Reversing expectations during discourse comprehension

Ming Xiang; Gina Kuperberg

doi:10.1080/23273798.2014.995679

. Author manuscript; available in PMC: 2016 Jul 1.

Published in final edited form as: Lang Cogn Neurosci. 2014 Nov 24;30(6):648–672. doi: 10.1080/23273798.2014.995679

Reversing expectations during discourse comprehension

Ming Xiang ^a, Gina Kuperberg ^b,^c

PMCID: PMC4405243 NIHMSID: NIHMS645274 PMID: 25914891

Abstract

In two ERP experiments, we asked whether comprehenders used the concessive connective, even so, to predict upcoming events. Participants read coherent and incoherent scenarios, with and without even so, e.g. “Elizabeth had a history exam on Monday. She took the test and aced/failed it. (Even so), she went home and celebrated wildly.”, as they rated coherence (Experiment 1) or simply answered intermittent comprehension questions (Experiment 2). The semantic function of even so was used to reverse real-world knowledge predictions, leading to an attenuated N400 to coherent versus incoherent target words (“celebrated”). Moreover, its pragmatic communicative function enhanced predictive processing, leading to more N400 attenuation to coherent targets in scenarios with than without even so. This benefit however, did not come for free: the detection of failed event predictions triggered a later posterior positivity and/or an anterior negativity effect, and costs of maintaining alternative likelihood relations manifest as a sustained negativity effect on sentence-final words.

Keywords: Discourse processing, concessive connectives, prediction, event structures, N400, P600, late negativity, ERP

General Introduction

Successful language comprehension draws heavily upon our experience in the real world. This real-world knowledge, stored within long-term memory, is recruited by the comprehension system to aid the construction of a discourse model. It tells us whether what we hear or read is plausible, implausible, true or false. Moreover, as language unfolds online, we continually draw upon this stored knowledge to facilitate our comprehension of sentences describing familiar events or states (Marslen-Wilson, Brown and Tyler, 1988; Warren and McConnell 2007; McRae et al. 1997), as well as discourse describing familiar relationships between events and states (Singer et al. 1996; Keenan et al., 1984; van Dijk and Kintsch, 1983).

At the same time, however, language offers us the remarkable ability to construct discourse models that do not necessarily conform to our real-world knowledge. Moreover, there is emerging evidence that such models can sometimes (although not always) facilitate the processing of incoming words during comprehension (Nieuwland and Van Berkum, 2006; Nieuwland and Martin, 2012; Nieuwland, 2013). For simplicity, we will refer to any such discourse model as an ‘alternative world model’— a separate set of events and relations that are established in some mental space that is different from the default real-world knowledge, stored within long-term memory.¹ For example, when reading Harry Potter, we may well expect to encounter events of magic and wizardry that are quite different from our everyday reality. And, if asked to entertain the possibility, “if humans were living on the moon…”, we would not be surprised to hear about facts and events that are quite different from what happens on planet Earth.

Concessive Connectives and “Even so”

Importantly, we do not only construct alternative world models when reading fiction or carrying out counterfactual reasoning. We use such models all the time during everyday communication through our use of small words or phrases, like but, however, even so, and although. These so-called concessive connectives set up an alternative world model by introducing a presupposition (Lakoff, 1971; Lagerwerf, 1998) or conventional implicature (Grice, 1975) that an upcoming proposition will contrast with, or contradict, a previously held assumption or expectation based on world knowledge (Blakemore, 2002; Lakoff, 1971). The rich inherent meaning of these lexical items also provides the comprehender with explicit information about how the upcoming proposition should be linked to its preceding discourse context. In this sense, they function to pragmatically constrain the incremental process of discourse comprehension, helping us to infer the relevance of any upcoming information (Blakemore, 2002; Wilson and Sperber,1993). This pragmatic constraining function sets concessive connectives apart from the fictional scenarios or counterfactuals mentioned above. These also set up alternative world models. However, they do not necessarily provide sufficient linguistic information about exactly how upcoming propositions will link to the discourse context.

Although concessive connectives are often used in everyday communication, there have been very few psycholinguistic studies examining their influences on online comprehension (but see Murray, 1994 and 1997). In the present study, we examine the effects of one particular concessive connective on word-by-word discourse comprehension—even so. Even so is a concessive connective that is commonly used to link propositions across sentence boundaries. It is derived from the scalar term, even, which introduces a presupposition that the event under discussion is very low in its likelihood, but asserts that the utterance is nevertheless true (Karttunen and Peters 1979). At the discourse level, even so inherits the scale-reversing property of even: it functions to establish an alternative world model in which the possible causal relationships between events are reversed from what would follow from our real-world knowledge. In addition, as discussed above, even so, like other concessive connectives, acts to pragmatically constrain the discourse context by ‘narrowing down’ the number of potential relationships to those that causally opposite-to-expected (see Noveck and Spotorno, 2013, for a more general discussion of this type of ‘narrowing down’ effect in communication, and discussed further below).

To illustrate the scale-reversing function of even so, consider the four types of three-sentence scenarios shown in Table 1. In the scenarios where the final sentence does not begin with “Even so” (the ‘plain scenarios’), coherence arises purely from the real-world relationship between the particular event described in the final sentence and the events and states described in the preceding context. For example, in the plain coherent scenario, “…Elizabeth took the test and aced it. She went home and celebrated…”, our real-world knowledge tells us that these events are causally related, whereas in the plain incoherent scenario, “…Elizabeth took the test and failed it. She went home and celebrated …”, it tells us that these events are unlikely to follow on from one another.

Table 1.

Example stimuli and characteristics.

Scenario Type (n=45 per condition)	Example	SSV of critical word^‡	Cloze^*	Constraint^*	Coherence ratings^{^}
1. Coherent	Elizabeth had a history exam on Monday. She took the test and aced it. She went home and celebrated wildly.	0.179 [0.078]	0.42 [0.32]	0.52 [0.26]	4.8 [0.2]
2. Incoherent	Elizabeth had a history exam on Monday. She took the test and failed it. She went home and celebrated wildly.	0.174 [0.079]	0.03 [0.09]	0.40 [0.24]	1.7 [0.4]
3. Even-so Coherent	Elizabeth had a history exam on Monday. She took the test and failed it. Even so, she went home and celebrated wildly.	0.174 [0.079]	0.31 [0.28]	0.44 [0.25]	3.3. [1.0]
4. Even-so Incoherent	Elizabeth had a history exam on Monday. She took the test and aced it. Even so, she went home and celebrated wildly.	0.179 [0.078]	0.04 [0.11]	0.40 [0.24]	2.4 [1.0]

Open in a new tab

Means are shown with standard deviations in square parentheses. The critical word in each of the example sentences is underlined (although this was not the case in the experiment itself).

^‡

LSA was used to calculate Semantic Similarity Values (SSVs) between the critical word and its preceding content words.

Cloze probability and constraint are represented as the proportion of total responses from 40 participants.

^{^}

Coherence ratings, on a 1-5 scale, were collected during the ERP recording session in Experiment 1. 5: very coherent; 1: incoherent.

