of HP knowledge scores). We expect item-level difficulty to have little impact on prediction for low-knowledge individuals. However, if strong knowledge of a sentence's contents renders prediction a less useful strategy, then, among high-knowledge individuals in particular, predictive engagement might be higher for generally less-well-known items.

In addition to LFPs, we also examine N400 amplitudes. Regardless of its relationship to engagement in active prediction, greater domain knowledge should yield reduced N400 amplitudes to the expected words (as we have demonstrated in all of our previous HP studies). In contrast, N400 amplitudes to the unexpected and implausible continuations should be invariant with knowledge, as this information is not supported by the context and is never expected; thus, this condition serves as a useful baseline. Of particular interest is whether N400 amplitudes to the unexpected-but-plausible words might differ as a function of HP knowledge. The simplest outcome would be that domain knowledge should not influence semantic access for these words, since, despite being somewhat contextually-related, they were designed to be unexpected for all participants. However, given that high knowledge individuals may construct richer HP-specific situation models of the events described in the text, it is possible that the state of their semantic networks might be more biased toward the HP-specific representation, and thus that they may show less facilitation for the unexpected-but-plausible words, perhaps especially in highly pervasive/easy sentences.

Finally, we also examined overall reading times at the critical words as a function of condition and domain knowledge. We expected that those with high domain knowledge might be more likely to read critical words in the HP sentences faster, and that this might particularly be the case for the expected words. In a more exploratory analysis, we also examined whether there might be differences in patterns of ERP results as a function of individuals' reading speed. For example, Payne and Federmeier (2017) found that individuals tended to exhibit LFPs to unexpected-but-plausible continuations of highly constraining sentences about general topics when they were reading relatively quickly (at a rate similar to the typical rate of presentation in ERP studies of sentence reading, roughly 400–500 ms per word).

Participants

We aimed to include at least 48 participants in our sample, based on the number of participants we have previously included in research investigating the link between individual differences in HP knowledge ERPs recorded during sentence reading (e.g., Troyer and Kutas, 2020, which also include HP sentences with three different types of critical words). To that end, 56 undergraduate students and/or members of the University of Illinois community (mean age = 19; range = 18–30; 32 women, 16 men, 8 not reporting) took part in the study for course credit or cash. Eight participants were excluded from data analysis due to excessive artifact, leaving a total of 48 participants whose data were included. All participants provided informed consent. The study was approved by the Institutional Review Board at the University of Illinois.

Materials

Sentence materials

During the EEG study, participants read three blocks of control sentence pairs followed by three blocks of HP sentence pairs. In each, a two-sentence context led up to a critical word which was variously the best/most typical continuation, an unexpected but plausible continuation, or an unexpected and implausible continuation (see below for details) and was always close to the end of the sentence. The same contexts (90 Control, 90 HP) were paired with three critical words, and three lists were created such that each participant read only one version of each sentence pair. A programming error affected our random assignment to list, resulting in over-sampling of some lists compared to others; upon noting the error, we ran more participants until there were a minimum of 8 in each list. Because of this issue, we include list as a factor in our statistical models (see below).

Control sentences

Ninety control sentence pairs (first sentence ranging 3–15 words, mean = 6.7; second ranging 3–11 words pre-critical-word, mean = 6.8, and ranging 2–3 words post-critical word, mean = 2.0) described everyday topics and events (examples in Table 1). All sentence pairs were highly constraining (mean cloze of best completion = 94.1%; range = 87–100%). These sentences were a subset of materials from a previously published study (DeLong and Kutas, 2020). Critical words were either the best completion (Expected), unexpected but plausible continuations (henceforth, Plausible), or unexpected and implausible continuations (henceforth, Implausible). Neither the plausible nor implausible words were ever produced during cloze norming.

Harry Potter sentences

Ninety Harry Potter (HP) sentence pairs were constructed as follows. Using freely available materials (including the text of the Harry Potter books, Harry Potter fan pages, Harry Potter trivia sites, and other Harry Potter-related websites), the first author along with an undergraduate research assistant created a set of sentence pairs that described situations that could take place in the Harry Potter universe but that were relatively open-ended vignettes. The first sentence always set up a context in the HP world (for example, “ At the start of the new school year, Harry needed some cash.”). The second sentence was written so that, in isolation, it would be an acceptable English sentence with no obvious reference to HP (e.g., “ He decided to stop by …”). We thus constructed the sentences so that, while there would be a standard/typical continuation consistent with the HP books (here, Gringotts —i.e., Harry Potter's bank), other continuations were possible. More examples are provided in Table 1.

After generating the sentence contexts, we conducted two cloze norming studies. In the first, we asked a group (N = 50) of age-matched participants recruited through Prolific.com (mean age = 24, range = 18–30; 35 female, 15 male) to read the two-sentence vignettes up to the word prior to where the critical word would occur. They were then asked to provide the word that would best continue/complete the sentence pairs. They were provided with the following example: “ Harry was walking to class. On the way, he stopped to talk to a… ” and instructed that they might continue this with “professor,” “friend,” or any other word they thought would best continue the sentence. These participants also completed our measure of Harry Potter knowledge (see below; mean score = 6.27 out of a possible 10 points; range: 3–10). On average, sentential constraint (here defined as the proportion of individuals who converged on the same best completion/continuation) was 33%, ranging 0–76%.

