Language, Cognition and Neuroscience

ISSN: (Print) (Online) Journal homepage: www.tandfonline.com/journals/plcp21

Better together: integrating multivariate with

univariate methods, and MEG with EEG to study
language comprehension

Lin Wang & Gina R. Kuperberg

To cite this article: Lin Wang & Gina R. Kuperberg (12 Jun 2023): Better together: integrating
multivariate with univariate methods, and MEG with EEG to study language comprehension,
Language, Cognition and Neuroscience, DOI: 10.1080/23273798.2023.2223783

To link to this article: https://doi.org/10.1080/23273798.2023.2223783

Published online: 12 Jun 2023.

Submit your article to this journal

Article views: 561

View related articles

View Crossmark data

Citing articles: 1 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=plcp21
LANGUAGE, COGNITION AND NEUROSCIENCE
https://doi.org/10.1080/23273798.2023.2223783

Better together: integrating multivariate with univariate methods, and MEG with
EEG to study language comprehension
Lin Wanga,b and Gina R. Kuperberga,b
a
Department of Psychiatry and the Athinoula A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical
School, Boston, MA, USA; bDepartment of Psychology, Tufts University, Medford, MA, USA

ABSTRACT ARTICLE HISTORY

We used MEG and EEG to examine the effects of Plausibility (anomalous vs. plausible) and Animacy Received 10 November 2022
(animate vs. inanimate) on activity to incoming words during language comprehension. We Accepted 5 June 2023
conducted univariate event-related and multivariate spatial similarity analyses on both datasets.
KEYWORDS
The univariate and multivariate results converged in their time course and sensitivity to ERP; representational
Plausibility. However, only the spatial similarity analyses detected effects of Animacy. The MEG similarity analysis; Animacy;
and EEG findings largely converged between 300–500 ms, but diverged in their univariate and predictive coding; N400;
multivariate responses to anomalies between 600–1000 ms. We interpret the full set of results P600; evidence accumulation
within a predictive coding framework. In addition to the theoretical significance, we discuss the
methodological implications of the convergence and divergence between the univariate and
multivariate results, as well as between the MEG and EEG results. We argue that a deeper
understanding of language processing can be achieved by integrating different analysis
approaches and techniques.

General introduction In a typical EEG/MEG experiment, we measure activity

To comprehend language, we must transform a continu- at multiple recording channels at each time point follow-
ous stream of sounds or letters into meaning. How do ing the onset of an “event” of interest, such as a critical
we accomplish this feat? How and when does the word in a sentence. Our goal is to understand how our
brain encode the semantic features of each incoming experimental variables influence this recorded activity
word during real-time comprehension? Does the so that we can make inferences about when and how
process of retrieving these semantic features always the brain extracts information from the linguistic input
produce an overall increase in brain activity? The exqui- as it unfolds in real time. To achieve this goal, we can
site temporal resolution of electroencephalography take one of two different types of analytic approaches.
(EEG) and magnetoencephalography (MEG) make these The first is to carry out a classic event-related analysis.
neuroimaging techniques ideally suited for addressing This approach has a long history in psycholinguistic
these questions. As information encoded within each research, dating from the discovery of the N400 in
new linguistic input is passed up and down the cortical 1980 (Kutas & Hillyard, 1980). The basic assumption of
hierarchy, large numbers of spatially aligned pyramidal an event-related analysis is that an experimental variable
neurons produce electric dipoles, generating electric of interest modulates the amplitude of activity measured
and magnetic fields (Buzsáki et al., 2012) that can be at individual adjacent channels. This assumption is natu-
detected at the head surface using EEG and MEG. At rally incorporated in a univariate General Linear Model,
each recording channel, the measured voltage or mag- which asks whether and when a particular variable
netic field reflects the sum of neural activity generated explains variance in the amplitude of a dependent vari-
by multiple underlying dipole sources, scaled by a able across multiple items (Baayen et al., 2008; Clark,
weighting factor that, in EEG, is influenced by the con- 1973). If the experimental variable is categorical, then
ductivities of head tissue, particularly the skull and the intercept term in this type of model simply corre-
scalp (Nunez, 1990), and, in MEG, is influenced by the sponds to the average amplitude of the response
folding pattern of the cortex (Hämäläinen et al., 1993). across all trials in the reference condition, and the

CONTACT Lin Wang [email protected] Department of Psychiatry and the Athinoula A. Martinos Center for Biomedical Imaging, Massa-
chusetts General Hospital, Harvard Medical School, Boston, MA 02129, USA; Gina R. Kuperberg [email protected] Department of Psy-
chology, Tufts University, Medford, MA 02155, USA
Supplemental data for this article can be accessed http://doi.org/10.1080/23273798.2023.2223783.
© 2023 Informa UK Limited, trading as Taylor & Francis Group
2 L. WANG AND G. R. KUPERBERG

model’s beta weight corresponds to the mean difference which focused on the magnitude of activity at individual
between experimental conditions (Smith & Kutas, 2015). voxels, these spatial patterns better reflected the “rep-
Indeed, the classic way of carrying out an event-related resentational information” encoded within these
analysis in EEG and MEG is simply to average the regions (Kriegeskorte & Bandettini, 2007).
measured activity at each recording channel, across all Initially, it was assumed that multivariate methods
trials within each experimental condition at each time were unsuited for analysing EEG/MEG data at the head
point following the onset of the critical event (Luck, surface. However, it has since become clear that
2014a). This yields a time course of activity at each unique spatial patterns produced within underlying neu-
channel – a sequence of “waveforms” that in EEG are roanatomical sources can give rise to unique spatial pat-
referred to as event-related potentials (ERPs), and in terns of electrical and magnetic activity that can be
MEG are referred to as event-related fields (ERFs). detected either across the full set of recording channels
These event-related responses can be described in (e.g. Cichy et al., 2014; Stokes et al., 2015) or across
terms of their amplitude, latency, duration, scalp topo- subsets of channels (e.g. Karimi-Rouzbahani et al.,
graphy, and, in the case of EEG, their polarity (whether 2021) at the head surface.
the voltage is negative-going or positive-going). If we In a multivariate analysis, the basic assumption is that
believe that a particular event-related response is the experimental variable of interest influences the geo-
reliably modulated by a particular type of psychological metric relationship amongst these spatial patterns. For
event, then we refer to it as an event-related component example, one type of multivariate approach – spatial
(Kappenman & Luck, 2012). similarity analysis, otherwise known as representational
At a neurophysiological level, the interpretation of an similarity analysis asks whether and when the simi-
event-related component is relatively straightforward: larity/dissimilarity amongst items, with respect to the
its amplitude (often averaged across a particular time experimental variable, can explain the similarity/dissim-
window and adjacent channels) is taken to reflect the ilarity amongst the spatial patterns produced by these
strength of time-locked (and phase-locked) neural items (Kriegeskorte et al., 2008; Nili et al., 2014). In a
activity produced by the underlying dipole source(s), typical representational similarity analysis stream, this
and any difference in amplitude that can be explained is examined by constructing a “Model” dissimilarity
by our experimental variable is taken to reflect differ- matrix and correlating it with a “Neural” dissimilarity
ences in the strength of this underlying neural activity. matrix at each time point following event onset, yielding
At a cognitive level, this difference is usually interpreted a time series of r values (e.g. Cichy et al., 2014).
as reflecting the effect of the experimental variable on a In contrast to event-related components, there is no
neurocognitive “process” (Kappenman & Luck, 2012). direct neurophysiological interpretation of these multi-
This classic event-related analysis approach has variate time series. They simply tell us whether and
yielded a large literature that has characterised a when the spatial patterns of neural activity produced
number of different event-related components, which by the stimuli can be discriminated by the experimental
have each been linked to different aspects of language variable of interest. From a cognitive perspective, this is
processing (see Swaab et al., 2012 and Dikker et al., usually taken to reflect whether and when “represen-
2020 for reviews of language-related ERP and MEG com- tational information” linked to this variable is neurally
ponents, respectively). For example, during sentence decodable.
comprehension, N400 event-related component1 is
usually taken to reflect the ease of “retrieving” or “acces-
The present study
sing” the semantic features of incoming words (Hagoort,
2013; Lau et al., 2008; Van Berkum, 2009). To sum up, there is a long history of using EEG and MEG
The second approach that we can take to analyse with event-related analyses to study the neural basis of
EEG/MEG data has a much shorter history in neurophy- online language comprehension. The combination of
siology and psycholinguistics. Instead of focusing on EEG/MEG with spatial similarity analysis and other multi-
the amplitude of activity measured at individual adjacent variate methods is more recent. Because these newer
recording channels, this approach aims to characterise multivariate approaches come from a different tradition
whole patterns of activity, measured across multiple and use different statistical methods, it is sometimes
recording channels. This multivariate analysis approach assumed that they detect neural activity that is function-
was first developed in fMRI to describe patterns of ally distinct and separable from that indexed by classic
activity observed across multiple voxels within neuroa- EEG/MEG components. However, this assumption is
natomical regions of interest (Haxby et al., 2001). It not always valid, and, despite close methodological
was argued that, in contrast to univariate methods, links between the two approaches, few psycholinguistic
 LANGUAGE, COGNITION AND NEUROSCIENCE 3