In the scenarios where the final sentence begins with “Even so” (the ‘even-so scenarios’), these relationships are reversed: coherence is evaluated in relation to the alternative world model set up under even so, rather than real-world knowledge. For example, the even-so coherent scenario, “…Elizabeth took the test and failed it. Even so, she went home and celebrated…”, is coherent, despite the fact that the relationships between the events/states described do not match our long-term real-world knowledge. And the even-so incoherent scenario, “…Elizabeth took the test and aced it. Even so, she went home and celebrated…”, is incoherent, despite the fact that the events described are consistent with our real-world knowledge.

In the present study, we asked participants to read discourse scenarios similar to those described above. Based on a large psycholinguistic literature showing that we are able to integrate multiple linguistic cues quickly and incrementally during online word-by-word language comprehension (e.g. Altmann and Steedman, 1988; MacDonald, Pearlmutter, and Seidenberg, 1994; Tanenhaus, Spivey-Knowlton, Eberhard, and Sedivy, 1995; van Berkum, 2009, Traxler, Bybee and Pickering, 1997), we expected that comprehenders would integrate even so relatively quickly to establish an alternative world model. Our main questions concerned whether, when and how comprehenders would use this alternative world model to predict and process incoming information as it unfolded, word by word.

Prediction, generative models and event-related potentials

The term prediction has been used by different researchers in different ways. While early models assumed that it necessarily entailed committing to specific lexical items (Forster, 1981) in a strategic, all-or-nothing fashion (we either predict or we don’t) (Becker, 1980 Becker, 1985), here we make no such assumptions. We use the term prediction and expectation interchangeably and conceive of prediction as influencing a Bayesian prior — an assessment of the probability of accessing information at a particular representational level ahead of encountering all the linguistic information required to activate, retrieve or compute this representation. We assume that predictions are generated at multiple levels of representation and that rather than being deterministic, they are probabilistic in nature: that is, multiple predictions at a particular representational level are held with differing probabilities that add up to 100% in total. Thus, a very strong prediction corresponds to a near-certain (e.g. 99 %) probability of belief in a particular upcoming representation, and a weak prediction corresponds to many low-probability beliefs in multiple possible upcoming representations.

We view such probabilistic predictions as a consequence of a dynamic hierarchical generative process by which our brains draw upon high-level stored representations and contextual information to construct a generative model that best explains the sensory input we encounter. This type of framework is proving powerful not only for understanding language processing (e.g. Farmer, Brown, & Tanenhaus, 2013; Feldman, Griffiths, & Morgan, 2009; Fine, Jaeger, Farmer, & Qian, 2013; Hale, 2001; Kleinschmidt & Jaeger, In press; Levy, 2008; Norris, 2006; Norris & McQueen, 2008, Kuperberg, 2014), but also many other aspects of perception and cognition (Clark, 2013; Jacobs and Kruschke, 2011; Griffiths et al., 2008; Rao & Ballard, 1999; Friston, 2005). According to this framework, probabilistic predictions are propagated from higher to lower level representational layers, and any residual error between these predictions and the input to each layer (implicit prediction error) serves as the feed-forward signal from lower to higher-level representational layers. This implicit prediction error (or Bayesian surprise) is, in turn, used to update our predictions in an ongoing attempt to refine the generative model and ‘explain away’ the bottom-up input (see Kuperberg, 2014).

In this study, we are primarily concerned with activity at three layers of representation (see Kuperberg, 2013 for discussion): (1) event sequences, which describe our knowledge about the necessary and likely temporal, spatial, causal and other relationships that link multiple events and states together to form sequences of events, sometimes known as scripts, frames or narrative schemas (Fillmore, 2006; Schank & Abelson, 1977; Sitnikova, Holcomb, & Kuperberg, 2008; Zwaan & Radvansky, 1998); (2) event structures, which describe our fine-grained knowledge about specific events (e.g. in an “arresting” event, it is more likely that a policeman arrests a burglar than the other way around, McRae, Ferretti, and Amyote, 1997), our knowledge about similar events (e.g. the similarities between a ‘teaching event’ and an “instructing” or ‘mentoring’ event), as well as our coarser-grained knowledge about the prototypical semantic-thematic roles (Agent, Patient, Experiencer, Stimulus etc) played by participants in actions and states (Jackendoff, 2002); and (3) semantic (or conceptual) features, which describes our knowledge of the perceptual features and functional properties of conceptual entities and categories, e.g. our knowledge that a “boy” has the properties of being <human>, <male>, <young> etc.

These representational layers interface with one another through statistical dependencies that describe the regularities between them, and, during language processing, predictions generated at higher layers influence the priors at lower layers through these dependencies. Thus, at any given time, the situation-level representation of context will interact with our stored knowledge of event sequences, influencing probabilistic predictions about upcoming event structures, which will, in turn influence probabilistic predictions about upcoming semantic features (indeed, this propagation of implicit predictions may sometimes continue down to lower representational layers, including word-form, pre-lexical and perceptual representations).

We conceive of even so as exerting its influence at the event sequence layer of representation by explicitly signaling to the comprehender to expect an opposite-to-expected causal relationship. Under the framework we just described, this prediction will propagate down to constrain the prior probability distribution at the event structure layer, narrowing it down from including many different kinds of event structures (with causal, spatial, temporal and other relationships with the context) that are each held with relatively low probabilities, to a specific type of event structure that is held with higher probability. This, in turn, will lead to strong (high probability) predictions about upcoming semantic features at the representation layer below.

One way of examining how these types of probabilistic predictions interact with incoming information as it unfolds in real time is through event-related potentials (ERPs) — an online neural measure of cognitive processing. It has been proposed that ERPs associated with auditory speech processing are a direct reflection of implicit prediction error within a hierarchical predictive coding system (Friston, 2005; Wacongne, Changeux, & Dehaene, 2012; Wacongne et al., 2011), and recent evidence suggests that this may also be true of semantic processing (Rabovsky & McRae, 2014), with different ERP components reflecting both the representational level of a prediction error (Kuperberg, 2014), as well as the certainty of our original predictions (Kuperberg, 2013, 2014). In this study, we focus on three sets of ERP components: the N400, the posterior late positivity or P600, and the late negativities.

The N400 is a negative-going waveform with a centro-parietal scalp distribution, seen between 300-500ms after word onset. It reflects changes in activity within semantic memory that are induced by incoming stimuli (Kutas and Federmeier, 2011), and it can be formalized as reflecting implicit prediction error at the level of semantic features (Rabovsky & McRae, 2014). Of particular relevance to the present study is evidence that the amplitude of the N400 can be influenced by implicit predictions generated at higher representational levels (e.g. event sequences or event structures). If an incoming word’s coarse-grained (Bornkessel-Schlesewsky & Schlesewsky, 2009; Paczynski & Kuperberg, 2011, 2012) or finer-grained (Ferretti, Kutas, & McRae, 2007; Metusalem et al., 2012; Paczynski & Kuperberg, 2012) semantic features match these implicit predictions, the N400 evoked by that word is attenuated. Moreover, pragmatic communicative cues (for example, commas, e.g. Nieuwland, Ditman, and Kuperberg, 2010, Experiment 2, or speech dysfluencies, e.g. Corley, MacGregor, and Donaldson, 2007) can play an important role in determining whether or not comprehenders are able to fully use a discourse context and stored event knowledge to generate implicit predictions leading to facilitated semantic processing during online comprehension.