We next conducted a second norming study in which we asked another group (N = 50) of age-matched participants recruited from Prolific.com (mean age = 25, range = 19–30; 36 female, 14 male) to read only the second sentence from each vignette, up to the word prior to where the critical word would occur. Thus, for the example described in the preceding section, participants would only read, “ He decided to stop by… ”. By design, these sentences were relatively open-ended, and participants did not typically converge on a single best completion (mean constraint = 15%; range = 4–62%).

To complete our set of sentence materials, we matched each sentence pair with three types of continuations: an Expected continuation (i.e., the most standard/typical continuation according to the HP books), an unexpected but Plausible continuation (by sampling from responses provided during the single-sentence norming study6), and an unexpected and Implausible continuation (judged by the authors to be anomalous in context). We then wrote short endings to complete the second sentence so that the critical words were sentence-medial. On average, the first sentence of each sentence pair was 9.2 words long (range = 3–16); the second sentence contained an average of 6.2 words prior to the critical word (range = 2–11) and an average of 4.0 words after the critical word (range = 2–8). Each participant read only one version of each sentence pair.

Measures of individual differences

Harry Potter knowledge quiz

We estimated participants' knowledge of Harry Potter from their score on a trivia-style quiz containing 10 multiple choice questions (see Troyer and Kutas, 2018, for more details).

Harry Potter self-report

We also asked participants questions about their experience with the Harry Potter book series and related materials, e.g., how many times they had read each of the HP books, seen each of the HP movies, listened to audio books, and/or participated in other activities related to HP (reading/writing fan fiction, engaging with web-based content, etc.). As an estimate of overall experience with HP, a numeric score was determined by summing the total number of times an individual had read each book, seen each movie, and so on. We report this number along with statistics on the number of times participants had read each book and seen each movie, on average.

General knowledge quiz

To estimate individual differences in ability to respond to trivia questions, participants additionally completed a 30-question trivia-style quiz asking questions about everyday topics relating to entertainment, literature, politics, religion, science, and history (see Troyer and Kutas, 2018, for more details).

Measures of print exposure

To estimate individual differences in print exposure, participants completed the author and magazine recognition tests (ART and MRT; Acheson et al., 2008). These asked participants to view checklists of 80 names (ART) or titles (MRT) and choose the ones they recognized as authors or magazines. Scores were calculated by subtracting the proportion of false alarms from the proportion of correct names checked.

Procedures

Ordering of tasks

During set-up for the EEG experiment, participants completed the ART and MRT. After EEG recording, participants completed questionnaires in the following order: HP knowledge quiz, HP self-report, and general knowledge quiz.

EEG experiment

Participants were instructed to remain relaxed and still during the EEG experiment to minimize muscle artifact. They were told they would be reading two-sentence stories across six blocks (first three blocks about general topics, remaining three blocks about Harry Potter). They were instructed that after each block they would complete a short memory task (see below). They were given practice trials to read (similar to the sentences in the control condition) before the first block of the study.

During the EEG experiment, participants sat approximately 100 cm in front of an LCD computer monitor. The background of the screen was black and words were presented in white type. Each trial began with a blank screen for 1 s. Then, the first sentence of each pair appeared on the screen until the participant pressed a button to advance to the next sentence. After their button press, a crosshair appeared in the center of the screen for a duration that varied randomly between 900 and 1,100 ms. Participants were instructed to focus on the crosshair and not move their eyes while it was on the screen. The second sentence was then presented one word at a time right above the crosshair. Participants pressed a button to advance through the second sentence, word by word. Each word was presented for a minimum of 100 ms plus an ISI of 300 ms, meaning that there was a minimum of 400 ms SOA. This ensured that participants could not skip over words without enough time to process them. After the sentence-final word disappeared, the crosshair stayed on the screen for a duration that randomly varied between 900 and 1,100 ms.

After each block of 30 sentences, participants completed a brief memory task that asked them to provide old/new judgments about words, some of which had been previously seen in the EEG experiment (from each condition) and some of which had not. On each trial, participants viewed a crosshair presented for a second. Next, a word appeared. Participants were instructed to place their fingers on four keys on a standard keyboard (“d,” “f,” “j,” and “k”) and respond with whether they judged the word to be old (“d”), maybe old (“f”), maybe new (“j”), or new (“k”). Because this task was not a main focus of the present investigation, we report only a d ′ measure of hits (correct responses to old/previously seen items) vs. false alarms (responding that a new/unseen word was old/previously seen, excluding lures), collapsing across conditions, as a check that participants were attending to the sentence materials.

EEG recording

The electroencephalogram (EEG) was recorded from 26 equally-spaced electrodes attached to an elastic electrode cap (see Ganis et al., 1996, Figure 9, for visualization and comparison with the International 10–20 system). Six additional electrodes were attached, including two placed lateral to the outer canthus of each eye (for use in monitoring eye movements), two placed under each eye (for use in monitoring blinks), and two placed behind the ears on each mastoid (for use as reference). Prior to recording, electrode impedances were lowered to under 5 kΩ. The signal was amplified using a BrainVision BrainAmp DC amplifier set at a bandpass of 0.02–250 Hz and a sampling rate of 1 kHz. During recording, the online reference was the left mastoid; after recording, all cap electrodes were re-referenced offline to an average of the left and right mastoid. In addition, the left horizontal eye channel was re-referenced offline to the right horizontal eye channel to create a bipolar channel to monitor for horizontal eye movements.