studies have systematically compared the results of bottom-up information matches prior predictions,
these two types of analyses in the same datasets. In failing to activate error units and produce a large univari-
addition, there has been little discussion in the psycho- ate response, it is still encoded within separate sets of
linguistic literature about how to synthesise the results “state units”. As such, it should still be possible to
of event-related and spatial similarity analyses to detect this expected information within the critical
inform theory. Finally, few studies have directly com- 300–500ms time window in which the bottom-up
pared univariate or multivariate findings between MEG input first makes contact with semantic memory.
and EEG. In this investigation, we aimed to close these To test this theory, we measured MEG and EEG as par-
gaps. We addressed three sets of questions. ticipants read three-sentence discourse scenarios. We
varied both the Plausibility and the Animacy of a critical
Question 1: when and how does the brain encode word in the final sentence, such that it was either
the Animacy of incoming words during online animate or inanimate, and either plausible (matching
language comprehension? the animacy constraints of the prior verb) or anomalous
Our first goal was to use multivariate and univariate (mismatching these animacy constraints). In addition to
methods with MEG/EEG to address a theoretical ques- carrying out event-related univariate analyses, we
tion: When and how does the brain encode the carried out multivariate spatial similarity analyses on
animacy of incoming words during online language the same datasets. Previous fMRI studies have shown
comprehension, and how is this influenced by the that multivariate methods can discriminate animate
prior context? and inanimate visual objects (Devereux et al., 2013; Krie-
A large literature using event-related approaches has geskorte et al., 2008; Proklova et al., 2016; Sha et al.,
established that, between 300–500ms, incoming words 2015) and words (Devereux et al., 2013). In addition, pre-
whose semantic features match these that were pre-acti- vious EEG/MEG studies have used representational simi-
vated by the prior context produce a smaller N400 larity analysis to show that animacy-based
event-related component than words whose semantic discriminations of visual objects occur rapidly (Carlson
features fail to match these predictions (ERP: Kuperberg et al., 2013; Cichy et al., 2014; Cichy & Pantazis, 2017;
et al., 2020; Kutas & Hillyard, 1980; Nieuwland et al., 2020; Khaligh-Razavi et al., 2018). However, no previous
MEG: Halgren et al., 2002; Helenius et al., 1998; Ihara study has used these methods to ask whether and
et al., 2007; Maess et al., 2006). For example, plausible when the brain distinguishes between animate and
nouns whose animacy-based semantic features match inanimate incoming words during language comprehen-
those that are pre-activated by a prior verb (Szewczyk sion. Based on principles of predictive coding, we pre-
& Schriefers, 2013; Wang et al., 2020) produce a dicted that the spatial similarity analysis would detect
smaller N400 than anomalous nouns whose features an effect of Animacy within the same 300–500 ms time
fail to match these verb-based predictions (Kuperberg window in which the event-related analysis detected
et al., 2020; Paczynski & Kuperberg, 2011, 2012; Szewc- an effect of Plausibility on the N400, and that this
zyk & Schriefers, 2013). effect would be detected regardless of whether the
In line with the idea that event-related components incoming words matched or mismatched the animacy-
index “processing”, the smaller N400 response to based constraint of the prior verb.
expected (versus unexpected) words is usually taken to In addition to asking when the animacy-based fea-
reflect the ease of semantic “retrieval” (Hagoort, 2013; tures of incoming words were encoded, we were also
Van Berkum, 2009) or lexico-semantic “access” (Lau interested in how these features were encoded. A long
et al., 2008). However, another way of understanding line of Cognitive Science research has established that
the N400 is as indexing the amount of new information the semantic features of animate items are more
that is encoded within the bottom-up input, i.e. the similar to one another than the semantic features of
amount of information that was not already predicted inanimate items (Garrard et al., 2001; McRae et al.,
by the prior context (Kuperberg, 2016). This is the 1997; Randall et al., 2004; Zannino et al., 2006). It has
premise of predictive coding – – a general computational been proposed that these inherent differences in seman-
theory of brain function that posits that unpredicted tic similarity can account for the brain’s sensitivity to
information encoded within the bottom-up input is animacy-based categorical structure (Devlin et al.,
detected by local “error units”, which produce “predic- 1998; Gonnerman et al., 1997; Moss et al., 1998; Taylor
tion error” (Friston, 2005; Mumford, 1992; Rao & et al., 2011; Tyler & Moss, 2001). On this account, the
Ballard, 1999), and an event-related response (Friston, similarity between the spatial patterns produced by
2005), see General Discussion for further elaboration. Cri- animate and inanimate nouns should mirror this internal
tically, according to predictive coding, even if new similarity structure. We therefore hypothesised that,
4 L. WANG AND G. R. KUPERBERG

between 300–500 ms, any spatial similarity effect of discriminated (cf. Kriegeskorte et al., 2008; Nili et al.,
Animacy would be driven by a greater similarity 2014). Therefore, directly comparing the results of the
amongst the spatial patterns produced by animate two methods is challenging.
than inanimate nouns. The experimental design and the spatial similarity
To test these hypotheses, we extracted the spatial analysis methods that we employed in the current
patterns produced by each critical noun at each time study allowed us to overcome this challenge. Our strat-
point following its onset and calculated the spatial simi- egy of constructing separate spatial similarity time
larity between these patterns across all pairs of nouns. series for each condition allowed us to directly
We then averaged these pairwise spatial similarity compare the shape and time course of these time
values (r values) within each of the four experimental series with the event-related waveforms produced by
conditions, giving rise to four spatial similarity time each condition. Moreover, by orthogonally manipulating
series. This allowed us to test our hypothesis regarding Animacy and Plausibility in the same set of linguistic
the directionality of the spatial similarity effect (animate stimuli, and examining the effect of both variables on
> inanimate), as well as to directly compare its timing the same univariate and multivariate measures, we
with the event-related effect of Plausibility. were able to determine when and how the univariate
and multivariate results converged and diverged. We
Question 2: how do spatial similarity and event- hypothesised that the event-related analyses would fail
related measures converge and diverge during to detect any spatial similarity effects of Animacy that
language comprehension? were detected by the spatial similarity analyses, but
Our second goal was more methodological in nature. As that the spatial similarity analyses would capture all
discussed earlier, event-related and spatial similarity the effects of Plausibility that were detected by the
analyses each aim to characterise different aspects of event-related analyses.
the underlying neural signal – the strength of the
dipoles versus the spatial patterns they produce. In prin- Question 3: how do the results of MEG and EEG
ciple, these two measures should be independent of converge and diverge during language
each other. However, both analysis methods face the comprehension?
challenge of distinguishing very small signals from Our third goal was to directly compare the results (both
ongoing activity that is not time-locked to stimulus univariate and multivariate) of data that were collected
onset, and independent of the experimental variable(s) using two different techniques – MEG and EEG.
of interest, i.e. “noise”. Therefore, in practice, estimates Because MEG and EEG both index postsynaptic
of spatial similarity are likely to covary with the ampli- activity that is produced by pyramidal neurons, they
tude of evoked responses. This covariation between uni- are sensitive to many of the same event-related com-
variate and multivariate responses has been discussed in ponents, including the N400. There is also evidence
the fMRI literature (e.g. Haxby et al., 2001; Jimura & Pol- that MEG and EEG are sensitive to some of the same
drack, 2012; Walther et al., 2016), and to some degree, in multivariate effects (Cichy & Pantazis, 2017; Wang
the EEG/MEG literature (Guggenmos et al., 2018). et al., 2020). However, there are also important differ-
However, the degree to which multivariate methods ences between the two techniques. Unlike the magnetic
are sensitive to the classic event-related components field, the electric field is smeared at the head surface
and effects that are typically produced during language (Geisler & Gerstein, 1961; Grynszpan & Geselowitz,
comprehension is unclear. 1973). In addition, MEG and EEG differ in their sensi-
One reason why we know so little about how event- tivities to underlying dipole sources with different orien-
related and multivariate measures converge and diver- tations (Cuffin & Cohen, 1979; Hämäläinen et al., 1993).
gence is that these methodologies use quite different Therefore, the two types of techniques may not always
analysis streams and methods of visualisation. As dis- detect the same neurocognitive mechanisms. For
cussed above, in a typical event-related analysis, we example, MEG is less sensitive than EEG to the domain-
average activity across items within each condition at general P300 event-related component (Siedenberg
each time point, with the goal of determining when et al., 1996).
one condition produces more or less activity than In the present study, we were particularly interested
another (Kappenman & Luck, 2012). In contrast, in a in comparing the MEG and EEG responses to anomalous
typical representational similarity analysis, we examine inputs in a later time window between 600–1000 ms.
correlations between a Model dissimilarity matrix and When a linguistic anomaly cannot initially be incorpor-
a Neural dissimilarity matrix at each time point, with ated into a high-level interpretation, the disruption of
the goal of determining when the conditions can be comprehension is thought to trigger multiple processes,
 LANGUAGE, COGNITION AND NEUROSCIENCE 5

including reprocessing the bottom-up input, reactivat- We began by analysing the data using a traditional
ing prior top-down predictions, and tracking conflict event-related analysis stream, averaging across trials
between these different sources of information to within each condition at each time point after critical
verify that the disruption stemmed from an external word onset. To test our hypotheses, we contrasted the
(rather than an internal) error (Kuperberg et al., under amplitude of the ERFs across the four conditions
review). within two time windows of interest: 300–500 ms (to
In EEG, the prolonged late error-based response capture the N400) and 600–1000 ms (to capture post-
manifests as a large posteriorly distributed late positivity N400 activity). Based on previous MEG studies, we
ERP effect, known as the P600 (Brothers et al., 2020; expected that between 300–500 ms, the amplitude of
Brothers et al., 2022; Kuperberg, 2007; Kuperberg et al., the N400 over left temporal sensors would be larger to
2003; Kuperberg et al., 2020; Kuperberg et al., under the semantically anomalous than the plausible nouns
review; van de Meerendonk et al., 2009). The P600 ERP (Halgren et al., 2002; Helenius et al., 1998; Ihara et al.,
component shares several functional properties with 2007; Maess et al., 2006). Based on the prior ERP litera-
the P300 (Coulson et al., 1998; Osterhout et al., 2012; Sas- ture (Brothers et al., 2020; Kuperberg, 2007; Kuperberg
senhagen et al., 2014; Sassenhagen & Fiebach, 2019), et al., 2020; van de Meerendonk et al., 2009; Van
and it may play a particularly important role in tracking Petten & Luka, 2012), we also considered the possibility
the source of linguistic errors (Brothers et al., 2022; that we would additionally see an effect of Plausibility
Kuperberg et al., under review). However, previous (anomalous > plausible) in the later 600–1000 ms time
MEG studies have not reported an analogous effect on window. However, because most previous MEG studies
anomalies in this late time window. This raises the possi- have not reported findings past the N400 time
bility that, due to their differential sensitivity to distinct window, the size and topography of any late effect of
neuroanatomical sources, MEG and EEG may also be Plausibility that we might see on these late ERFs were
differentially sensitive to distinct aspects of the pro- unclear.
longed error-based response between 600–1000 ms. We then carried out a spatial similarity analysis on the
To date, few psycholinguistic studies have directly same MEG dataset. We aimed to address two questions.
compared the results of EEG and MEG, using the same The first was whether this analysis would pick up on the
stimuli and tasks. We therefore carried out the same same underlying neural activity detected by the event-
experiment using MEG (reported in Study 1) and EEG related analysis. If so, then the timing of the rise and
(reported in Study 2). This allowed us to qualitatively fall of the spatial similarity r values should track that of
compare both the event-related and spatial similarity both early and late event-related components. In
results across the two techniques. We hypothesised addition, this analysis should reveal similar effects of
that the MEG and EEG results would largely converge Plausibility (anomalous > plausible) as the event-related
within the 300–500 ms time window, but that they analysis within both the N400 (300–500 ms) and post-
would show differences in their sensitivities to the N400 (600–1000 ms) time windows.
activity produced by semantic anomalies in the later Our second question was whether and when the
600–1000 ms time window. spatial similarity analysis would detect additional
effects of the Animacy on the critical nouns that were
not detected by the event-related analysis. As noted in
Study 1: MEG the General Introduction, previous MEG studies have
shown that multivariate methods can discriminate
Introduction
between animate and inanimate visual objects (Carlson
We recorded MEG while participants read multiple et al., 2013; Cichy et al., 2014; Cichy & Pantazis, 2017;
three-sentence scenarios. We orthogonally manipulated Khaligh-Razavi et al., 2018). However, previous studies
the Plausibility and the Animacy of critical nouns in the have not asked whether and when these methods are
final sentence, such that they either matched or violated able to discriminate the animacy of linguistic inputs
the animacy-based constraints of the preceding verb. during sentence-level comprehension, or what drives
This 2 × 2 design gave rise to four experimental con- the discrimination in the spatial patterns produced by
ditions in which the critical nouns were: (1) plausible the animate versus inanimate concepts. To address
and animate (e.g. “ … greeted the guests … ”), (2) plaus- these questions, we averaged the pairwise spatial simi-
ible and inanimate (e.g. “ … clamped the wires … ”), (3) larity values (r values) within each experimental con-
anomalous and animate (e.g. “ … clamped the *guests dition at each time point following critical word onset,
… ”), and (4) anomalous and inanimate (e.g. “ … and constructed four spatial similarity time series. This
greeted the *wires … ”). enabled us to determine both the timing and
6 L. WANG AND G. R. KUPERBERG

directionality of any spatial similarity effects, as well as to across contextual constraint. The scenarios were coun-
directly compare the spatial similarity time series with terbalanced across four lists, with approximately 50 sen-
the event-related time series. tences of each condition within each list.