Although the N400 is influenced by predictions generated at the event structure layers, and it reflects implicit prediction error at the level of semantic features, it is actually not directly sensitive to prediction errors at the event structure layer itself (see Kuperberg, 2013 for discussion and Federmeier et al, 2007; Lau et al, 2013; Van Petten and Luka, 2012 for consistent evidence). Rather, the costs of disconfirmed predictions at the level of event structure manifest as waveforms that peak after the N400 time window.

The first of these is a posteriorly-distributed late positive-going ERP component — the so-called P600, which peaks between 500-800ms after stimulus onset (Kuperberg, 2007). Although the P600 was originally characterized as the ERP component produced by words that violated the syntactic constraints of predicted events structures (Hagoort, Brown, & Groothusen, 1993; Osterhout & Holcomb, 1992; 1993), it is now clear that it can also be evoked by violations of strong semantic constraints on event structure (see Kuperberg, 2007 for a review). Specifically, it is triggered when the sentential or discourse context encourages a strong near-certain prediction for a particular event structure (a prediction with near-100% probability), and this conflicts with the event structure that is computed by initial attempts to integrate the bottom-up input. It may reflect prolonged attempts to (re)-analyze the context and input to come up with a new discourse model (Kuperberg, 2007, Paczynski & Kuperberg, 2012). In Bayesian terms, it may reflect ‘unexpected surprise’ that triggers a switch to a new generative cause at the level of event sequences that better explains (and allows us to rapidly adapt to) the input (see Courville, Daw, & Touretzky, 2006; Qian, Jaeger, & Aslin, 2012; Yu, 2007; Yu & Dayan, 2005; discussed by Kuperberg, 2013, 2014).

The second set of waveforms that can be seen when event structure predictions are disconfirmed by an input is a group of late negativities, which also peak past the N400 time window and often have a more widespread or frontal distribution than the N400. Unlike the P600, these late negativities are not evoked by violations of a single near-certain event structure prediction. Rather, they are seen when the context constrains for one event structure with medium-high probability and another event structure with lower probability, and the bottom up input leads the less probable event structure to be selected (e.g. Lee and Federmeier, 2006, 2009, Baggio, Lambalgen, and Hagoort, 2008; Bott, 2010; Wlotko and Federmeier, 2012; Paczynski, Jackendoff and Kuperberg, 2014; Wittenberg, Paczynski, Wiese, Jackendoff, and Kuperberg, 2014). In this sense, they correspond to implicit prediction error at the level of event structures (see Kuperberg, 2014).

In the present study, we examined these ERPs to ask four questions about how and when predictions established by even so are used during online discourse processing.

1. Reversed and enhanced semantic expectations under even so?

Our first question was whether, under even so, our implicit predictions about upcoming semantic features, if any, would be based on an alternative world model or based on the stored long-term real-world knowledge. Previous studies using other constructions that set up alternative world models have provided different answers to this question. Sometimes, stored long-term real-world knowledge appears to dominate the initial stages of semantically processing incoming words, as indexed by the N400. For example, in a study of counterfactuals, Fergurson et al. (2008, Experiment 2) asked participants to read sentences like, “If cats were not carnivores, families could feed their cats a bowl of fish/carrots…”. They showed that the real-world consistent (but alternative world inconsistent) word, fish, elicited a smaller N400 than the real-world inconsistent (but alternative world consistent) word, carrots (see also Ferguson et al. 2008, Experiment 1, and Ferguson and Sanford, 2008, for related eye tracking results). At other times, however, an alternative world model can override long-term real-world knowledge to facilitate semantic processing of incoming words (e.g. Nieuwland and Van Berkum, 2006; Nieuwland and Martin, 2012; Nieuwland, 2013; also see Hald, Steenbeek-Planting, and Hagoort 2007, and Ferguson and Breheny, 2011). For example, in another study of counterfactuals, Nieuwland and Martin (2012; also see Nieuwland, 2013) asked participants to read sentences like, “If NASA had not developed its Apollo Project, the first country to land on the moon would have been Russia/America surely”, and showed that the N400 was smaller to Russia than to America.

One factor that seems to be crucial in determining whether we draw upon long-term real-world knowledge or an alternative world model to facilitate subsequent semantic processing is whether the discourse context is pragmatically constraining—that is, whether it provides explicit information about how the upcoming proposition will be linked to it. This is nicely illustrated by the counterfactual experiments described above. In the study by Ferguson et al. (2008, Experiment 2), the preceding discourse context does not constrain for a particular event or state: people will have quite different opinions (or little to say) about the most likely thing that people will feed a non-carnivorous cat. In the studies by Nieuwland and Martin (2012) and Nieuwland et al. (2013), however, the discourse is more constraining: given how well-known the America-Russia space race is, most people will expect the upcoming event to describe the opposite of what actually happened.

Returning to the present study, we hypothesized that the pragmatic constraining function of even so would lead comprehenders to draw upon the alternative world model established and generate strong (high probability) predictions about an upcoming real-world inconsistent event and semantic features consistent with this event, leading to more semantic facilitation of congruous incoming words. Indeed, given this pragmatic constraining function, these predictions might be stronger (higher probability) than those generated in the plain scenarios.

To test these hypotheses, we examined the modulation of the N400 across two contrasts. First, we compared the coherent and incoherent even-so scenarios. If comprehenders reverse their expectations under the scale-reversing function of even so, then we should see a smaller N400 (more semantic facilitation) to critical words like celebrated in the even-so coherent (but real-world inconsistent) scenarios (e.g. “…Elizabeth took the test and failed it. Even so, she went home and celebrated…”). Second, we contrasted the even-so and the plain coherent scenarios. If comprehenders enhance their semantic expectations under the pragmatic constraining function of even so, then we should also see a smaller N400 to celebrated in the even-so coherent scenarios than in the coherent plain scenarios (e.g. “…Elizabeth took the test and aced it. She went home and celebrated…”).

2. Costs of disconfirmed event structure predictions under even so?

Our second question was whether we would see any evidence in the ERP waveform of encountering input that disconfirmed any high probability predictions of upcoming event structures established under even so. As noted above, the costs of violating event structure predictions do not manifest directly on the N400 itself (Federmeier et al, 2007; Lau et al, 2013; Van Petten and Luka, 2012; Kuperberg, 2013), but appear later, on components that peak past the N400 time window — a stage at which the event structure has been fully computed from the bottom-up input.