Data analysis

EEG

Filtering, artifact correction, and epoching

Data were digitally lowpass filtered at 30 Hz. For plotting purposes only, we filtered the data using a lowpass filter of 10 Hz. We extracted single-trial epochs of EEG data from the continuous recordings from 200 ms prior to the onset of a critical word until 900 ms post-critical word. For each electrode, the 200 ms pre-stimulus baseline was subtracted from the waveform. Trials contaminated by eye movements, blinks, muscle activity, blocking, or other artifact were removed from subsequent analysis. This resulted in an exclusion of approximately 10% of trials across conditions: 10.6% Control-Expected; 10.5% Control-Plausible; 9.8% Control-Implausible; 9.8% HP-Expected; 10.1% HP-Plausible, 10.2% HP-Implausible.

Time windows and regions of interest

For statistical analyses, we measured ERPs across two time windows: (1) 250–500 ms after critical word onset, corresponding to the typical peak of the N400 (and following Troyer and Kutas, 2018) and (2) 500–800 ms after word onset, during which late frontal positivities (LFPs) are typically observed (Federmeier et al., 2007, 2010). For N400 amplitude measurements, we selected a priori a centro-parietal region of interest (ROI) where N400 are typically most prominent, averaging across 8 electrodes placed at midline central, midline parietal, left and right mediocentral, left and right mediolateral central, and left and right medial occipital locations (see Federmeier and Kutas, 1999a, for a visualization of these electrodes, therein labeled MiCe, LMCe, RMCe, MiPa, LDPa, RDPa, LMOc, and RMOc); this same ROI was used in Troyer and Kutas (2018). For LFP amplitude measurements, we used an a priori (Federmeier et al., 2010) region of interest over frontal electrodes, including midline prefrontal, left lateral prefrontal, right lateral prefrontal, left medioprefrontal, and right medioprefrontal (MiPf, LLPf, RLPf, LMPf, RMPf). Analyses were performed separately for the control and HP blocks.

Statistical analysis

For the N400, we were primarily interested in effects of contextual support (Implausible vs. Expected) and plausibility (Implausible vs. Plausible); for completeness, we also examined Expected vs. Plausible continuations (“expectedness” of Plausible continuations).

For the LFP, we were primarily interested in the difference between the two types of unexpected words (as the LFP is known to depend on plausibility). For completeness, we also examined each unexpected item type vs. expected words.

Our main research questions centered around whether HP knowledge would modulate these ERPs. In testing this, it was important to also rule out the possibility that such a relationship was due to other individual differences co-occurring in the group. To do this, we examined whether HP knowledge influenced the neural response to effects of contextual support in the control sentences. Additionally, in cases where HP knowledge was shown to modulate effects of interest, we also examined whether other measured individual differences better accounted for these group differences. For the latter, we tested models incorporating verbal fluency (VF), general knowledge test (GKT) performance, and an aggregate measure of print exposure (combining ART/MRT).

For all ERP statistical analyses, we used linear mixed-effects models (Baayen et al., 2008) fit to ERPs from each time window and corresponding ROI (N400, LFP). For the majority of analyses (all but the analyses including trial-level measures), mixed-effects models were fit to participant-averaged data and included by-participant random intercept terms. Given that our counter-balancing lists were not evenly balanced across participants (with more participants having received List 1 than the other two lists), we also included List as a fixed effect in statistical models. All analyses used sum coding for categorical variables and z-transformed values for continuous variables. For statistical inferences about predictors of interest, we conducted model comparison using Chi-square tests on nested models, comparing a full model with a reduced model that omitted the term being tested, using the anova() function in R version 4.0.2, with f 2 effect sizes computed for significant effects based on marginal r 2 values (constituting variance explained by fixed effects) computed using the r2_nakagawa() function from the “performance” package. This approach is standard practice in psycholinguistic studies with mixed-effect designs (see Schad et al., 2020). For example, to test interaction effects, we compared models including an interaction (e.g., condition X HP knowledge) against models with only additive fixed effects (condition + HP knowledge). To test effects of single fixed effects (e.g., condition), we compared models using the effect vs. models with only an intercept term (and no fixed effects). We provide representative examples of these model specifications in the Supplementary materials. We report the full models most important for our main points in the main text, using the Satterthwaite option for denominator degrees of freedom for F statistics. We additionally report partial effect sizes (η 2) for each predictor, computed using the F_to_eta2 function from the “effectsize” package.

Finally, we also examined the ERPs to HP sentences (N400s and LFPs) as a function of trial-level cloze probability (i.e., item-level difficulty), determined by an offline norming study from a separate group of participants. For these analyses, we fit mixed-effects models incorporating by-participant and by-item random intercept terms to trial-level ERPs. To examine effects of item difficulty within relatively high- vs. low-knowledge individuals, we additionally examined “high” and “low” subgroups of our participants based on a median split of HP knowledge scores.