Similarity structure of the critical words

Methods
Semantic similarity between pairs of animate and
Participants inanimate nouns. To confirm that the animate critical
The MEG study was carried out at the Martinos Centre at nouns were semantically more similar to each other
Massachusetts General Hospital. We report data from 32 than the inanimate nouns, we quantified the semantic
participants (16 females, mean age: 23.4; range: 18–35). similarity between all pairs of animate nouns and all
Thirty-three participants originally took part in the pairs of inanimate nouns using word2vec (Mikolov
experiment, but the data of one participant was et al., 2013). Word2vec is a neural network model that
excluded because of technical problems. Written learns distributed semantic representations of words
consent was obtained from all participants following by training it on the Google News corpus of ∼100
the guidelines of the Mass. General Brigham Institutional billion words. Each word is represented as a 300-dimen-
Review Board (IRB). All participants were right-handed sional vector, and words with similar meanings are
and had normal or corrected-to-normal vision. All were located close to each other in this high-dimensional
native speakers of English, with no additional language space. We calculated the semantic similarity between
exposure before the age of five. Participants were not each pair of words by measuring the cosine distance
taking psychoactive medication and were screened to between their corresponding 300-dimensional word
exclude the presence of psychiatric and neurological vectors, using the Genism natural language processing
disorders. library in Python.
To test for a statistical difference in semantic similarity
Experimental design between the animate and inanimate nouns pairs, we
Each participant read 200 three-sentence discourse used a permutation approach. Specifically, we calculated
scenarios. The first two sentences set up a discourse the mean difference in similarity between the two
context, and the third sentence had a fixed structure: groups of nouns and used this as our test statistic. We
an adjunct phrase (ranging between 1–4 words), a then randomly reassigned the similarity values to
subject, verb, determiner, direct object noun (the “criti- either the animate or inanimate group and recalculated
cal noun”), followed by three additional words. In half the mean difference in similarity between the two
the scenarios, the verb in the third sentence constrained groups. We repeated this process 1000 times to create
for an animate direct object noun (e.g. “ … greeted the a null distribution of mean differences. We considered
… ”), and in the other half, the verb constrained for an our observed test statistic to be statistically significant
inanimate direct object noun (e.g. “ … clamped the if it fell within the top or bottom 2.5% of the null distri-
… ”). In half the scenarios, the critical noun matched bution. This analysis confirmed that the semantic simi-
the verb’s animacy constraints (its selectional restric- larity amongst pairs of animate nouns (mean ± SD:
tions), rendering the scenarios plausible, and in the 0.13 ± 0.11) was significantly greater than amongst
other half, it violated these constraints, rendering the pairs of inanimate nouns (mean ± SD: 0.11 ± 0.12, p =
scenarios anomalous. The animate nouns were slightly 0.001).
longer (mean number of letters ± SD: 7.51 ± 2.14) than
the inanimate nouns (6.56 ± 2.13, t(698) = 4.56, p < 0.001, Similarities in word length and frequency between
d = 0.35). In addition, the animate nouns were slightly pairs of animate and inanimate nouns. We also quan-
less frequent (mean log frequency ± SD: 0.74 ± 0.92; tified the similarity in length (number of letters) and log
based on Brysbaert & New, 2009) than the inanimate frequency (Brysbaert & New, 2009) amongst the animate
nouns (0.94 ± 0.82, t(698) = −3.07, p = 0.001, d = −0.23). and inanimate nouns by computing the absolute differ-
A detailed description of how the stimuli were con- ence in these values for each pair of items. For each of
structed and normed can be found in Wang et al. these variables, we took the difference between the
(2020) and Kuperberg et al. (2020). As discussed there, mean similarity value (i.e. the absolute difference
we also varied the lexical constraint of the prior contexts value) in each group of nouns as our test statistic and
such that, in each of the four conditions, half the critical used a permutation test (1000 permutations) to deter-
nouns followed high constraint contexts and half fol- mine if there were significant differences between the
lowed low constraint contexts. However, for the two groups. We found that the pairs of animate nouns
purpose of the analyses reported here, we collapsed were slightly less similar to one another (i.e. showed
 LANGUAGE, COGNITION AND NEUROSCIENCE 7

greater differences) than the pairs of inanimate nouns, After applying a bandpass filter at 0.1–30 Hz, the data
both in word length (animate: 2.41 ± 1.84, inanimate: were segmented into −100–1000 ms epochs relative to
2.34 ± 1.91, p = 0.001) and frequency (animate: 1.05 ± critical word onset. Epochs in which amplitudes
0.78, inanimate: 0.91 ± 0.71, p = 0.001). exceeded pre-specified cutoff thresholds (4e-10 T/m
for gradiometers and 4e-12 T for magnetometers) were
Experimental procedure removed, leaving 40 - 42 trials (on average) in each of
The discourse scenarios were presented over eight the four conditions.
blocks, with short breaks in between each block.
Before the experimental session, a practice session was Event-related analysis
conducted to familiarise participants with the stimulus At each recording site, we combined the activity pro-
presentation and tasks. duced by the two gradiometer sensors by computing
In each scenario, the first two sentences were each their root mean square value. We carried out an event-
presented as a whole for 3900 ms (100 ms interstimulus related analysis using Fieldtrip, an open-source Matlab
interval, ISI). Then, after a white fixation (“++++”) for toolbox (Oostenveld et al., 2011). In each participant,
550 ms (100 ms ISI), the third sentence appeared word- we calculated the evoked responses, relative to critical
by-word, with each word presented for 450 ms word onset, by averaging activity across trials within
(100 ms ISI). Following the sentence-final word, a pink each of the four conditions at each sampling point fol-
question-mark (“?”) appeared for 1400 ms (100 ms ISI), lowing critical word onset, using a −100–0 ms baseline.
which cued participants to press one of two buttons In MEG, the topographic distribution of the N400
with their left hand to indicate whether or not the scen- effect at the head surface has been well characterised:
ario made sense. In addition, following a proportion of Several previous studies have shown that it is produced
trials (24/200), which were semi-randomly distributed primarily over left temporal sensors (e.g. Lau et al., 2013;
across blocks, participants responded to a yes/no com- Wang et al., 2018; Wang et al., 2023). Therefore, to
prehension question that appeared on the screen for compare the magnitude of the N400 across conditions,
1900ms (100 ms ISI). After each trial, a blank screen we averaged activity across nine left temporal sites
appeared with a variable duration, ranging from 100 to (highlighted in the topographic plot in Figure 1A)
500 ms, followed by a green fixation (“++++”) for between 300–500 ms. These mean values served as the
900 ms (100 ms ISI) during which participants were dependent variable in a repeated-measures ANOVA,
encouraged to blink. with two within-subject variables: Plausibility (plausible,
anomalous) and Animacy (animate, inanimate).
MEG data acquisition and preprocessing We also carried out statistical analyses on the mean
MEG data were acquired with a Neuromag VectorView activity between 600–1000 ms. In this time window, pre-
system (Elekta-Neuromag Oy, Finland) with 102 triplets vious studies using ERPs have reported late effects of
of sensors. Each triplet was comprised of two orthogonal Plausibility (anomalous > plausible) on a positive-going
planar gradiometers and one magnetometer. EEG data component known as the P600. However, most previous
were acquired at the same time as the MEG data (see MEG studies have failed to examine activity within this
Study 2). MEG signals were digitised at 1000 Hz, with later time window, and so the topography of any late
an online bandpass filter of 0.03 - 300 Hz. Vertical and MEG effect was unclear. We therefore took a mass uni-
horizontal electrooculography (EOG) data, as well as variate statistical approach to determine whether and
electrocardiogram (ECG) data were collected with where there were differences across conditions in
bipolar recordings. Impedances were kept at less than evoked activity within this late time window, carrying
30kΩ. out tests at all 102 combined gradiometer sensors. To
The data were preprocessed using the Minimum test for the main effects of each variable, we carried
Norms Estimates (MNE) software package in Python out dependent-samples t-tests, collapsing across the
(Gramfort et al., 2014). We first identified “bad” sensors two levels of the other variable. To test for an interaction
in each participant in each block, resulting in the between Plausibility and Animacy, we computed differ-
removal of seven (on average) out of the 306 MEG ence waveforms (inanimate minus animate) separately
sensors. The data of these bad MEG sensors were later for the anomalous and plausible conditions, and con-
interpolated using spherical spline interpolation (Perrin trasted these two difference waveforms. To correct for
et al., 1989). Eye-movement and blink artifacts were multiple comparisons, we used a cluster-based permu-
automatically removed (Gramfort et al., 2014), and tation test (Maris & Oostenveld, 2007): All spatially adja-
Signal-Space Projection (SSP) correction (Uusitalo & cent data samples that exceeded a preset uncorrected
Ilmoniemi, 1997) was used to correct for ECG artifacts. significance threshold of 5% were taken as a cluster,
8 L. WANG AND G. R. KUPERBERG

Figure 1. Results of the MEG event-related and spatial similarity analyses (n = 32). A. Grand-average ERF waveforms following the
onset of critical nouns at a left temporal site (MEG1512 + 1513) and a left occipital sensor site (MEG1932 + 1933). The duration of
the critical nouns (0 - 450 ms) and the subsequent word (550 - 1000 ms) are marked with gray bars on the x-axes. B. Topographic
maps showing the distributions of the effects of Plausibility (collapsed across the animate and inanimate conditions) between 300–
500 ms (left) and 600–1000 ms (right). MEG sensors that showed significant differences are indicated with white dots. C. Bar graphs
showing the amplitude of the MEG event-related fields (ERFs) in each condition, averaged between 300–500 ms at a representative
left temporal sensor (left), and between 600–1000 ms at a representative right frontal sensor (right). Error bars represent ±1 standard
error of the mean. Between 300–500 ms, the anomalous nouns produced a significantly larger N400 than the plausible nouns. The
amplitude of the N400 did not differ significantly between the animate and inanimate nouns, either in the plausible or the anomalous
condition. Between 600–1000 ms, the anomalous nouns produced slightly larger ERF than the plausible words at right frontal MEG
sensors. Again, the magnitude of this ERF did not differ significantly between the animate and inanimate nouns in either the plausible
or the anomalous condition. D. Grand-average spatial similarity time series following the onset of critical nouns in each of the four
conditions. The duration of the critical nouns (0 - 450 ms) and the subsequent word (550 - 1000 ms) are marked with gray bars on the
x-axis. E. Bar graphs showing the mean MEG spatial similarity values for each condition, averaged between 300–500 ms (left) and
between 600–1000 ms (right). Error bars represent ±1 standard error of the mean. Between 300–500 ms, the anomalous nouns pro-
duced larger spatial similarity values than the plausible nouns, mirroring the ERF findings. In addition, the animate nouns produced
significantly larger spatial similarity values than the inanimate nouns in both plausible and anomalous conditions. The same pattern
was observed between 600–1000 ms.