We considered two possibilities. The first was that even so would lead comprehenders to generate a single strong high certainty prediction for one specific type of event that has a real-world inconsistent causal relationship with the preceding event. On this account, if integration of the bottom-up input yields a real-world consistent causal relationship, the resulting conflict would trigger prolonged attempts to (re)-analyze the context and input to come up with a new discourse model (Kuperberg, 2007 & 2013). This might manifest as a larger posteriorly distributed late positive-going P600 to targets like celebrated in incoherent even-so scenarios than in incoherent plain scenarios.

The second possibility was that, rather than commit with near-100% certainty to a single real-world inconsistent event structure under even so, comprehenders would consider the possibility of encountering a real-world consistent event structure with some lower probability. On this account, integration of the bottom-up input would ultimately select the less probable but real-world consistent structure over the more probable but real-world inconsistent structure, and this selection cost might manifest as a larger late anteriorly-distributed sustained negativity to targets in the incoherent even-so scenarios than in the incoherent plain scenarios (e.g. Baggio, Lambalgen, and Hagoort, 2008; Bott, 2010; Paczynski, Jackendoff & Kuperberg, In press; Wittenberg, Paczynski, Wiese, Jackendoff, & Kuperberg, 2014). Note that the two possibilities outlined above are not mutually exclusive because the certainty of event structure prediction might vary between participants and/or between trials.

3. Wrap-up costs of assessing overall discourse coherence against the alternative world established under even so

Our third question was whether even so would lead to independent costs associated with assessing a discourse model that is coherent under a set of likelihood relations that differ from our default real-world knowledge. These costs might not necessarily be apparent at the point of the critical word itself. However, they might manifest later as a sustained negativity on the sentence-final word—the point at which overall discourse coherence is ‘wrapped up’ and evaluated.

A larger sustained negativity is often seen on the final words of sentences that are implausible (versus plausible) in relation to real-world knowledge, even when the implausibility or anomaly occurs mid-sentence (e.g. Hagoort and Brown, 2000; Hagoort, 2003; Ditman, Holcomb and Kuperberg, 2007; De Grauwe, Swain, Holcomb, Ditman and Kuperberg, 2010). This is presumably because it is harder to assess overall coherence if the overall discourse model mismatches long-term real-world knowledge than if it matches real-world knowledge. If similar wrap-up costs are incurred when the comprehension system evaluates overall coherence against a set of reversed likelihood relationships established under even so (the alternative world model), this would predict a larger sustained negativity effect on sentence-final words in the even-so scenarios than in the plain scenarios, even when both are coherent. Moreover, it should be even harder to come up with a final representation of meaning when overall coherence is evaluated against a set of reversed likelihood relationships, and the scenario turns out to be incoherent, predicting the largest sustained negativity on the final words of the even-so incoherent scenarios.

4. Effects of task

Our final set of questions concerned the effects of task on comprehending both the plain and the even-so scenarios. Task can influence processing in several different ways. It can influence the degree to which comprehenders attend to different aspects of discourse, including the semantic relationships between propositions. This can, in turn, influence the strength/certainty of our predictions, which, as discussed above, can influence the neural mechanisms engaged at multiple stages of processing. To examine the effects of task in this study, we carried out two experiments. In Experiment 1, participants were asked to explicitly rate the coherence of each discourse scenario. In Experiment 2, participants were asked to read the scenarios and to answer intermittent comprehension questions about their content.

Experiment 1

In Experiment 1, participants read the four types of three-sentence scenarios described in Table 1. The factors of Coherence and Even-so were fully crossed, and we measured ERPs on both critical words (e.g. celebrated) as well as on the final words of the third sentence. After each scenario, participants explicitly judged the coherence of the final sentence in relation to the previous context. This encouraged them to pay close attention to the internal semantic relationships between the propositions.

Importantly, we matched general schema-based semantic relationships between the critical word and the ‘bag of words’ in the context across all four conditions using Latent Semantic Analysis (LSA, see Methods). This allowed us to determine whether readers based their expectations on specific types of relationships between events, e.g. the fact that students are likely to celebrate after doing well on exams (Yang et al. 2007; St. George et al. 1997; Kuperberg et al. 2011), or on more general, unstructured word associations based around a particular schema (Schank and Abelson, 1977), e.g. the general association between successful/failed exams and parties afterwards (see Otten and Van Berkum, 2007; Paczynski and Kuperberg, 2012 for discussion).²

Our starting point was the plain scenarios. We asked whether, with these task instructions, readers would generate predictions about likely upcoming events/states, leading to facilitated semantic processing of incoming words whose semantic features were associated with these events. Based on a previous study in which we contrasted causally coherent and incoherent plain discourse scenarios while participants carried out a similar explicit coherence judgment task. (Kuperberg et al., 2011), we hypothesized that we would see a smaller N400 on critical words in the plain coherent than in the plain incoherent scenarios.

Having established how the plain scenarios were processed, we then turned to the effects of even so. We make the following hypotheses based on the discussion in the General Introduction: (1) If comprehenders are able to draw upon the alternative world model previously generated under even so, and use this to reverse their predictions about the real-world likelihood of upcoming events, the N400 should be smaller to critical words in the coherent than the incoherent even-so scenarios, just as in the plain scenarios; indeed, if predictions established under even so are stronger (higher probability) than in the plain scenarios, then the N400 to critical words in the coherent even-so scenarios should be even smaller than than in the coherent plain scenarios. (2) If even so leads to strong reverse predictions about upcoming events, then disconfirmation of these event predictions by the bottom-up input should lead to prolonged neural costs (past the N400 time window) to critical words in the even-so incoherent versus the plain incoherent scenarios. Finally, (3) if, during sentence-final wrap-up, evaluating coherence against the temporary alternative world model established under even so incurs more costs than evaluating coherence against long-term real-world knowledge, then sentence-final words in the even-so coherent scenarios should produce a larger negativity than sentence-final words in the plain coherent scenarios; moreover, this negativity should be still larger on sentence-final words in the incoherent even-so scenarios.

Methods

Construction and characterization of stimuli

One-hundred-and-eighty sets of two-sentence scenarios were constructed, each with four conditions (45 scenarios per condition), see Table 1. In all scenarios, a critical word appeared in the final sentence but before the sentence-final word. The number of words between the critical word and the sentence-final word varied between 0 and 3 across trials, but was matched between conditions within a scenario. The number of words between the critical word and the sentence initial word (excluding “Even so”) mostly varied between 1 to 4 words (with only a few scenarios with more than 4 words), but once again this was matched between conditions within a scenario.

The four conditions were constructed by crossing two factors: Even-so (the presence or absence of the phrase, “Even so” at the beginning of the final sentence: plain or even-so scenarios) and Coherence (the coherence of the critical word in relation to its preceding context: coherent or incoherent scenarios). In the plain coherent and the plain incoherent conditions, the final sentence was identical: differences in coherence arose because of differences in the first two sentences. In the even-so scenarios, coherence was reversed: the original plain coherent scenarios became the even-so incoherent scenarios, and the original plain incoherent scenarios became the even-so coherent scenarios. The critical word was thus identical in all four conditions.