Behavioral data

Measures of individual differences

Table 2 presents descriptive statistics of all measures of individual differences. Our main measure of HP knowledge was derived from the 10-question HP trivia quiz. The median score for this measure was 5 out of 10 (M = 5.40, SD = 2.2, Range = [1–10]).

Table 3 presents correlations between measures of individual differences. In addition to being correlated with other measures of HP experience, HP knowledge scores were also correlated with scores on the general knowledge test. This is a pattern we have previously observed in studies examining HP knowledge.

Reading times

Reading times to critical words of each type in Control and HP sentences are presented in Figure 1. Across conditions, critical words in HP sentences were read more quickly than critical words in control sentences [ χ (1) 2 = 45.26, p < 0.0001, f 2 = 0.075]. Among critical words in control sentences, Expected words were read more quickly than either Plausible (t (47) = −2.15, p = 0.037, Cohen's d = −0.31, 95% CI = [−0.60, −0.02]) or Implausible (t (47) = −3.00, p = 0.004, Cohen's d = −0.43, 95% CI = [−0.73, −0.14]) words. There was no interaction between reading times and HP knowledge for the control sentences (p = 0.663). Among critical words in HP sentences, there were no differences in reading times among the three condition types and no interaction with HP knowledge (p s >0.258). Given the lack of significant effects of condition on reading time in HP sentences, we report variation in ERPs according to reaction time (as well as sentence type, condition, and HP knowledge) only in the Supplementary Figures 1–3. Overall, there were no major differences in ERP patterns based on reading times.

Memory task performance

Among Control sentences, the mean hit rate (correct “old” responses) was 67% (SD = 20%) and the false alarm rate (incorrect “old” responses) was 24% (SD = 18%); d ′ = 2.26. Among HP sentences, the mean hit rate was 72% (SD = 21%) and the false alarm rate was 21% (SD = 15%); d ′ = 2.74. Performance was thus well above chance for each sentence type, indicating that participants were attending to the sentences.

ERP data

Grand-averaged ERP data

Grand-averaged ERP data for both Control and HP sentences are presented in Figure 2. As typically seen in ERP studies averaging across words presented with RSVP, across conditions, we see the canonical waveform morphology across the majority of electrodes with an N1, followed by a P2, followed by an N400.

HP sentences

N400 time period (250–500 ms)

Analyses on N400 amplitudes to critical words in HP sentences across all participants revealed a main effect of Condition [ χ (1) 2 = 42.22, p < 0.0001, f 2 = 0.232]. As we predicted, Expected words elicited reduced N400s compared to both Implausible [ χ (1) 2 = 27.47, p < 0.0001, f 2 = 0.296] and Plausible [ χ (1) 2 = 19.94, p < 0.0001, f 2 = 0.157] critical words; Plausible words also elicited reduced N400 amplitudes compared to Implausible [ χ (1) 2 = 7.79, p = 0.005, f 2 = 0.040] critical words.

Critically, HP knowledge interacted with condition [ χ (1) 2 = 6.58, p = 0.037, f 2 = 0.026; Figure 3a]. As expected, there was a significant interaction between HP knowledge and condition on the N400 effect of contextual support [Expected vs. Implausible; χ (1) 2 = 4.25, p = 0.039, f 2 = 0.032] (Table 4). HP knowledge also affected expectedness effects [Expected vs. Plausible; χ (1) 2 = 3.92, p = 0.048, f 2 = 0.023]. There was no significant effect of knowledge on the N400 effect of plausibility [Plausible vs. Implausible; χ (1) 2 = 0.41, p = 0.522]. Simple correlations between HP knowledge and each of the pairwise effects revealed the same pattern (Figure 3b). These effects were driven by a significant correlation between HP knowledge and N400 amplitudes to Expected words (r = 0.26, p = 0.037, 95% CI = [0.021, 1.00]). There was no significant correlation between HP knowledge and N400s to either of the other two continuation conditions (Figure 3c). Thus, as in previous studies (e.g., Troyer and Kutas, 2018, 2020), our measure of HP domain knowledge had a specific influence on N400 amplitudes when recorded to words in supportive contexts that required specific knowledge about the HP world to understand.

We also asked whether a model incorporating additive effects of all individual difference measures (of HP knowledge, general knowledge, and print exposure), as well as interactive effects between these measures and condition, might better explain our findings that HP knowledge modulated N400 effects of contextual support (Expected vs. Implausible). Nested model comparison between this larger model and the simpler model (which did not contain any effects of general knowledge or print exposure) indicated that the simpler model (Table 4) was preferred, suggesting that measured individual differences beyond HP knowledge did not better explain the N400 effect of contextual support (p = 0.130). For completeness, we present the results of the full model in the Supplementary Table 1.

In addition, we conducted trial-level analyses to investigate joint effects of HP domain knowledge and item-level difficulty (i.e., pervasiveness). Across participants, we observed a significant interaction of item difficulty and condition on N400 amplitudes [ χ (2) 2 = 17.94, p < 0.001, f 2 = 0.006], which, critically, was modulated by HP knowledge [ χ (2) 2 = 11.50, p = 0.042, f 2 = 0.003]. Planned follow-up tests investigated effects of item difficulty for relatively high knowledge compared to low knowledge subgroups, based on a median split. Item difficulty and condition had an interactive effect on N400s to critical words in HP sentences among the high knowledge [ χ (1) 2 = 18.20, p = 0.002, f 2 = 0.045] but not the low knowledge group (p = 0.20; Figure 4).