and individual t-statistics within each cluster were cluster-level statistics for each randomisation, and enter-
summed to yield cluster-level test statistics. We built a ing the largest cluster-level statistic into the null distri-
null distribution by randomly assigning condition bution. We then compared the observed cluster-level
labels within each participant 1000 times, calculating test statistics against the null distribution and took any
 LANGUAGE, COGNITION AND NEUROSCIENCE 9

clusters falling within the highest or lowest 2.5th percen- words produced a slightly larger ERF than the plausible
tile to be statistically significant. words at right frontal MEG sensors (cluster-based per-
mutation test: p = 0.014), although the magnitude of
Spatial similarity analysis this effect was small, see Figure 1B (right). No clusters
We carried out a spatial similarity analysis on the prepro- revealed a significant effect of Animacy or an interaction
cessed MEG data using custom-written scripts. In each between Plausibility and Animacy within this late time
participant, on each trial, and at each time point follow- window.
ing critical word onset, we extracted a vector that
characterised the spatial pattern of neural activity pro- MEG spatial similarity results
duced across all sensors at all 102 recording sites (i.e. Description of the MEG spatial similarity time series.
306 sensors in total – 204 gradiometer and 102 magnet- As shown in Figure 1D, the overall shape of the spatial
ometer), and computed the similarity between the similarity time series mirrored that of the ERF time
spatial patterns produced by each pair of critical nouns courses: The rise and fall of the spatial similarity values
by correlating these spatial vectors (Pearson r). We between 100–200 ms corresponded to the M170 ERF
then averaged these r values across all pairs of trials component, and the rise and fall of the spatial similarity
within each of the four conditions to create four values between 300–500 ms, particularly to the anoma-
spatial similarity time series. lous nouns, corresponded to the N400. The rise and fall
For statistical analyses, we averaged these spatial of spatial similarity values between 650–750 ms corre-
similarity values across the same 300–500 ms and 600– sponded to the M170 evoked by the following word.
1000 ms time windows that were used for the event-
related analysis, and carried out repeated measures 300–500ms. As shown in Figure 1D and E, the spatial
ANOVAs with Plausibility (plausible, anomalous) and similarity values, averaged between 300–500 ms, were
Animacy (animate, inanimate) as within-subject vari- significantly larger to the anomalous than to the plaus-
ables. To visualise the data, we averaged the time ible nouns (Main effect of Plausibility: F(1,31) = 50.06, p
series for each condition across all participants, yielding < 0.001, eta2 = 0.617). In addition, both the plausible
four grand-average spatial similarity time series. and anomalous critical words produced spatial similarity
values that were slightly larger to the animate than to
the inanimate nouns (Main effect of Animacy: F(1,31) =
Results
7.86, p = 0.009, eta2 = 0.202). There was no interaction
Event-related responses between Plausibility and Animacy (F(1,31) = 0.053, p =
Description of ERF time courses. As shown in Figure 1A, 0.820, eta2 = 0.002).
between 100–200 ms, the critical words in all four con-
ditions produced an M170 ERF component (Tarkiainen 600–1000ms. Both the spatial similarity effects
et al., 1999) at bilateral occipital sensors, particularly described above continued into the 600–1000 ms
over the left hemisphere. Between 300–500 ms, the criti- window, see Figure 1D and E. The spatial similarity
cal words, particularly the anomalous words, produced values remained larger to the anomalous than plausible
an N400 that peaked at 400 ms, and was maximal over nouns (Main effect of Plausibility: F(1,31) = 16.384, p <
bilateral temporal sensors, particularly over the left 0.001, eta2 = 0.346), and slightly larger to the animate
hemisphere. Within the later 600–1000 ms time than inanimate nouns, regardless of Plausibility (Main
window, there was no clear evidence of a distinct wave- effect of Animacy: F(1,31) = 5.073, p = 0.032, eta2 =
form, although the M170 evoked by the following word 0.141). Again, there was no interaction between Plausi-
was visible at occipital sensors between 650–750 ms. bility and Animacy (F(1,31) = 0.38, p = 0.54, eta2 = 0.012).

300–500ms. As shown in Figure 1 (left), the N400 ERF Follow-up exploratory analyses
was significantly larger to the anomalous than the plaus- To sum up, the spatial similarity analyses picked up on
ible nouns (Main effect of Plausibility: F(1,31) = 44.33, p < the same effects of Plausibility that were detected by
0.001, eta2 = 0.583). There was no evoked effect of the classic event-related analyses. In addition, the
Animacy (F(1,31) = 2.57, p = 0.119, eta2 = 0.077), or any spatial similarity analyses detected effects of Animacy
interaction between Plausibility and Animacy (F(1,31) = (animate > inanimate) that were not detected by the
0.34, p = 0.854, eta2 = 0.001).2 event-related analyses. To further explore the relation-
ship between the univariate and multivariate findings,
600–1000ms. Within this later time window, the cluster- we carried out two additional analyses, focusing on
based permutation tests showed that the anomalous the effect of Animacy in the plausible sentences
10 L. WANG AND G. R. KUPERBERG

between 300–500 ms (which we replicated using EEG in 2007; Maess et al., 2006; Wang et al., 2023). Here, we
Study 2). We addressed two questions. show that it was followed by a smaller evoked effect
First, was the failure to detect a univariate effect of of Plausibility between 600–1000 ms at right frontal
Animacy within this time window simply because we recording sites. We further show that these same
constrained our univariate analysis to left temporal effects of Plausibility were detected by the spatial simi-
sensors? As discussed under Methods, the reason why larity analyses.
we chose to average activity across these sensors is The spatial similarity analysis additionally revealed an
that previous MEG studies have shown that this is effect of Animacy. This effect was smaller than the effect
where the N400 effect is largest at the head surface of Plausibility, but it showed a similar time course,
(e.g. Lau et al., 2013; Wang et al., 2018; Wang et al., appearing between 300–500 ms. It was driven by a
2023). However, few MEG studies have tested for an greater similarity amongst the spatial patterns for
effect of Animacy on univariate activity. It is therefore animate than inanimate nouns. These differences in
possible that by taking this analysis approach, we spatial similarity cannot be explained by differences in
missed a univariate effect of Animacy over other similarity amongst the animate and inanimate nouns in
sensors, which could have driven the multivariate their word length or frequency, which went in the oppo-
effect of Animacy observed in the 300–500 ms time site direction (see Study 1 Methods: Similarities in word
window. To determine whether this was the case, we length and frequency between pairs of animate and inan-
tested for an effect of Animacy on the mean response imate nouns). Instead, they mirror the difference
between 300–500 ms in the plausible sentences at all amongst the animate and inanimate nouns in their
102 (combined) gradiometer sensors, and used cluster- semantic similarity structures (animate > inanimate)
based permutation tests to control for multiple compari- (Garrard et al., 2001; McRae et al., 1997; Randall et al.,
sons (Maris & Oostenveld, 2007), i.e. the same approach 2004; Zannino et al., 2006). These findings have impor-
that we took to carry out the univariate analysis in the tant theoretical implications, which we consider in
600–1000 ms time window. This mass univariate analysis detail in the General Discussion.
failed to reveal any spatial clusters showing an effect of In this Discussion section, we will focus on why the
Animacy. multivariate and univariate measures converged and
Having excluded the possibility that the spatial simi- diverged from a methodological perspective.
larity effect of Animacy was driven by a univariate
effect at any group of individual channels, we can con- Convergence between the spatial similarity and
clude that it was driven by differences in similarity event-related results
amongst the spatial patterns produced across multiple The results of the spatial similarity analyses converged
channels. This led to a second question: Would we be with the results of the event-related analysis in two
able to detect the same spatial similarity effect if we ways. First, the overall spatial similarity time series
used only the subset of sensors that are known to con- revealed a sequence of “waveforms” whose timing
tribute to the N400 evoked response? To address this coincided with the event-related waveforms. Second,
question, we repeated the spatial similarity analysis, the spatial similarity analysis captured all the effects
but instead of using all 306 sensors at all 120 recording detected by the event-related analysis. For example,
sites, we used only the sensors at the nine left temporal between 300–500 ms (the N400 time window), the
sites that we used in the univariate analysis (i.e. 27 spatial similarity values were larger for the anomalous
sensors in total – 18 gradiometer and 9 magnetometer). than for the plausible nouns, mirroring the larger N400
This subset analysis indeed revealed the same effect of event-related response to anomalous than plausible
Animacy on spatial similarity values: the spatial patterns nouns.
produced by pairs of animate plausible nouns were again From a mathematical perspective, the Pearson’s cor-
more similar than those produced by pairs of inanimate relation between a pair of spatial vectors should be inde-
plausible nouns, t(31) = 2.29, p = 0.029, d = 0.787. pendent of each vector’s magnitude. It may therefore
not be obvious why the similarity between pairs of
spatial patterns, which describe neural activity across
Discussion
multiple channels, covaried with the amplitude of
The event-related and spatial similarity analyses of the evoked activity at adjacent channels.
MEG data both revealed effects of Plausibility. The The reason for this is that evoked responses and
larger evoked N400 response to anomalous than plaus- spatial similarities are derived from the same very
ible nouns replicates findings from previous studies small neural signals that are embedded in noise. By
(Halgren et al., 2002; Helenius et al., 1998; Ihara et al., “noise”, we mean any measurement that is of no
 LANGUAGE, COGNITION AND NEUROSCIENCE 11