Lexical predictability of critical words and lexical constraint of discourse contexts: Offline cloze norming

The 180 sets of scenarios were counterbalanced across four lists using a Latin Square design. For each scenario, the critical word and all words following it were removed and replaced by an ellipsis, e.g. “Elizabeth had a history exam on Mon. She took the test and aced it. She went home and ….”. Cloze ratings of these stems were conducted as an online survey using SurveyMonkey.com, with participants recruited from Tufts University and other neighboring areas. Participants were asked to read the scenario stems and to complete the unfinished last sentence by writing down the most likely ending. Initially, 40 native English speakers (30 female, 10 male, average age: 23.3) participated (ten per list). Cloze probabilities for each of the four scenario types were calculated based on the percentage of respondents who produced a word that matched the critical word exactly. Based on the initial results, we modified 13 of the scenarios that didn’t show any difference between coherent and incoherent scenarios, and carried out a second cloze study on these 13 new items with another set of 40 native English speakers (29 female, 11 male, average age =26.2). We used these 13 new items to replace the old 13 items, and then recalculated the cloze probability for each item across the entire stimulus set.

Cloze probabilities for each scenario type are given in Table 1. A 2 × 2 repeated measures ANOVA with Coherence and Even-so as within-items factors revealed a main effect of Coherence (F(1,179)=329, p < .001) and a main effect of Even-so (F(1,179)=22, p<.001). There was also an interaction between the two factors (F(1, 179)=21, p<.001). Follow-up paired t-tests examining effects of Coherence at each level of Even-so showed that, as expected, the even-so and the plain coherent scenarios had significantly higher cloze probabilities than their corresponding incoherent conditions (plain coherent vs. plain incoherent: t(179)=16, p<0.001; even-so coherent vs. even-so incoherent: t(179)=13, p<0.001), but that the difference in cloze probability between the coherent and incoherent conditions was larger in the plain than in the even-so scenarios. Follow-up t-tests examining effects of Even-so at each level of Coherence showed that the cloze probability of critical words in the plain coherent scenarios was significantly higher than in the even-so coherent scenarios (t(179)=5, p<0.001), but that there was no significant difference in cloze probability between the plain and even-so incoherent scenarios (t(179)=0.16, p>0.8).

In addition to calculating cloze probabilities, we also calculated the contextual lexical constraint for each type of scenario context by finding the most common completion across participants who saw that context, regardless of whether or not it matched the critical word, and tallying the number of subjects who provided this completion, see Table 1. For example if, for a given scenario, our designed critical word was “disappointed” and 3 out of 10 people provided “disappointed” as their answer, then the cloze probability would be 0.3 for this scenario, but if 5 out of the 10 people provided “confused” as their answer, then the lexical constraint probability for this context would be 0.5. A 2 × 2 repeated measures ANOVA with Coherence and Even-so as within-items factors again revealed a main effect of Coherence (F(1, 179)=22, p < .001), a main effect of Even-so (F(1, 179)=5.7, p<.05), and an interaction between the two (F(1, 179)=7.4, p<.01). Follow-up paired t-tests showed that the lexical constraint of the plain coherent contexts was greater than the plain incoherent contexts (t(179)=5, p<0.001), while the difference between the even-so coherent and incoherent contexts was only marginally significant (t(179)=1.9, p<0.06). In addition, the lexical constraint of the plain coherent contexts was greater than the even-so coherent scenario contexts (t(179)=3.6, p<0.001), but there was no difference in lexical constraint between the plain and even-so incoherent scenario stems (t(179)=0.76, p>0.9).

Latent Semantic Analysis

Latent Semantic Analysis (LSA, a measure of semantic relatedness, Landauer and Dumais 1997; Landauer et al. 1998) was carried out on the final stimulus set, on a term-to-term basis, to examine the Semantic Similarity Values (SSVs) between the critical word and previous content words across the context of each scenario. A paired t-test revealed no significant differences between the plain coherent and plain incoherent scenarios (t(179)=1.55, p > 0.10). Note that, because of how the stimuli were constructed, SSVs were the same for the plain coherent and even-so incoherent conditions, and for the plain incoherent and even-so coherent conditions. See Table 1 for SSVs in all four scenario types.

Set-up of lists for the ERP experiment

The final set of experimental scenarios was divided into four lists, counterbalanced using a Latin Square design. During the ERP experiment, each participant viewed only one list and therefore only one condition of each scenario, but across all participants, each scenario and critical word was seen in all four conditions. Each list had 180 scenarios, 45 from each condition. The order of items was randomized within each list separately.

Participants in the ERP experiment

Twenty-nine native English speakers initially participated in the ERP study (two participants were subsequently excluded for extensive ocular and muscular artifacts, see below). All participants were undergraduate students recruited from local universities. They were all right-handed, as assessed using the Edinburgh handedness inventory (Oldfield 1971), with normal or corrected-to-normal vision, and no history of neurological disorders. Participants were paid for their participation and gave full consent according to the guidelines of the Tufts University Human Subjects Committee. The 27 subjects included in the data analysis had an average age of 20 years (SD: 1.7) and 14 were males.

Stimulus presentation

Participants sat in a quiet and dimly-lit room, separated from the experimenter and control computers. Their task was to rate each scenario on a 1 to 5 scale, based on how naturally the third sentence followed on from the previous two sentences. For half of the participants, a score of 1 meant “it does not follow at all” and 5 meant “it follows very naturally”; and for the other half, the scoring was reversed for counterbalancing purpose. Before starting the experiment, each participant read twelve practice scenarios to ensure that they understood the task.

Stimuli were presented on a computer monitor, in white font, centered on a black background Participants were randomly assigned to one of the four lists. Each trial began with the word “READY” on the screen, which cued the participant to press a button to begin reading the three-sentence scenario. The first two context sentences were presented one after another as whole sentences. Participants read these two sentences at their own pace, pressing a button to move on to the second sentence. They then saw a fixation cross (“+”) in the middle of the screen for 500ms, followed by a blank screen for 100ms, and then the last sentence was presented word by word. Each word was centered in the middle of the screen, and was presented for 350ms, followed by an interstimulus interval (ISI) of 150ms. In the even-so scenarios, the phrase “Even so” was presented as a whole. Because it consists of two words, it appeared on the screen for 400ms followed by an ISI of 150ms. The last word of the final sentence appeared with a period and was presented for 800ms. A 400ms ISI followed this final word, and then a “?” appeared on the screen which cued participants to make their rating responses. Participants indicated their responses by pressing one of the five buttons on the response pad.

ERP recording

The EEG response was recorded from 29 electrodes (Electro-Cap International, Inc., Eaton, OH; see Figure 1 for montage). Additional electrodes were placed below the left eye and at the outer canthus of the right eye to monitor vertical and horizontal eye movements. There were also two mastoid electrodes (A1, A2) and the EEG signal was referenced to the left mastoid online. The EEG signal was amplified by an Isolated Biometric Amplifier (SA Instrumentation Co., San Diego, California) with a band pass of 0.01-40 Hz. It was continuously sampled at 200Hz and the impedance was kept below 5kOhm.