To better understand these effects, we also investigated the extent to which item difficulty and HP knowledge interacted to predict N400 amplitudes to words in each condition. Item difficulty and HP knowledge had an interactive effect only on N400 amplitudes to the Plausible words [ χ (1) 2 = 6.12, p = 0.037, f 2 = 0.006], with the effect of item difficulty being significant among high-knowledge individuals [ χ (1) 2 = 4.37, p = 0.037, f 2 = 0.010] and absent when considering only the relatively low-knowledge individuals (p = 0.965). There were no interactive effects between item difficulty and HP knowledge on N400 amplitudes to the Expected words nor to the Implausible words. In sum, high-knowledge individuals showed reduced N400 amplitudes to (i.e., facilitation of) Plausible words continuing sentences describing relatively less pervasive/less well-known events from the HP books, suggesting they were less likely to engage in active prediction when the sentence contents were well-known.

Late frontal positivity time period (500–800 ms)

Across participants, there was an effect of condition on LFPs [ χ (2) 2 = 17.03, p < 0.001, f 2 = 0.070]: LFPs to Implausible words were smaller than those to either Expected [ χ (1) 2 = 14.23, p < 0.001, f 2 = 0.082] or Plausible [ χ (1) 2 = 10.25, p = 0.001, f 2 = 0.067] critical words, which did not differ in amplitude [ χ (1) 2 = 0.31, p = 0.576]. When considering all three conditions, there was no significant interaction between HP knowledge and condition [ χ (2) 2 = 3.44, p = 0.180]. However, analyses focusing on HP knowledge modulations of each effect type separately revealed the predicted numerical pattern in the key Plausible vs. Implausible contrast, albeit one which only approached significance [ χ (1) 2 = 2.82, p = 0.093], with larger LFPs to Plausible compared to Implausible prediction violating words among those with higher knowledge (Figure 5a). This pattern received support in the simple correlations between HP knowledge and the size of each of the pairwise effects (Figure 5b), which showed the same pattern; the Implausible minus Plausible LFP showed a numerically inverse relationship with domain knowledge (r = −0.24, p = 0.051), being larger for those with higher domain knowledge. Correlations of knowledge with mean LFP amplitudes in individual conditions suggested this pattern was largely driven by the relationship between HP knowledge and LFPs to plausible continuations (Figure 5c).

We additionally conducted trial-level analyses to investigate joint effects of HP domain knowledge and item-level difficulty (i.e., pervasiveness) in the LFP time period. There was no significant three-way interaction between item difficulty, condition, and HP knowledge on LFP amplitudes [χ 2 (5) = 7.18, p = 0.207]. Our analyses focusing on effects of item difficulty for the high knowledge and low knowledge subgroups (as above) revealed an effect of item difficulty specific to the Plausible condition within high-knowledge individuals [ χ (1) 2 = 5.28, p = 0.022, f 2 = 0.1 02]: LFPs were larger for Plausible words in more difficult (i.e., less well-known) compared to easier (i.e., more well-known) sentences among high-knowledge individuals (Figure 4). Item difficulty did not have a significant effect on any other condition among either the high or low knowledge groups. In sum, individuals with high knowledge showed evidence for prediction, indexed by LFPs, specifically when items were overall more difficult/relatively less well-known.

Control sentences

Replicating previous studies (DeLong et al., 2014a; DeLong and Kutas, 2020), we observed a main effect of condition on N400 amplitudes [ χ (1) 2 = 69.83, p < 0.0001, f 2 = 0.352], with Expected critical words eliciting reduced N400 amplitudes compared to Implausible [ χ (1) 2 = 43.51, p < 0.0001, f 2 = 0.472] and Plausible [ χ (1) 2 = 30.45, p < 0.0001, f 2 = 0.279] critical words; Plausible words also elicited reduced N400 amplitudes compared to Implausible words [ χ (1) 2 = 14.74, p < 0.001, f 2 = 0.037]. There was also a main effect of condition on LFPs [ χ (2) 2 = 22.29, p < 0.0001, f 2 = 0.145]; however, the pattern was somewhat different from that previously observed in the studies by DeLong and colleagues (DeLong et al., 2014b; DeLong and Kutas, 2020). Compared to Implausible words, LFPs were larger in amplitude for both Expected [ χ (1) 2 = 15.38, p < 0.0001, f 2 = 0.168] and Plausible [ χ (1) 2 = 21.34, p < 0.0001 f 2 = 0.157] critical words, which did not significantly differ from one another (p = 0.898). Thus, under conditions of self-paced reading, we observed a somewhat different pattern than that previously attested, with greater LFP activity for plausible words only when compared to implausible (and not to expected) words.7 Critical for our purposes, there was no interaction between condition and HP domain knowledge during either time period (p s >0.291), supporting our interpretation that individual differences seen among the HP sentences do not reflect general differences between participants and instead reflect specific differences in domain knowledge.