theoretical interest to the experimenter, i.e. that is detect the same increases and decreases in signal
neither time-locked to the experimental stimuli, nor strength as increases and decreases in the similarity in
modulated by the experimental variables of interest.3 the spatial patterns produced across multiple pairs of
To extract this signal from noise, both types of analyses words.
rely on the assumption that the signal of interest is con- Second, the shared sensitivity of the event-related
sistently produced across multiple trials of the same con- and spatial similarity analyses to common sources of
dition, while noise fluctuates randomly across trials. signal and noise also explains why whenever the
This idea has been discussed extensively in the event- event-related analyses detected a larger response to
related literature where the goal is to determine whether the anomalous than the plausible words, pairs of anom-
and how an experimental variable influences the alous nouns produced more similar patterns than pairs
strength of the underlying neural signal produced by of plausible words. To understand why the evoked and
individual items within a particular condition. To spatial similarity effects (i.e. the differences between
isolate this signal, one can “filter out” noise by averaging conditions) covaried, the key factor to consider is the
activity across multiple items (Luck, 2014a), as we did in signal-to-noise ratio in each condition. In the presence
the current study. Alternatively, one can use a General of signal and no noise, two identical spatial patterns
Linear Model, which models noise within the error would produce a Pearson’s r value of 1. For example, if
term (Baayen et al., 2008; Clark, 1973; Smith & Kutas, two anomalous words each produced twice as much
2015). In a spatial similarity analysis, the same basic prin- signal at the same three measurement channels as two
ciple applies except that, instead of estimating the plausible words, then, in the absence of noise, the
strength of the underlying neural signal produced by spatial similarity between the two anomalous words
individual items, the goal is to estimate the similarity would be the same as between the two plausible
between the spatial patterns of signals produced by words (in both cases, Pearson’s r is 1). However, any
pairs of items. Once again, to isolate this signal, one noise will result in an underestimation of the true
can either filter out noise by averaging spatial similarity spatial similarity value (r < 1), such that the weaker the
values across multiple pairs of items, as we did in the signal-to-noise ratio, the greater the reduction of the
present study, or one can use a single trial model- estimated value. Therefore, on the assumption that
based approach, such as correlating a Model dissimilar- noise is held constant, the larger signal produced by
ity matrix with a Neural dissimilarity matrix (e.g. Krieges- the two anomalous words will result in a greater
korte et al., 2008; Nili et al., 2014). signal-to-noise ratio. Therefore, our estimation of the
The sensitivity of these two types of analyses to over- spatial similarity between these two anomalous words
lapping sources of signal and noise can explain both will be more accurate, with an r value that is larger
aspects of the convergence we observed between the (closer to the true Pearson’s r value of 1) than between
event-related and spatial similarity results. First, it the two plausible words.
explains why, regardless of experimental condition, the These considerations raise the question of whether it
rise and fall of the spatial similarity values tracked the is possible to carry out a multivariate analysis that is not
rise and fall (relative to baseline) of the evoked M170 sensitive to univariate activity. For example, with the
and N400 ERF components. At baseline, when no input current design, is there a type of multivariate analysis
was presented, there was no signal – only noise. There- that would detect the effects of Animacy, but not the
fore, the expected value of both the event-related rise and fall of condition-non-specific evoked activity,
response and Pearson’s r was zero. However, as the infor- or the effects of Plausibility on the evoked response?
mation associated with each critical word became avail- One approach that some researchers have taken to
able to each level of the cortical hierarchy, it produced reduce the influence of univariate activity is to remove
transient increases and decreases in the strength of condition-non-specific responses before carrying out a
the neural signal. Because each word consistently pro- multivariate analysis. This so-called “cocktail blank
duced the same transient increases and decreases in removal” removes the “common” mean pattern of
activity within the same neural sources, the event- activity across all conditions by subtracting it from
related analysis was able to detect a consistent rise activity produced by each individual condition before
and fall in the amplitude of the magnetic field across computing Pearson’s correlations (Haxby et al., 2001;
multiple items at particular sets of recording sensors. Op de Beeck et al., 2006). However, this method can
Because these increases and decreases in activity had introduce false dependencies when examining differ-
consistent scalp topographies (i.e. they were consist- ences in spatial similarity between conditions (see Die-
ently produced at some recording channels but not drichsen et al., 2011 and Garrido et al., 2013 for
others), the spatial similarity analysis was able to discussion).
12 L. WANG AND G. R. KUPERBERG

A second approach, previously attempted in the fMRI The reason why multivariate measures, such as spatial
literature, is to remove univariate effects (differences similarity analysis, are sensitive to effects that are invis-
between conditions) before carrying out a multivariate ible to univariate measures is that they can capture sys-
analysis. Here, for each condition, within each spatial tematic differences between conditions in how patterns
region, the mean response across all voxels and all of activity, measured across multiple channels, vary
trials is subtracted from the activity produced by each across trials. In the present study, the spatial similarity
individual trial at each voxel (Jimura & Poldrack, 2012; analysis was able to exploit the fact that, across trials,
LaRocque et al., 2013). However, this approach intro- the spatial patterns produced by pairs of animate
duces other practical and conceptual difficulties (for dis- nouns were consistently more similar than those pro-
cussion, see Hebart & Baker, 2018). duced by pairs of inanimate nouns.
A third approach for dissociating multivariate and
univariate activity is to use a different measure of
spatial similarity – the Euclidean distance, instead of Study 2: EEG
Pearson’s correlations (or Pearson’s distance, i.e. 1 - r).
Introduction
In contrast to these Pearson’s measures, the Euclidean
distance is invariant to condition-non-specific responses In this second study, we presented participants with the
(Guggenmos et al., 2018; Walther et al., 2016). In same set of stimuli, but this time measured EEG rather
addition, at least in fMRI, there is evidence that this than MEG. We had two goals. The first was to replicate
measure can be insensitive to univariate differences the broad pattern of convergence and divergence
between conditions (e.g. see Walther et al., 2016, between the evoked and spatial similarity effects using
Figure 8). However, to yield results that can be meaning- a complementary technique. The second was to deter-
fully interpreted, it is necessary to use a cross-validated mine whether, where and how the EEG and MEG univari-
Euclidean distance measure (Guggenmos et al., 2018; ate and multivariate measures converged and diverged.
Walther et al., 2016).4 Cross-validation requires the Although MEG and EEG are both sensitive to tangen-
dataset to be split into at least two partitions. While par- tial dipole sources, only EEG can detect radial dipoles
titioning the data is feasible in studies of visual percep- (Ahlfors et al., 2010; Hämäläinen et al., 1993; Nunez,
tion in which the same visual stimulus is usually 1990). It is therefore able to detect activity that is invis-
presented multiple times, this would introduce signifi- ible to MEG (Ahlfors et al., 2010; Siedenberg et al.,
cant challenges in studies of language processing. This 1996). In the present study, we were particularly inter-
is because we typically present multiple different linguis- ested in whether EEG would detect additional activity
tic items within each experimental condition, and each produced by anomalous nouns in the post-N400 time
individual linguistic item is usually presented only window, between 600–1000 ms. Although MEG did
once to avoid psychological confounds. Therefore, to reveal a late effect of Plausibility over right frontal
compute the cross-validated Euclidean distance, we sensors in this late time window (anomalous > plausible),
would need to match items across partitions on numer- the size of this effect was small. In contrast, previous EEG
ous potentially confounding variables. Moreover, cross- studies have shown that between 600–1000 ms, anoma-
validation becomes impossible if we are interested in lous words that disrupt the process of deep comprehen-
isolating specific neural patterns that are unique to indi- sion produce a robust posteriorly distributed positive-
vidual items (e.g. Wang et al., 2018). going ERP component, known as the “semantic P600”
(Brothers et al., 2020; Kuperberg, 2007; Kuperberg
Divergence between the spatial similarity and et al., 2003; Kuperberg et al., 2020; van de Meerendonk
event-related results et al., 2009). This ERP component is thought to be func-
In contrast to the effect of Plausibility, the effect of tionally related to the P300 ERP component, to which
Animacy was only detected by the multivariate spatial MEG is known to be relatively insensitive (Ahlfors et al.,
similarity analysis. We saw no evidence that this effect 2010; Siedenberg et al., 1996). We were therefore inter-
was detected by the univariate analysis, either when ested in directly comparing the magnitude of the ERP
we collapsed across left temporal sites, or when we and ERF effects produced by the same set of anomalous
carried out a mass univariate analysis at all sensors. (versus plausible) words in this late time window, as well
Moreover, the multivariate effect was observed, regard- as exploring any differences between EEG and MEG in
less of whether the absolute amplitude of the event- how Animacy modulated spatial similarity in this same
related responses (and spatial similarity values) was rela- time window.
tively small (on the plausible nouns) or large (on the Finally, we were interested in whether EEG would
anomalous nouns). detect an interaction between Plausibility and Animacy
 LANGUAGE, COGNITION AND NEUROSCIENCE 13

in the earlier N400 time window. Although the MEG Each individual’s EEG data was segmented into −100–
study showed no evidence of this interaction, a previous 1000 ms epochs, relative to noun onset. We applied an
ERP study with a similar design, but using a different set Independent Component Analysis to correct for eye
of stimuli, showed that animate nouns produced a movement artifacts (Bell & Sejnowski, 1997; Jung et al.,
smaller effect of Plausibility than inanimate nouns (Pac- 2000). Remaining artifacts were removed based on
zynski & Kuperberg, 2011, Experiment 1). This interaction visual inspection, leaving 39 - 40 artifact-free trials in
was largely driven by a less negative waveform to the each of the four conditions across the two EEG datasets.
anomalous animate than the anomalous inanimate
nouns. In the present study, we were interested in Event-related analysis
whether this ERP finding would replicate. The spatial and temporal characteristics of both the
N400 and P600 ERP effects have been very well charac-
terised in previous studies. For statistical analyses, we
Methods
therefore selected spatiotemporal regions of interest
We collected EEG data in two separate cohorts of partici- based on those used in a recent study that examined
pants. The first cohort was the same set of 32 partici- both these ERP components (Kuperberg et al., 2020).
pants who participated in Study 1 at Massachusetts We operationalised the N400 as the average activity
General Hospital. In these participants, EEG data were between 300–500 ms across six centro-parietal elec-
acquired at the same time as the MEG data, using a trode sites (C3, Cz, C4, CP1, CPz, CP2). We operationa-
70-channel MEG-compatible scalp electrode system lised the P600 as the average activity between 600–
(BrainProducts, München), with an online reference at 1000 ms across six posterior electrode sites (P3, Pz, P4,
the left mastoid. Signals were digitised at 1000 Hz, O1, Oz, O2). For each component, we carried out a
with an online bandpass filter of 0.03 - 300 Hz. repeated-measures ANOVA with two within-subject vari-
The second EEG dataset was collected in a new cohort ables: Plausibility (plausible, anomalous) and Animacy
of 40 participants (19 females, mean age: 21.5 years; (animate, inanimate).
range: 18–32 years) at Tufts University, using a subset
of the scenarios used in Study 1 (160 out of 200). Spatial similarity analysis
Written consent was obtained from these participants, As in the MEG study, in each participant, at each time
based on guidelines from the Tufts University Social, point, we correlated the vectors that characterised the
Behavioural, and Educational Research Institutional spatial patterns of activity between every pair of trials
Review Board. These EEG data were acquired using within each of the four conditions, and then averaged
Biosemi ActiveTwo acquisition system from 32 active the resulting pairwise spatial similarity values to
electrodes using a modified 10/20 system montage. produce four spatial similarity time series. We then aver-
Signals were digitised at 512 Hz, with a bandpass of aged these spatial similarity values between 300–500 ms
DC-104 Hz. The EEG data were referenced offline to and between 600–1000 ms, and tested for significant
the average of the left and right mastoid channels. differences across conditions by carrying out 2 × 2
To maximise statistical power, we combined these repeated measures ANOVAs as described above. To visu-
two EEG datasets, yielding a dataset of 72 participants alise the EEG spatial similarity results and compare them
in total.5 with the ERPs, we constructed four grand-average
spatial similarity time series.
EEG data preprocessing
The two EEG datasets were preprocessed using Fieldtrip
Results
(Oostenveld et al., 2011). Because they were acquired
with different online filtering settings (0.03 - 300 Hz vs. ERP results
DC - 104 Hz), we applied an offline low-pass filter of Description of ERP time courses. As shown in Figure 2A,
30 Hz to the first EEG dataset and an offline band-pass the critical nouns evoked an N1/P1, which peaked at
filter of 0.1 - 30 Hz to the second EEG dataset. In ∼100 ms over fronto-central (N1) and posterior (P1) elec-
addition, because the two datasets were acquired with trode sites, followed by a P2 component, peaking at
different sampling rates (1000 Hz vs. 512 Hz), we ∼200 ms over fronto-central sites. These early com-
down-sampled both datasets to 500 Hz. In the first ponents were followed by the N400, which was larger
dataset, we removed, on average, seven “bad” EEG chan- to the anomalous than the plausible continuations, and
nels (out of the 70 channels), which were subsequently was visible between 300–500 ms over centro-parietal
interpolated using spherical spline interpolation (Perrin sites. Between 600–1000 ms, the anomalous nouns pro-
et al., 1989). duced a large posteriorly distributed positive-going slow
14 L. WANG AND G. R. KUPERBERG