Electrode montage, showing each 3-electrode region used for analysis. Regions in dark grey were part of the mid-regions omnibus ANOVA and regions in light grey were part of the peripheral regions omnibus ANOVA.

ERP analysis

Trials contaminated with eye artifact (with max-min amplitudes exceeding 70μv, as well as visual inspection) or amplifier blockage were excluded from analyses. After artifact rejection, averaged ERPs, time-locked to critical words, were obtained by calculating the mean amplitude (relative to a 100ms pre-stimulus baseline). At the critical word, we carried out analyses across two time windows. To capture the peak of the N400 in all four conditions, we used a 350-450ms time window. To capture the late positivity/P600 in all four conditions, and to avoid component overlap with the earlier N400 effect, we used a 600-800ms time window. Examination of the waveforms also revealed a late, sustained anteriorly-distributed negativity effect, which we captured between 800-1000ms. At the sentence-final word, visual inspection of the waveform revealed a prolonged negativity, which was captured with a 300-1000ms time window.

We initially carried out two omnibus repeated-measures ANOVAs in which the scalp was subdivided into several 3-electrode regions along its anterior–posterior distribution, at both mid and peripheral sites (each region contained three electrode sites, see Figure 1). In the mid-regions omnibus ANOVA, the within-subject variables were Coherence (2 levels: coherent, incoherent), Even-so (2 levels: plain, even-so), and Region (5 levels: prefrontal, frontal, central, parietal, occipital). In the peripheral regions omnibus ANOVA, the within-subjects variables were Coherence (2 levels: coherent, incoherent), Even-so (2 levels: plain, even-so), Region (2 levels: frontal, parietal) and Hemisphere (2 levels: left, right). For further follow-ups, we focused on the subgroup of regions that showed most modulation across conditions (smallest p values; largest F values in omnibus ANOVAs analyses carried out in each region).

In each subgroup of regions, we carried out 2 (Coherence) × 2 (Even-so) × Region ANOVAs; any interactions between Coherence and Even-so were followed up by (a) by examining the effects of Coherence at each level of Even-so, and (b) by examining the effects of Even-so at each level of Coherence. In all analyses, the Greenhouse and Geisser (1959) correction was applied to repeated measures with more than one degree of freedom, and a significance level of alpha = .05 was used for all comparisons.

Results

Behavioral results

The coherence ratings for each of the four scenario types are given in Table 1. A 2 × 2 ANOVA confirmed a significant main effect of Coherence (F(1, 26)=131, p<.001). It also revealed a main effect of Even-so, reflecting higher overall coherence ratings in the plain than the even-so scenarios (F(1, 26)=33, p<.001). In addition, there was a significant interaction between these two variables (F(1, 26)=33, p<.001). Planned follow-up comparisons examining the effects of Coherence on the plain and even-so scenarios separately indicated that, as expected, the coherent scenarios were always rated as significantly more coherent than the incoherent scenarios (plain: coherent vs. incoherent: t(26) = 32, p < .001; even-so: coherent vs. incoherent: t(26) = 2.45, p < .05), although the difference was larger in the plain than in the even-so scenarios. Follow-ups examining the effects of Even-so in the coherent and incoherent scenarios separately showed that the coherent plain scenarios were rated as more coherent than the coherent even-so scenarios (t(26)=7.6, p<.001), and the incoherent plain scenarios were rated as more incoherent than the incoherent even-so scenarios (t(26)=-2.9, p<.01). In other words, the plain coherent and plain incoherent scenarios were rated as more coherent and incoherent respectively than their corresponding even-so scenarios, which received ratings that were in between these two extremes.

Event related potentials

Critical Word

At the critical word, 20% of trials were rejected for artifact (plain coherent: 19%; plain incoherent: 19%; even-so coherent 20%; even-so incoherent: 20%). A 2 × 2 within-subjects ANOVA showed that the rejection rate did not differ between the coherent and incoherent scenarios (no main effect of Coherence F(1,26)<.1, p>.9), or between the even-so and plain scenarios (no effect of Even-so F(1,26)=.9, p>.3). There was also no interaction between these two factors (F(1,26)<.1, p>.9).

N400: 350-450ms

There was a main effect of Coherence, reflecting a widespread N400 effect across the even-so and plain scenarios (mid-regions: F(1,26)=40.8, p<.001; peripheral regions: F(1,26)=36.7, p<.001), which was largest in frontal, central and parietal mid-regions (interaction between Coherence and Region in the mid-regions analysis, F(4, 104)=4.3, p<.01, with follow-ups showing the effect of Coherence in each of these three regions, Fs>23, p<.001). To determine how effects of Coherence were modulated by Even-so, we collapsed across these three frontal, central and parietal mid-regions (9 electrode sites in total).

In this 9-electrode region, Coherence was modulated by Even-so (Coherence × Even-so interaction: F(1,26)=5.4, p<.05). There were significant N400 effects of Coherence in both the plain scenarios (F(1,26)=7, p<.05) and the even-so scenarios (F(1, 26)=40, p<.001). However, the magnitude of the N400 effect in the even-so scenarios was larger than in the plain scenarios, see Figure 2B. This larger N400 Coherence effect was driven by a smaller N400 to coherent critical words in the even-so than the plain coherent scenarios (F(1, 26)=5.3, p<.05),³ see Figure 3A (note that voltage map in Figure 3A shows a positivity between 350-450ms because the plain condition was subtracted from the even-so condition). In contrast, there was no difference in the N400 evoked by incoherent critical words in the plain and even-so incoherent scenarios (F(1, 26)<1, p>.4), see Figure 3B.

Grand-averaged waveforms to critical words in Experiment 1 (coherence rating task) showing effects of Coherence at electrodes Fz, Cz and Pz.

**Panel A**: waveforms to critical words in plain coherent (black solid) and plain incoherent (red dotted) scenarios.

**Panel B**: waveforms to critical words in even-so coherent (black solid) and even-so incoherent (red dotted) scenarios.

Voltage maps show differences in ERPs between incoherent and coherent critical words (incoherent *minus* coherent) between 350-450ms (N400) and 600-800ms (P600).

Grand-averaged waveforms to critical words in Experiment 1 (coherence rating task) showing effects of Even-so at electrodes Fz, Cz and Pz.

**Panel A**: waveforms to critical words in coherent plain (black solid) and coherent even-so (blue dotted) scenarios.

**Panel B**: waveforms to critical words in incoherent plain (black solid) and incoherent even-so (blue dotted) scenarios.

Voltage maps show differences in ERPs between even-so and plain critical words (even-so *minus* plain) between 350-450ms and 600-800ms.

Late Posterior Positivity/P600: 600-800ms

Collapsed across the even-so and plain scenarios, there was a P600 effect over parietal (left, right and mid) and mid-occipital regions (interactions between Coherence and Region in the mid-regions analysis, F(4, 104)=9.5, p<.001, and in the peripheral regions analysis, F(3,78)=14.2, p<.001, with follow-ups showing the effect of Coherence in each of these four regions, all Fs > 6.8, ps < 0.05). To determine how the effect of Coherence was modulated by Even-so, we collapsed across these four regions: left, mid and right parietal and mid-occipital regions (12 electrode sites in total).