Here, we asked whether variation in domain knowledge would influence neural processes underlying prediction during language comprehension. We extended previous findings showing that domain knowledge has immediate effects on the use of context during language processing. In particular, N400 amplitudes to the most expected words in HP sentences about possible HP-related events were reduced for relatively high- compared to low-knowledge individuals. By contrast, when considering all stimuli together, HP knowledge had little or no impact on processing contextually unexpected words, whether plausible or implausible—although this pattern was modulated by the relative difficulty of the individual items in a manner that provides insight into how knowledge is deployed during comprehension.

A key goal of this study was to examine the potential impact of domain knowledge on late frontal positivities (LFPs) to plausible prediction violations, providing a more direct index that prediction has occurred. Our strongest evidence comes from item-level difficulty effects we observed among high-knowledge readers: High-knowledge individuals were particularly prone to producing large LFP effects (i.e., larger LPFs to plausible but unexpected compared to predictable or anomalous continuations) in sentences describing facts that were harder/less likely to be known. This suggests that active prediction is a strategy deployed to deal with particular comprehension conditions—including those in which individuals with existing relevant knowledge may be using prediction to learn more about less-well-known content. We discuss the implications of these findings below.

The relationship between domain knowledge and prediction

Of primary interest, our results suggest that prediction is stronger/more likely among high- compared to relatively lower-knowledge individuals when reading about events that might happen in the HP universe. Importantly, this relationship was moderated by overall item-level difficulty. Among low-knowledge individuals, item difficulty did not seem to have much of an effect on LFPs to unexpected-but-plausible items. This finding is not surprising, given that these individuals are less likely to know much about any of the events/topics described in the sentences. By contrast, item difficulty was important in determining whether or not (or perhaps the extent to which) unexpected-but-plausible words elicited LFPs among high-knowledge individuals. For words continuing sentences that described events that were well-known among a group of peers, high-knowledge individuals showed minimal evidence of LFPs. By contrast, high knowledge readers did elicit reliable LFPs to unexpected-but-plausible words continuing less well-known (i.e., more difficult) sentences. Taken together, these patterns suggest that individuals with greater knowledge about HP were more likely to engage in active prediction than those with less knowledge, but not always. Specifically, high-knowledge individuals seemed to engage actively in prediction—thus needing to engage mechanisms to override that prediction when confronted with an unexpected, plausible word—only when they were reading sentences that were less likely to be well known.

There are several reasons why this might be the case. One possibility is that high-knowledge individuals might have been less confident in their knowledge for the less-well-known sentences. Another is that the specific knowledge might not have been available to them at all, perhaps because they did not remember that particular fact. In either case, the less-well-known sentences might have encouraged high-knowledge individuals to actively engage in attempting to guess what the next word might be, perhaps to ease comprehension or, more likely, to facilitate learning. By contrast, when high-knowledge individuals were asked to comprehend sentences that reflected well-known or “over-learned” information, they might not have engaged as deeply in comprehension, perhaps because they were not as invested in learning any new information from the sentence contents.

Thus, when linguistic content draws on domain knowledge that is less pervasive/less well-known, high-knowledge individuals might engage more deeply, leading them to engage in active prediction of upcoming words. When, instead, they encounter unexpected but plausible words, they would then need to hold on to a message-level meaning (that would have resulted from the predicted word) in order to make the revision required to process the contextualized meaning of the actually-encountered word. This is precisely what we believe LFP activity to reflect: some combination of having to shift one's mental representation toward an unexpected input while having to maintain aspects of the (initial) mental representation that allowed one to generate a more dominant prediction (see discussion in Chiu et al., under revision).

There is an alternative interpretation we would also like to consider which takes a somewhat different view of the relationship between LFPs and prediction than the current literature does. It is possible that there are separable mechanisms underlying the act of predicting a word (and/or its specific semantic content) and the act of immediately revising one's mental model based on a (plausible) prediction disconfirmation. If these mechanisms are indeed separable, then it is possible that high-knowledge individuals do typically engage in prediction, even for “overlearned”/well-known content, but that when they encounter unexpected (and potentially plausible) words in highly-constraining (i.e., “easy” or well-known) sentences, they treat the so-called “plausible” words as anomalous and reject the opportunity to integrate them into their working representation of the message-level content. If this is the case, they would not be expected to show LFPs to the “plausible” words—because the high knowledge individuals would not judge them to be plausible. This raises the interesting possibility that having high domain knowledge may interrupt processes that allow individuals to quickly revise their mental models.

Thus, there are at least two reasons why high-knowledge individuals may not have shown strong LFP activity for well-known sentences. We lean toward the simplest interpretation, which is that they were not predicting in these cases. However, we recognize that high-knowledge individuals might also be less willing to incorporate highly atypical (though plausible) continuations into their ongoing message-level representations, which could explain the absence of LFPs in these cases. It could then follow that individuals did predict in these cases, but then did not revise their predicted representations, or it could be that individuals did not engage in prediction because they recognized ahead of time (whether consciously or unconsciously) that they would be unwilling to revise a prediction.

These differing possibilities could be empirically tested using downstream assays of memory. Words deemed contextually anomalous/implausible have previously been shown to disrupt memory for sentence-level meaning but can lead to stronger episodic memory at the individual word level, which often engenders greater late posterior components that reflect aspects of episodic memory formation Hubbard and Federmeier, (2024). Thus, if individuals with high domain knowledge immediately deem atypical continuations to highly-pervasive information to be implausible, these words might impair memory for message-level content but spare (or even enhance) episodic memory for the words themselves. This is a hypothesis that we plan to empirically test in future work.