Figure 2. Results of the EEG event-related and spatial similarity analyses (n = 72). A. Grand-average ERP waveforms following the
onset of critical nouns at a midline central electrode (Cz) and a midline posterior electrode (Oz). The duration of the critical nouns
(0 - 450 ms) and the subsequent word (550 - 1000 ms) are marked with gray bars on the x-axes. B. Topographic maps show the dis-
tributions of the effects of Plausibility separately for the animate and inanimate conditions between 300–500 ms (left and middle) and
for the collapsed animate and inanimate conditions between 600–1000 ms (right). The centro-parietal and posterior EEG electrodes
that showed significant differences are marked with white dots. C. Bar graphs showing the amplitude of the event-related potentials
(ERPs) in each condition, averaged between 300–500 ms across central electrode sites (left), and between 600–1000 ms across pos-
terior electrode sites (right). Error bars represent ±1 standard error of the mean. Between 300–500 ms, the anomalous nouns produced
a significantly larger (more negative-going) N400 than the plausible nouns. There was no significant difference in the amplitude of the
N400 evoked by the animate and inanimate nouns, in either the plausible or the anomalous condition. However, in the anomalous
condition, the animate nouns produced a smaller negativity than the inanimate nouns. As discussed in the main manuscript, we
suggest that this may reflect an earlier onset of the P600 produced by the animate (versus inanimate) anomalies. Between 600–
1000 ms, the anomalous nouns produced a larger P600 than the plausible nouns. The amplitude of the P600 did not differ significantly
between the animate and inanimate nouns in either the plausible or anomalous condition. D. Grand-average spatial similarity time
series following the onset of critical nouns in each of the four conditions. The duration of the critical nouns (0 - 450 ms) and the
subsequent word (550 - 1000 ms) are marked with gray bars on the x-axis. E. Bar graphs showing the mean EEG spatial similarity
values in each condition, averaged between 300–500 ms (left) and between 600–1000 ms (right). Error bars represent ±1 standard
error of the mean. Between 300–500 ms, the anomalous nouns produced larger spatial similarity values than the plausible nouns. In
the plausible condition, the animate nouns produced significantly larger spatial similarity values than the inanimate nouns. In the
anomalous condition, however, the spatial similarity values were significantly larger in the inanimate than the animate condition.
Between 600–1000 ms, the anomalous nouns produced significantly larger spatial similarity values than the plausible nouns. In
the plausible condition, the animate nouns produced larger spatial similarity values than the inanimate nouns, although this effect
only approached significance. In the anomalous condition, the spatial similarity values were significantly larger in the inanimate
than the animate condition.
 LANGUAGE, COGNITION AND NEUROSCIENCE 15

wave – the P600 effect. Within this late time window, the = 0.041, d = 0.25). On the plausible nouns, the spatial
P1 evoked by the word following the critical noun was similarity analysis detected an effect of Animacy that
visible at occipital electrodes, peaking at ∼650 ms. was not detected by the ERPs, with greater spatial simi-
larity values for the plausible animate than the plausible
300–500ms. As shown in Figure 2 (left), between 300– inanimate nouns (t(71) = 2.15, p = 0.035, d = 0.25).
500 ms, the N400 was larger to the anomalous than
the plausible nouns (Main effect of Plausibility: F(1,71) = 600–1000ms. Mirroring the late ERP effect, the spatial
132.65, p < 0.001, eta2 = 0.651). There was also a signifi- similarity values were larger for the anomalous than
cant interaction between Plausibility and Animacy the plausible nouns (Main effect of Plausibility: F(1,71) =
(F(1,71) = 15.91, p < 0.001, eta2 = 0.183). This interaction 27.75, p < 0.001, eta2 = 0.28). In addition, the spatial simi-
was driven by a smaller negativity to the anomalous larity analysis again detected effects of Animacy that
animate than the anomalous inanimate nouns (t(1,71) = were not detected by the event-related analysis.
5.41, p < 0.001, d = 0.64), but no difference between However, within this later 600–1000 ms time window,
the plausible animate and plausible inanimate nouns the effect of Animacy went in a different direction,
(t(1,71) = −0.27, p = 0.786, d = −0.03). depending on whether the nouns were plausible or
anomalous, reflected by a significant interaction
600–1000ms. Between 600–1000 ms, the anomalous between Plausibility and Animacy (F(1,71) = 14.10, p <
nouns produced a larger late posteriorly distributed 0.001, eta2 = 0.17). On the plausible nouns, the spatial
P600 than the plausible nouns (Main effect of Plausibility: similarity values were larger for the animate than for
F(1,71) = 108.48, p < 0.001, eta2 = 0.604). There was no inanimate nouns (although this effect only approached
main effect of Animacy (F(1,71) = 0.51, p = 0.479, eta2 = significance: t(71) = 1.63, p = 0.099, d = 0.20). For the
0.007), and no interaction between Plausibility and anomalies, however, inanimate nouns produced larger
Animacy (F(1,71) = 1.87, p = 0.176, eta2 = 0.026) in this spatial similarity values than animate nouns (t(71) =
time window. 3.99, p < 0.001, d = 0.47).

EEG spatial similarity results Follow-up exploratory analyses

Description of the EEG spatial similarity time series. As Just as for MEG, we carried out two additional analyses
shown in Figure 2D, the shape of the EEG spatial simi- to further explore the relationship between the univari-
larity time series mirrored that of the ERP waveforms, ate and multivariate effects of Animacy between 300–
with peaks at ∼100 ms (corresponding to the N1/P1), 500 ms after the onset of the plausible nouns.
at ∼200 ms (corresponding to the P200), and at First, to ensure that we did not miss any evoked effect
400 ms (corresponding to the N400). We also observed of Animacy over electrode sites that were not included
a slow rise and fall of spatial similarity values between in our a priori selected regions of interest, we tested
600–1000 ms, which corresponded to the slow rise and for an effect of Animacy on the mean evoked activity
fall of the P600 waveform. We note, however, that between 300–500 ms at all electrode sites, and used
whereas the polarity of the P600 ERP component is cluster-based permutation tests to control for multiple
opposite to that of the earlier N400 component, spatial comparisons (Maris & Oostenveld, 2007). This analysis
similarity values have no polarity, and so these values failed to reveal any spatial clusters showing a significant
are positive in both the 300–500 ms and 600–1000 ms event-related effect of Animacy.
time windows. Superimposed on the slow rise and fall Second, we tested for a spatial similarity effect of
of spatial similarity values between 600–1000 ms, one Animacy between 300–500 ms using only the subset of
can see a rise and fall of spatial similarity values that centro-parietal electrode sites that we used in the uni-
mirrors the early ERP components produced by the fol- variate analysis to examine the N400. In contrast to
lowing word. MEG, this analysis did not reveal a significant spatial
similarity effect (t(71) = 0.76, p = 0.45, d = 0.18).
300–500ms. Between 300–500 ms the spatial similarity
values were larger for the anomalous than the plausible
Discussion
nouns (Main effect of Plausibility: F(1,71) = 40.33, p <
0.001, eta2 = 0.362). In addition, mirroring the ERP This EEG study replicated the basic pattern of conver-
N400 results, Plausibility interacted with Animacy gence and divergence between the event-related and
(F(1,71) = 8.838, p = 0.004, eta2 = 0.11) due to smaller spatial similarity results that we observed in the MEG
spatial similarity values for the anomalous animate study (Study 1). First, the spatial similarity time series
than the anomalous inanimate nouns (t(71) = −2.084, p again picked up on the rise-and-fall morphology of
16 L. WANG AND G. R. KUPERBERG