In this 12-electrode parietal-occipital region, there was a three-way interaction between Coherence, Even-so and Region (F(2, 52)=4.9, p<.05). Follow-ups showed a P600 effect of Coherence in both the plain scenarios (at left and right parietal regions, F(1,26)s>4.7, ps<.05, Figure 2A), and in the even-so scenarios (at the right parietal region and the mid-occipital region, F(1,26)s>6.4, ps<.05, Figure 2B). Once again, the effect was larger in the even-so scenarios than in the plain scenarios. This time, however, the larger effect in the even-so scenarios was driven by a larger late positivity to incoherent even-so than incoherent plain critical words (Figure 3B, in the occipital region, F(1,26)=8.2, p<.01, and in the right parietal region, F(1,26)=4.3, p<.05); there was no significant difference in the late positivity evoked by coherent critical words in the even-so and plain coherent scenarios in any of these regions (all ps>.05, Figure 3A).

Late Anterior Negativity: 800-1000ms

Analysis within this time window revealed an Even-so × Coherence × Region interaction in the mid-regions analysis (F(4, 104)=16.6, p<.001). Some of this effect was driven by the late positivity effect continuing into the late time window at the posterior-parietal site. But in the prefrontal region (electrode sites FP1, FP2, FPz), there was a larger negativity to critical words in the even-so incoherent scenarios than in the other three scenarios (all ts(26)>2.6, ps<.05, see Figure 2 and 3).

Sentence-final word

At the sentence-final word, 28% of trials were rejected for artifact (plain coherent 24%; plain incoherent: 27%; even-so coherent: 29%; and even-so incoherent: 32%). There was a near-significant effect of Coherence (F(1, 26)=4.0, p=.056) because there were slightly more rejected incoherent than coherent trials, and an effect of Even-so (F(1,26)=9.8, p<.01) because there were slightly more rejected even-so than plain trials, but there was no interaction between the two factors.

Sentence-final negativity: 300-1000ms

Activity on the sentence-final words was captured over a prolonged 300-1000ms time window, shown in Figure 4. As expected, there was a larger sustained negativity on sentence-final words in the incoherent than the coherent scenarios (main effects of Coherence: mid-regions, F(1,26)=24, p<.001; peripheral regions, F(1,26)=15, p<.01 analyses). In addition, there was a main effect of Even-so, with a larger negativity on sentence-final words in the even-so than the plain scenarios (mid-regions: F(1,26)=24, p<.001; peripheral regions: F(1,26)=15, p<.01). Finally, in the mid-regions analysis, there was an interaction between Even-so, Coherence and Region (F(4, 104)=3.4, p<.05). To follow-up this three-way interaction, we carried out pair-wise comparisons at two 6-electrode regions: posterior (combining the parietal and occipital 3-electrode regions) and frontal (combining the frontal and central 3-electrode regions).

Grand-averaged waveforms to sentence-final words in Experiment 1 (coherence rating task). ERPs to sentence-final words in all four conditions are shown at electrodes Fz, Cz and Pz. Plain coherent: black solid line; Plain incoherent: blue dashed line; Even-so coherent: green solid line; Even-so incoherent: red dotted line.

In the posterior 6-electrode region, the final words of the even-so incoherent scenarios evoked the largest negativity (differing significantly from the three other conditions, Fs > 13, ps < .01), while the final words of the plain coherent scenarios evoked the smallest negativity. The final words of the plain incoherent and even-so coherent scenarios each evoked medium-sized sustained negativities, which were smaller than in the even-so incoherent scenarios (Fs > 13, ps < .01), but larger than in the plain coherent scenarios (Fs > 4, ps < .05). In the frontal 6-electrode region, however, there was no difference in the amplitude of the negativities evoked by sentence-final words of the even-so coherent, even-so incoherent and plain incoherent scenarios (Fs<3, ps>.05), which all evoked a more negative waveform than the sentence-final words of the coherent plain scenarios (Fs > 6, ps < .05).

Discussion

In this experiment, participants made explicit judgments about the coherence of each scenario. In both the plain and the even-so scenarios, we saw both N400 and P600 effects of coherence on the critical words. In the even-so scenarios, however, both these effects were larger than in the plain scenarios. In addition, in the even-so scenarios, we also saw an effect of coherence on a late anterior negativity between 800-1000ms. At the sentence-final word, we saw effects of both Coherence and Even-so. Finally, the pattern of these ERP findings across the four conditions dissociated from participants’ offline behavioral coherence judgments of the same sentences. We discuss each of these findings in turn.

The plain scenarios

The significant N400 effect of Coherence in the plain scenarios is consistent with our previous findings (Kuperberg et al., 2011), in which we contrasted causally coherent and incoherent plain discourse scenarios while participants carried out a similar explicit coherence judgment task. Just as in this previous study, the N400 in this study was attenuated on the coherent critical words, even though its general schema-based semantic relatedness with the preceding context (as operationalized by LSA) was matched with the incoherent scenarios (see Table 1). This tells us that, when participants actively attend to discourse relationships, they are able to construct a full discourse model from the context, use this model to access stored information about event relationships, and generate predictions about likely upcoming events and their associated semantic features, leading to facilitated processing of incoming words whose semantic features match these predictions.

In the present study, the N400 effect was followed by a small P600 effect, which we did not see in our previous study (Kuperberg et al., 2011). The reason for this is unclear, but one possibility is that the presence of the even-so scenarios themselves encouraged comprehenders to engage in additional analysis in attempts to make sense of all incoherent sentences (see footnote 6).

The even-so scenarios

The N400 attenuation to critical words in the coherent versus incoherent even-so scenarios indicates that readers were able to draw upon the alternative world model set up under even so rather than stored long-term real-world knowledge, to facilitate semantic processing of upcoming words. Moreover, our finding that the N400 Coherence effect was actually larger in the even-so than the plain scenarios, and that this was due to an attenuation of the N400 to the even-so coherent (versus plain coherent) critical words, indicates that the event structure and semantic predictions, set up under even so, were stronger (higher probability) than those generated in the plain scenarios on the basis of default world knowledge. We argue that this is because even so narrowed down the number of potential upcoming event structures to those that were causally real-world inconsistent, unlike in the plain scenarios where they may have considered multiple possible upcoming event structures (based on causal, spatial and temporal relationships with the preceding context), each with lower probability.