The relationship between domain knowledge and mental models

Though N400s are not directly yoked to prediction, our pattern of N400 effects among HP sentences informs an understanding of how domain knowledge influences real-time dynamics of accessing or building mental models, which may contribute to predictive processing. In both control and HP sentences in our study, N400 amplitudes were reduced for plausible but unexpected words compared to anomalous words. Interestingly, when considering all items together, there was no reliable relationship between HP domain knowledge and N400 amplitudes to unexpected but plausible words, suggesting that domain knowledge had a specific influence in making domain-related content available, but did not much influence the availability of everyday content. Interestingly, however, high-knowledge individuals' N400s showed a sensitivity to item-level difficulty, such that N400s were relatively larger (i.e., not as reduced/facilitated) for unexpected-but-plausible words in contexts describing events that were “easier”/more well-known compared to those in more difficult contexts.

We propose that this reflects a difference in how individuals might construct mental models in real time (as they comprehend language) as a function of their degree of relevant background knowledge. For example, consider our low-pervasiveness (difficult) sentence pair, “ Hermione was looking for Madam Pomfrey. The first place she tried was the hospital …”. In our norming study, the most commonly provided continuation, “ hospital,” was provided by only 6% of our participants. By contrast, most participants (76%) provided the best continuation in the sentences, “ The students were going to the candy store. They wanted to buy some chocolate frogs …”. We suspect that, even among our high-knowledge individuals, it may have been easier to construct an appropriate mental model in the second case. There are many scenes/events in the Harry Potter books and movies that depict students eating or shopping for chocolate frogs, so when individuals first encounter the setting of a candy store, it may be easy to bring strongly domain-related information to mind, thus affording those with high knowledge a rich “Harry Potter” lens on the sentence contents. By contrast, in the former sentence pair, readers would need to quickly remember/know and understand that Madam Pomfrey is the wizarding school nurse and thus would be likely to be found in the hospital. This information is not as pervasive—i.e., not as well-known among our norming participants, and less central to the events that occur in the HP book series and movies.

If this proposal is accurate, it would follow that even high-knowledge individuals might be less likely to access rich situation models when understanding sentences about less (compared to more) pervasive information. They might thus rely on more on typical, everyday knowledge to mentally simulate the contents of the sentences. In the example where Hermione is searching for someone called Madam Pomfrey, they might instead think about multiple locations where one might search for an adult at school (including a mailroom, which constituted the unexpected-but-plausible continuation for that sentence). Future work could test this proposal by examining the extent to which event-related knowledge is available as a function of more vs. less pervasive contexts (both with respect to domain knowledge and more general knowledge; see discussion in Knoeferle, 2025).

Thus, the present results suggest a small/minimal facilitatory effect of domain knowledge on unexpected, domain-unrelated but linguistically plausible words. By contrast, domain knowledge has been shown to more strongly facilitate processing of contextually anomalous words that are yet domain-related as well as conceptually related to the specific content of the sentence. Troyer and Kutas (2020) asked young adults to read short vignettes describing fictional “facts” about people/things/places/events from the HP series. In that study, sentence-final critical words were either expected (i.e., correct, based on the stories), anomalous and unrelated, or anomalous but related to the contents of the sentence (either because they were related thematically to the overall event being described, related categorically to the critical word, or both). For example, participants might read, “ Harry Potter's parents were killed by Voldemort. Harry was left with a lightning { scar / Lily / Fudge }.”, where “ scar ” is the correct, expected continuation, “ Fudge ” is incorrect/anomalous and unrelated to the sentence's contents, and “ Lily ” is incorrect/anomalous but is related to the events described by the sentence: Lily is the name of Harry's mother. HP domain knowledge was found to systematically reduce N400 amplitudes recorded to the expected words (replicating prior results; Troyer and Kutas, (2018) and, critically, to the related —but not the unrelated—anomalous words. These findings strongly suggest that domain knowledge can rapidly allow individuals to bring specific mental/situation model to mind such that linguistically inappropriate but contextually related information is readily available.

Taken with these previous results, our present findings suggest that when individuals have the strongest knowledge of sentences (i.e., when individuals are high-knowledge and the sentence contents are generally well-known), they may experience a “domain override” such that they rely specifically on domain-specific knowledge and do not engage as deeply with lexico-semantic information that does not fit their domain-specific mental/situation model. In other words: those with strong domain knowledge might more strongly activate the domain knowledge and not activate more general knowledge based on the statistics of general/everyday language. By contrast, those with weak domain knowledge might rely relatively more on their general linguistic knowledge to read sentences when they lack domain-specific background knowledge—for example, when reading about a character called Peter Pettigrew changing shape, they likely bring to mind knowledge that when fictional characters change shape, they often take the form of another living being (whether that be a rat or a zombie).