both early and late event-related components (here, the techniques in the effects that they detected on the
N1/P1, P2, N400 and P600), and it again revealed effects anomalous nouns.
of Plausibility (anomalous > plausible) in both the 300–
500 ms and 600–1000 ms time windows. Second, the MEG was relatively insensitive to the P600 ERP effect.
spatial similarity analysis again revealed additional Perhaps the most striking difference between EEG and
effects of Animacy that were not detected by the MEG was in the magnitude of the late evoked effect pro-
event-related analysis: Between the 300–500 ms, the duced by the anomalous (versus plausible) nouns
spatial patterns produced by the plausible animate between 600–1000 ms. Consistent with many previous
nouns were again more similar to each other than studies (Brothers et al., 2020; Kuperberg, 2007; Kuperberg
those produced by the plausible inanimate nouns, mir- et al., 2003; Kuperberg et al., 2020; van de Meerendonk
roring their semantic similarity structures, although, as et al., 2009), EEG detected a robust positive-going P600
discussed further below, the effect of Animacy on effect within this time window, with an effect size of d
spatial similarity values went in the opposite direction = 1.23 at posterior electrode sites. In contrast, the MEG
on the anomalous nouns. evoked effect within this time window was much
In the Discussion of Study 1, we focused on why the smaller, with an effect size of d = 0.20 (estimated from
univariate and multivariate findings converged and the right frontal sensors that showed a significant
diverged from a methodological perspective. In this Dis- cluster in the mass univariate analysis). We also note
cussion section, we will focus on why the present EEG that because MEG gradiometer sensors do not carry infor-
findings converged and diverged from the previous mation about the polarity of the magnetic field (i.e.
MEG findings, also from a methodological perspective. whether the magnetic flux is leaving or entering the
We will discuss the theoretical significance of the full head), the late MEG effect was of the same polarity as
set of results in the General Discussion. the earlier N400 effect.
We suggest that the reason why EEG is so much
Convergence between the MEG and EEG results more sensitive than MEG to the P600 at the scalp
Both MEG and EEG techniques are sensitive to tangen- surface is that it is able to detect widespread sources
tially-oriented dipoles located within cortical sulci. There- within frontal and parietal cortices that may contribute
fore, the broad similarities between the MEG and EEG to this component. In principle, activity at the scalp
results are not surprising. Our finding that both MEG surface can potentially stem from two types of dipole
and EEG detect the well-known effect of Plausibility on sources within these fronto-parietal regions: Radially
the N400 replicates many previous studies (MEG: oriented dipoles, which are produced within the
Halgren et al., 2002; Helenius et al., 1998; Ihara et al., crowns of gyri and the bottom of sulci, and tangen-
2007; Maess et al., 2006; Wang et al., 2023; ERP: Kuperberg tially orientated dipoles, which are produced within
et al., 2020; Kutas & Hillyard, 1980; Nieuwland et al., 2020). the walls of different sulci.6 MEG, however, is relatively
Our finding that EEG was able to detect the same multi- blind to these sources. Radially orientated dipole
variate effect of Animacy as MEG is consistent with pre- sources typically do not result in a detectable magnetic
vious studies reporting a similar convergence between field (Cuffin & Cohen, 1979; Hämäläinen et al., 1993),
MEG and EEG multivariate effects (Cichy & Pantazis, and if two tangential dipoles face in the same direc-
2017; Wang et al., 2020). It is particularly encouraging tion, which is common in frontal and parietal cortices,
that EEG was able to detect a spatial similarity effect of they will each produce magnetic fields that cancel
Animacy despite the fact that there were fewer EEG elec- each other out (Ahlfors et al., 2010; Ahlfors et al.,
trodes (70 electrodes in 32 participants, and 32 electrodes 2010). In contrast, EEG is highly sensitive to both
in the remaining 40 participants) than MEG sensors (306 types of sources: in both cases, the electrical dipoles
sensors). We note, however, that in contrast to MEG, will summate to produce large electric fields (Nunez,
EEG failed to detect the multivariate effect of Animacy 1990).
between 300–500 ms when we used only a subset of A similar argument has been made for why scalp-
recording sites, probably because the EEG signal is recorded EEG is so much more sensitive than MEG to
more smeared than MEG due to the high resistance of the well-known P300 evoked response (the P3b;
the skull relative to the scalp (Geisler & Gerstein, 1961; Ahlfors et al., 2010; Siedenberg et al., 1996), which is
Grynszpan & Geselowitz, 1973). also thought to reflect widespread activity within
frontal and parietal regions (Soltani & Knight, 2000).
Divergence between the EEG and MEG results Indeed, several researchers have noted that the P600
In addition to the similarities between the EEG and MEG ERP component is functionally related to the P300
results, we saw some differences between the two (Coulson et al., 1998; Kuperberg et al., under review;
 LANGUAGE, COGNITION AND NEUROSCIENCE 17

Osterhout et al., 2012; Sassenhagen et al., 2014; Sassen- produced between 600–1000 ms had the same polarity
hagen & Fiebach, 2019). as the earlier N400, and (b) because MEG is relatively
These differences between EEG and MEG have poten- insensitive to the P600 effect, as discussed above (for
tial functional implications. When a linguistic anomaly further discussion of component overlap in EEG versus
cannot be integrated into a high-level interpretation, it MEG, see Wang et al., 2023, Supplementary Materials,
is thought to trigger multiple error-based processes Discussion Section 2).
within the 600–1000 ms time window (see Kuperberg
et al., under review). Thus, the differential sensitivity of On anomalous nouns, MEG and EEG produced spatial
MEG and EEG to distinct neuroanatomical sources similarity effects of Animacy that went in opposite
between 600–1000 ms means that the two techniques directions. When the critical nouns were plausible (i.e.
may also be differentially sensitive to distinct aspects when they matched the animacy-based constraints of
of the prolonged error-based response within this time the preceding verb), both MEG and EEG showed that
window. We elaborate further on this idea in the pairs of the animate nouns produced more similar
General Discussion. spatial patterns than pairs of the inanimate nouns, mir-
roring their semantic similarity structures (animate >
Between 300–500 ms, EEG but not MEG produced a inanimate). However, on the anomalous nouns, the
smaller evoked response to anomalous animate MEG and EEG spatial similarity effects of Animacy went
than anomalous inanimate nouns. We also saw differ- in the opposite direction: In MEG, the effect on the
ences between the EEG and MEG results in the earlier anomalies went in the same direction as on the plausible
300–500 ms time window. Similar to the event-related nouns (animate > inanimate). However, in EEG, the
MEG analysis, the ERP analysis showed that the anoma- spatial patterns were more similar for the inanimate
lous continuations produced a larger N400 than the anomalous than the animate anomalous nouns; that is,
plausible continuations. However, in contrast to MEG, instead of mirroring the similarity structure of animacy
but replicating a previous ERP study using a similar features that were encoded within the anomalous
design (Paczynski & Kuperberg, 2011, Experiment 1), bottom-up input itself (animate > inanimate), the effect
we observed an interaction between Plausibility and mirrored the semantic similarity structure of the
Animacy because the anomalous animate nouns (e.g. “ animacy features that were predicted, based on the
… clamped the *guests … ”) produced a smaller nega- animacy-based constraints of the prior verb (inanimate
tivity than the anomalous inanimate nouns (e.g. “ … > animate). The fact that EEG and MEG detected
greeted the *wires … ”). different spatial similarity effects of Animacy on the
We suggest that the smaller negativity produced by anomalous continuations is particularly notable in the
the animate anomalous nouns within the 300–500 ms late 600–100 ms window where there was no difference
time window reflected an early-starting positive-going in the amplitude of the evoked response produced by
P600, which was triggered because, in addition to violat- the animate and inanimate anomalous nouns.8 Once
ing the selectional constraints of the prior verb, the again, these findings suggest that the two techniques
animate anomalies also violated the comprehender’s may be differentially sensitive to distinct underlying
canonical expectations that inanimate direct object neural sources activated by the anomalies within this
noun-phrases should follow animate subject noun- late time window. We discuss the theoretical signifi-
phrases – expectations based on the so-called cance of this finding in the General Discussion.
“animacy hierarchy” (Paczynski & Kuperberg, 2011; Sil-
verstein, 1976). In EEG, this early-starting positivity-
General discussion
going P600 would have artificially reduced the ampli-
tude of the N400 produced by the anomalous animate We measured MEG and EEG as participants read dis-
nouns at the scalp surface as a result of component course scenarios with animate or inanimate critical
overlap; that is, because the N400 and late posterior words that were either plausible (matching the
positivity/P600 ERP components both have posterior animacy-based constraints of the prior verb) or anoma-
scalp distributions, if they overlap in time, their opposite lous (mismatching these constraints). We examined the
polarities will cancel each other out (Brouwer & Crocker, effects of both Plausibility and Animacy by carrying
2017; Kuperberg et al., 2007).7 This would explain why out both event-related and spatial similarity analyses
we saw no such interaction between Plausibility and on both datasets.
Animacy in MEG where spatiotemporal component The univariate event-related analyses revealed effects
overlap is less of an issue (a) because the magnetic of Plausibility (anomalous > plausible) both between
field does not have polarity, and so any ERF effect 300–500 ms (the N400) and in the later 600–1000 ms
18 L. WANG AND G. R. KUPERBERG

time window. The multivariate spatial similarity analyses animacy-based selectional constraints of the preceding
picked up on the same event-related waveforms and the verb to pre-activate upcoming semantic features
same effects of Plausibility. In addition, it detected (Szewczyk & Schriefers, 2013; Wang et al., 2020). For
effects of Animacy that were not detected by the example, the context “ … he greeted the … ” led to the
event-related analyses. pre-activation of semantic features associated with
We also saw convergence and divergence between <animate > entities (e.g. <can breathe>, <can move>).
MEG and EEG findings. Between 300–500 ms, the MEG It was therefore relatively easy to retrieve/access the
and EEG univariate and multivariate findings largely con- semantic features of plausible nouns that matched
verged. However, between 600–1000 ms, we saw differ- these prior predictions, resulting in a relatively small
ences between the effects detected by MEG and EEG in N400. However, when an anomalous noun was encoun-
the anomalous condition. tered, all these semantic features had to be retrieved de
In Study 1, we offered a detailed discussion of why novo, resulting in a larger N400 (Kuperberg et al., 2020;
the multivariate and the event-related measures con- Paczynski & Kuperberg, 2011, 2012; Szewczyk & Schrie-
verged and diverged from a methodological perspec- fers, 2013).
tive. In Study 2, we offered a detailed discussion of the
convergence and divergence between the MEG and The multivariate effect of Animacy between 300–
EEG findings, also from a methodological perspective. 500ms
In this General Discussion, we focus on the theoretical The effect of Animacy revealed by the multivariate simi-
interpretation of our findings. We argue that the full set larity analysis in this time window is more novel.
of results can be understood within a predictive Although previous MEG/EEG studies have shown that
coding framework in which animacy-based semantic representational similarity analysis is able to distinguish
features of incoming words are encoded between the animacy of visual objects (Carlson et al., 2013; Cichy
300–500 ms regardless of their predictability, but this et al., 2014; Cichy & Pantazis, 2017; Khaligh-Razavi et al.,
only produces prediction error and a larger univariate 2018), this is the first study to show that these methods
response when these features fail to match prior predic- can discriminate between animate and inanimate
tions. We suggest that, between 600–1000 ms, a failure incoming words during real-time language comprehen-
of the predictive coding algorithm to settle on an anom- sion. Crucially, our results also offer new insights into
alous interpretation triggers both reprocessing of the the underlying mechanisms and temporal dynamics of
bottom-up input, as well as a re-activation of prior this process.
predictions. First, we show that this multivariate distinction fol-
We conclude by considering the broader implications lowed the basic time course of the N400 event-related
of our findings, offering some general recommendations component. This provides additional evidence that
for how best to integrate these different analysis 300–500 ms provides a critical time window in which
approaches (univariate and multivariate) and techniques the perceptual features of incoming words first make
(MEG and EEG) to gain deeper insights into the neural contact with distributed knowledge stored within
basis of language comprehension. semantic memory (Federmeier & Laszlo, 2009; Kutas &
Federmeier, 2011).
Second, we show that the spatial patterns produced
Theoretical interpretation
by pairs of animate nouns were more similar than
The univariate N400 effect of Plausibility those produced by pairs of inanimate nouns. This is
The univariate effect of Plausibility on the N400 (anoma- important because it supports a longstanding proposal
lous > plausible) replicates many previous findings (e.g. in Cognitive Science that categorical distinctions might
MEG: Halgren et al., 2002; Helenius et al., 1998; Ihara emerge implicitly from differences in patterns of covaria-
et al., 2007; Maess et al., 2006; Wang et al., 2023; ERP: tion amongst the “microfeatures” that comprise
Kuperberg et al., 2020; Kutas & Hillyard, 1980; Nieuwland different concepts (Hinton et al., 1986; Kemp & Tenen-
et al., 2020). The N400 is thought to reflect the ease of baum, 2008; Rogers & McClelland, 2008), including
accessing/retrieving the semantic features associated animacy (Devlin et al., 1998; Gonnerman et al., 1997;
with new lexical inputs (Kuperberg, 2016; Kutas & Feder- Moss et al., 1998; Taylor et al., 2011). Specifically, a
meier, 2011). If these semantic features have already large body of previous work has established that pairs
been pre-activated by the prior context, then they will of animate words are more similar to one another and
be easier to retrieve and so the amplitude of the N400 have more semantic features in common than pairs of
will be reduced. In the present study, before observing inanimate words (Garrard et al., 2001; McRae et al.,
the critical noun, the brain was able to use the 1997; Randall et al., 2004; Zannino et al., 2006). For
 LANGUAGE, COGNITION AND NEUROSCIENCE 19