The larger P600 on critical words in the incoherent even-so than in the incoherent plain scenarios suggests that, when these high probability predictions were violated, prolonged neural costs were incurred. More specifically, we argue that, under even so, at least on some trials, comprehenders predicted a real-world inconsistent event with high certainty (near 100% probability), and that the P600 was triggered by conflict between this event structure and the structure produced by initial attempts to integrate incoming critical words. We suggest that the P600 itself reflected prolonged attempts to integrate the context and critical words to construct a new discourse model (see Kuperberg, 2007; Paczynski & Kuperberg, 2012; Kuperberg, 2013).⁴

In addition to producing the late posteriorly distributed positivity/P600 effect, we also saw evidence of a late sustained anteriorly distributed negativity effect on critical words in the incoherent even-so (versus the incoherent plain) scenarios. We suggest that, at least on some even-so trials, comprehenders kept alive the possibilities of both a real-world inconsistent and a real-world consistent upcoming event structure. That is, the late sustained anterior negativity effect reflected the maintenance of these alternative event structures and the process of selecting the (less probable) real-world consistent event structure as the bottom-up information from the critical word was integrated (see Baggio, van Lambalgen, and Hagoort, 2008; Paczynski, Jackendoff & Kuperberg, 2014; Wittenberg, Paczynski, Wiese, Jackendoff, & Kuperberg, 2014; and Lee and Federmeier, 2009; for similar interpretations of late sustained negativities in other situations).

The sentence-final word

Our findings at the sentence-final word suggest that the costs of generating an alternative world model under even so did not come for free: in addition to the prolonged sentence-final negativity effect in the incoherent (versus coherent) plain scenarios, we also saw a larger sentence-final negativity effect in the even-so coherent scenarios than in the plain coherent scenarios. Moreover, at posterior sites, the effects of Coherence and Even-so appeared to be additive: the sentence-final negativity effect was maximal in the even-so incoherent scenarios, indicating that still more wrap-up costs were incurred when the integration of the sentence-final word mismatched the alternative world model that had been originally anticipated under even so (at frontal sites, the negativities evoked by the sentence-final words in the even-so coherent, even-so incoherent and plain incoherent scenarios were all of the same magnitude).

Dissociation between online ERP findings and offline behavioral data

The pattern of ERP findings at the critical word dissociated from the pattern of offline discourse coherence ratings across the four conditions: despite clear semantic facilitation (a smaller N400) on critical words in the even-so versus the plain coherent scenarios, participants rated the same even-so coherent scenarios as less coherent than the plain coherent scenarios (3.3 vs. 4.8, Table 1). Moreover, despite prolonged neural costs associated with processing the even-so (versus plain) incoherent critical words, participants rated the same even-so incoherent scenarios as less incoherent than the plain incoherent scenarios (2.4 vs. 1.7, Table 1).

We think that these dissociations arose because participants’ offline retrospective coherence judgments about the coherence of the even-so scenarios were influenced by their long-term real-world knowledge. For example, the even-so incoherent scenario, “Elizabeth took the test and aced it. Even so, she went home and celebrated” sounds quite odd, but the relationship between the events themselves matches real-world knowledge, and this may have led participants to rate it as less incoherent than the incoherent plain scenarios. Similarly, although the even-so coherent scenario, “Elizabeth took the test and failed it. Even so, she went home and celebrated” sounds quite natural, the real-world relationship between the two main events mismatches our real-world knowledge; given time to think about its coherence, this mismatch may have led participants to rate it as less coherent than the plain coherent scenarios.

We also suggest that similar factors influenced the cloze ratings of the even-so scenarios, which also dissociated from the online ERP data. For example, when asked to produce a specific word after the context, “Alice was walking home from work at night. A stranger was following her. Even so, she felt…”, although 5 out of the 10 participants that were given this particular item produced the word “safe”, one participant actually produced “uncomfortable”, suggesting that, during these offline cloze judgments, some participants might have ignored “even so” altogether and continued the sentence on the basis of the default real-world event relations.

Another factor to consider is that, on a substantial subset of the items, although participants clearly had committed to a continuation that was semantically inconsistent with the default real-world knowledge, they did not necessarily converge on exactly the same lexical item. For example, for the scenario, “John tried out for the comedy troupe. His lines were all mumbled and unintelligible. Even so, the director asked him to…”, the target critical word was “join”, but some participants produced other coherent options, such as “stay”.

Together, these findings underline two important points. First, online neural activity does not always mirror participants’ offline coherence judgments and cloze ratings: our ERP results suggest that, with limited time before the next word appeared, online semantic facilitation of the mid-sentence critical words, as reflected by the amplitude of the N400, was driven by the alternative world model established under even so, whereas our offline ratings, and, as discussed above, neural activity at the sentence-final word, were partially driven by long-term real-world knowledge. Second, they show that enhanced prediction for a particular type of event doesn’t necessarily imply prediction or commitment to one specific lexical item. We will return to the relationships between cloze ratings, lexical prediction, event structure prediction, semantic facilitation and N400 modulation in the General Discussion.

Experiment 2

In this second experiment, we presented a new set of participants with the same experimental stimuli. This time, however, participants were asked simply to read the scenarios and answer intermittent, randomly-dispersed comprehension questions about these scenarios.

Once again, our starting point was the plain scenarios. In Experiment 1, and in our previous study examining the use of real-world causal knowledge across sentence boundaries (Kuperberg et al. 2011), we showed that, when explicitly asked to judge coherence, readers are able to retrieve and use quite specific knowledge about likely relationships between real-world events to predict upcoming events and facilitate semantic processing of incoming words consistent with these events, as reflected by an attenuation of the N400. This, however, does not imply that we always use event knowledge to facilitate semantic processing of incoming words during online comprehension; there are times in which semantic processing can be driven primarily by more general stored knowledge about unstructured schema-based semantic relationships between words or concepts (see Otten & Van Berkum, 2007; Kuperberg et al. 2011; Paczynski & Kuperberg, 2012; and Lau, Holcomb & Kuperberg, 2013 for discussion). Whether we use context to go beyond this more general unstructured schema-based knowledge depends on whether we are able to establish a deep discourse representation of the context and use this to retrieve and predict stored upcoming event representations before the semantic features of the incoming word becomes available. This, in turn, depends on many factors (see General Discussion). One of these factors is task demands.

Thus, in Experiment 2, our first question was whether, in the plain scenarios, without an explicit task requirement to judge coherence, readers would still be able to construct a deep discourse model to predict upcoming events and facilitate semantic processing of incoming critical words. Alternatively, semantic processing of critical words might simply be driven by an interaction between whatever contextual representation had been constructed at the time the bottom-up semantic input became available and more general, unstructured schema-based stored semantic relationships. As LSA was matched between the coherent and incoherent plain scenarios, this would predict no difference between the N400 produced by critical words in the two conditions.

Having established how the plain scenarios were processed with a comprehension task, we then asked about the effect of even so. Our question here was whether readers would still be able to reverse and enhance their semantic expectations, as they did in Experiment 1. If so, then we should see similar semantic facilitatory effects on the N400 to those seen in Experiment 1: a smaller N400 on critical words in even-so coherent versus even-so incoherent scenarios, and a smaller N400 on critical words in even-so coherent versus plain coherent scenarios, despite these being matched on schema-based lexical relationships. We might also see evidence of costs of disconfirmed event structure predictions, manifested as prolonged ERP effects past the N400 time window. Finally, we asked whether, in the absence of an explicit coherence judgment, we would still see a sentence-final wrap cost associated with even so.