The relationship between mental models and prediction

Taken together, our findings suggest that the ability to quickly build and maintain a rich mental model may have a complex relationship with prediction. On the one hand, some researchers have suggested that the ability to build a rich mental model may be a necessary precursor of (or at least an important contributor to) prediction in language comprehension. For example, Brothers et al. (2020) found evidence that individuals engaged more in prediction when discourse contexts described rich events that required integration across sentence boundaries in order to accurately arrive at an appropriate prediction. However, the present findings suggest that there may be a more nuanced picture in which the ability to construct rich mental models sometimes allows individuals to access likely upcoming words “for free” without necessitating (potentially costly) engagement in active prediction. This suggests that there may be a “sweet spot” for prediction. It is likely that individuals need to be motivated enough to become actively engaged in predictive processes (cf. Lau et al., (2013); Brothers et al., (2017). It thus stands to reason that reading a rich context may be likely to engage readers and lead them to predict upcoming words and/or events compared to reading impoverished contexts (though see Federmeier et al., 2010 and Chiu et al., under revision for evidence the readers can predict even in minimal single-word or word-pair contexts). However, it is also possible that when the sentence context allows comprehenders to immediately access rich knowledge structures (including those supporting the availability of detailed mental/situation models), they rely more strongly on this information in a way that makes prediction of a specific word (in sequence) a less useful strategy (or at least one that individuals are less likely to engage in, no matter the specific reason; see discussion in Troyer et al., 2022).

That said, we do find evidence that those with high domain knowledge show the greatest propensity for prediction, notably when the sentence contents described difficult/less-well-known events. We suggest that the availability of (some) knowledge plays an important role in actively generative upcoming predictions, at least at the word level. This accords with findings that prediction and the corresponding LFP response to prediction violations seem to be most prevalent among highly constraining sentences and weaker or non-existent among low-constraint sentences (Federmeier et al., 2007; reviewed in our introduction). By contrast, when individuals are able to quickly bring to mind perhaps more fully-fleshed-out, rich situation models of what the language describes—as is likely to be the case for experts of a closed-world fantasy series like the one studied here—they may be less likely to guess (predict) because they “know” what should come next.

Limitations and future directions

The present study has several limitations that could be addressed in future work. First, we elected to use a self-paced reading task given previous reports that it might be an important variable for understanding prediction and sorting LFP responses (Payne and Federmeier, 2017). However, reading time did not much relate to ERPs in the present study. Instead, using a self-paced task might have introduced more noise, influencing our ability to detect LFP effects—especially given their protracted nature and the fact that their timecourse overlaps with the (variable) onset time of presentation for the word following the critical word. More generally, the variability of reading times across all of the words in the second sentence of our materials might have contributed additional noise to our data, potentially making it harder to detect ERP effects overall. We therefore plan to conduct an extension of this work using rapid serial visual presentation (RSVP) to ensure that all participants receive materials on the same time course.

In addition, compared to many of our previous ERP studies investigating effects of domain knowledge on word-by-word reading, we used relatively fewer words per condition (30 here compared to ~50 in most previous studies). This choice was made so that we could include the same number of Control as HP sentences without requiring an overly lengthy experimental session for participants. Future work investigating prediction/LFPs as a function of domain knowledge could benefit from a more specific focus on the domain of interest, thus allowing experimenters to increase the number of stimuli per condition.

The combination of self-paced reading and smaller trial numbers may have meant that we were underpowered to detect small effects of HP knowledge on LFP effects. Our study tested a number of participants (N = 48) equivalent to or larger than the number of participants in our previous HP studies, which have repeatedly shown robust effects of domain (HP) knowledge on N400 effects of contextual support and other semantic manipulations, and we again found robust effects of individual differences on the N400. However, there was not the same breadth of knowledge on which to base our sample size for LFP effects, which, different from N400s, are non-obligatory, being apparent in moments where individuals are engaged in active prediction, and which can thus fluctuate over the course of an experiment (e.g., Payne and Federmeier, 2017; Jongman et al., 2023). We therefore aim to include a larger sample size in future work investigating individual differences in LFPs.

Nonetheless, our results present novel evidence that domain knowledge can influence the predictive processes underlying real-time language comprehension. In future work, we plan to further explore the possibility that strong knowledge of language content may lead individuals to disengage from (at least some aspects of) predictive mechanisms in order to “stick to their guns”—and may thus make them less likely to update their mental models in real time.

Conclusions

In the present study, we investigate one variable—domain knowledge—that seems likely to influence the extent to which individuals engage in prediction, as well as the nature of those predictions, during language comprehension. Our findings suggest nuanced relationships between the extent of an individual's existing domain knowledge and the circumstances under which they engage in active prediction, as evidenced by interactive relationships between the comprehender and the overall pervasiveness/difficulty of the language. Moreover, the extent of an individual's domain knowledge may also immediately impact how “immersed” they are in considering domain-specific mental models compared to their ability to rapidly engage more everyday semantic and linguistic associations with the sentence; evidence for this comes from the finding that unexpected-but-plausible sentence continuations comprising everyday, non-domain-specific information are less available when individuals are likely to possess strong, domain-specific knowledge that may override otherwise plausible alternate continuations. We speculate that the ability to generate such rich, detailed mental models may obviate the need for individuals to actively engage in prediction, since they may instead immediately activate the necessary information “for free.” Thus, knowledge seems to be an important variable to consider when assessing the comprehension conditions that engender prediction (and those that do not).