example, the two animate words, “guests” and “sailors”, The idea that these spatial patterns stemmed from
share many semantic features (e.g., <sentient>, <can widely distributed cortical regions would be consistent
move>, <can breathe>), while the two inanimate with evidence that animacy-based categorical structure
words, “wire” and “fruit”, are each associated with dis- can emerge from patterns of covariation across semantic
tinct semantic features (e.g. <straight > for “wire”, features within a widely distributed semantic system
<tasty > for “fruit”). Using Word2Vec (Mikolov et al., (Taylor et al., 2011; Tyler & Moss, 2001), including evi-
2013), we confirmed that this was the case for the dence for animacy-based categorisation deficits in
current stimulus set (see Study 1 Methods: Semantic patients with non-focal neuropathologies such as mid-
similarity between pairs of animate and inanimate stage Alzheimer’s disease (Gonnerman et al., 1997; see
nouns). On the assumption that the unique set of micro- also Devlin et al., 1998 for discussion).
features associated with each individual word is rep-
resented by a unique set of neuroanatomical sources Better together: synthesising the univariate and
that produce a unique spatial pattern of electric and multivariate results between 300–500ms
magnetic fields at the head surface (Wang et al., 2018), Taken together, therefore, the univariate and multi-
the spatial patterns produced by pairs of animate variate results complement each other. Both methods
words will be more similar than those produced by tell us that during sentence comprehension, the brain
pairs of inanimate words. encodes or accesses the semantic features of incoming
A spatial similarity analysis at the scalp surface words between 300–500 ms after their onset. The
cannot tell us which underlying sources produced the spatial similarity results tell us that this occurs regardless
effect. However, we know that within the N400 time of their predictability. The event-related results tell us
window, lexical representations, which are thought to that only the process of encoding unpredicted implausi-
be represented in multiple regions of the left temporal ble semantic features produces changes in the overall
cortex (Hirshorn et al., 2016; Lau et al., 2008; Wool- strength of neural activity (anomalous > plausible). We
nough et al., 2021), play a crucial role in mapping now need a computational theory that can explain
orthographic and phonological forms onto semantic both sets of findings.
features (Wang et al., 2023). These semantic features One such theory is predictive coding – a compu-
are thought to be represented within multimodal tational account of how the brain approximates Baye-
regions that are widely distributed across the cortex sian inference (Friston, 2005; Mumford, 1992; Rao &
(Huth et al., 2016; Martin & Chao, 2001). They may Ballard, 1999). According to predictive coding, as new
also be “bound together” as conceptual represen- bottom-up information reaches each level of the cortical
tations within the anterior and medial temporal lobe hierarchy, it activates two types of connectionist units –
(Lambon-Ralph et al., 2017), and there is some evi- “state units” and “error units”. State units that encode
dence that semantic features may be topographically contextual information at higher levels of the hierarchy
coded within these temporal regions (e.g. Chao et al., continually generate predictions that attempt to
1999; Haxby et al., 2001). We therefore speculate that “explain” expected information at the level below. Any
the scalp-recorded spatial patterns stemmed from new bottom-up information that matches these top-
both sources within the left temporal lobe as well as down predictions is suppressed within lower-level
more widespread cortical regions. error units, while any residual unpredicted information
The idea that these spatial patterns stemmed from that cannot be explained activates these error units, pro-
left temporal regions would be consistent with our ducing “prediction error” (Friston, 2005; Mumford, 1992;
finding in the MEG study that the same spatial similarity Rao & Ballard, 1999; Spratling, 2016). This lower-level
effect was produced when we repeated the analysis prediction error is then used to update the information
using only the subset of left temporal MEG sensors encoded within the higher-level state units, which, in
that reflect lexico-semantic activity between 300– turn, produce more accurate top-down predictions. By
500 ms within the left temporal cortex (Wang et al., minimising prediction error over multiple iterations of
2023). It would also be consistent with evidence that the predictive coding algorithm, the brain converges
patients with localised temporal lesions can show on the set of representations that best explain the
animacy-based category specific deficits (e.g. Warring- bottom-up input.
ton & McCarthy, 1987; Warrington & Shallice, 1984), Several researchers have suggested that the N400
and that the animacy of objects and words can be may reflect prediction error produced during predictive
decoded based on spatial patterns produced within coding (Bornkessel-Schlesewsky & Schlesewsky, 2019;
the left temporal lobe (Devereux et al., 2013; Fairhall & Kuperberg et al., 2020; Rabovsky & McRae, 2014; Xiang
Caramazza, 2013; Liuzzi et al., 2020). & Kuperberg, 2015). In recent work, we have used a
20 L. WANG AND G. R. KUPERBERG

predictive coding computational model to show that the raising the possibility that these two techniques were
effect of predictability on the N400, as well as its rise- differentially sensitive to different aspects of this late
and-fall morphology, can be simulated by the magni- error-based activity.
tude of prediction error produced by error units at Between 600–1000 ms, MEG detected only a small
lexical and semantic levels of representation, i.e. lexico- evoked response to the anomalies. In previous work,
semantic prediction error (Nour Eddine, 2021; Nour we source localised this late evoked response to lower-
Eddine et al., 2022).9 level cortical regions, providing evidence that it
To understand how this framework can explain both reflects bottom-up reprocessing of the anomalous input
the univariate and multivariate findings between 300– (Wang et al., 2023). Consistent with this idea, the MEG
500 ms in the present study, consider the state of the multivariate analysis revealed more similar spatial pat-
language system before the incoming critical word terns to the anomalous animate than inanimate nouns
was encountered. At this point, the brain will have pre- within this 600–1000 ms time window. As discussed
activated expected upcoming animacy-based semantic above, this mirrors the internal semantic similarity struc-
features, based on the animacy constraints of the prior ture of anomalous input. These findings therefore
verb. Consistent with this idea, we have previously suggest that the animacy-based semantic features
shown that when a verb constrains for upcoming associated with the anomalous bottom-up inputs were
animate nouns (e.g. “ … he greeted the … ”), the brain re-activated within this later time window, and that
produces spatial patterns that are more similar than MEG was sensitive to this late bottom-up activity.
when a verb constrains for upcoming inanimate nouns In contrast, between 600–1000 ms, the EEG detected
(e.g. “ … he clamped the … ”) (Wang et al., 2020). a much larger univariate ERP effect on the anomalies –
According to predictive coding, upon encountering a the posteriorly distributed semantic P600 ERP com-
plausible noun that confirms these prior predictions, ponent (Brothers et al., 2020; Brothers et al., 2022; Kuper-
these animacy-based features are reinstated between berg, 2007; Kuperberg et al., 2003; Kuperberg et al., 2020;
300–500 ms within state units. However, any activation Kuperberg et al., under review; van de Meerendonk et al.,
of error units (i.e. prediction error) is immediately sup- 2009). Moreover, within this later time window, the EEG
pressed by the prior predictions. This explains why we multivariate effect of Animacy on the anomalous nouns
found a spatial similarity effect of Animacy (animate > went in the opposite direction, with more similar pat-
inanimate) between 300–500 ms on the plausible critical terns observed in the anomalous inanimate than the
words (reflecting state activity), even though these anomalous animate condition; that is, within this later
words produced a relatively small event-related N400 time window, instead of mirroring the internal semantic
response (reflecting a relatively small prediction error similarity structure of observed anomalous animate and
within error units). In contrast, upon encountering an inanimate nouns, these patterns mirrored the semantic
anomalous noun whose animacy features cannot be similarity structure of the patterns that were expected,
explained by prior predictions, at the same time as based on the constraints of the prior verb (e.g. “ … he
state units are accessing the animacy features associated greeted the … <animate>”, “ … he clamped the …
with the bottom-up input, error units produce predic- <inanimate>”). These findings suggest that the incorrect
tion error because this newly encoded semantic infor- prior predictions were reactivated during this time
mation is not suppressed by prior predictions. This can window and that EEG was relatively more sensitive
explain why the anomalous inputs produced a larger than MEG to this late feedback activity.
N400 evoked response than the plausible inputs. The P600 ERP component itself may function to track
these conflicting sources of bottom-up and top-down
Responses to the anomalous nouns between 600– information in this late time window. Specifically, we
1000ms: divergence between MEG and EEG have recently argued that the P600 reflects a decision
findings variable that accumulates evidence that the original dis-
In the present study, the implausible nouns were not just ruption in comprehension stemmed from an external
unpredictable; they were anomalous, and therefore error in the bottom-up input, as opposed to an internal
could not be incorporated into the comprehender’s error in processing (Kuperberg et al., under review). This
high-level interpretation between 300–500 ms. When a type of confidence tracking may play an important role
linguistic error disrupts the process of comprehension, in ensuring that the brain responds optimally to linguis-
this can trigger a number of prolonged processes tic errors over multiple time scales. Similar evidence
between 600–1000 ms. We found that MEG and EEG accumulation decision-making mechanisms have been
diverged in both their univariate and multivariate linked to the well-known P300 ERP component (the
responses to the anomalies in this later time window, P3b or pre-decisional centro-parietal positivity; see
 LANGUAGE, COGNITION AND NEUROSCIENCE 21

Twomey et al., 2015), which, as noted earlier, shares implications. We offer three take-home messages, and
several functional properties with the P600 (Coulson provide recommendations for how to effectively use
et al., 1998; Osterhout et al., 2012; Sassenhagen et al., multivariate together with univariate methods, as well
2014; Sassenhagen & Fiebach, 2019).10 It is therefore par- as MEG together with EEG, to deepen our understanding
ticularly noteworthy that, like the P600, the P300 event- of language processing.
related response is larger in EEG than in MEG (Sieden-
berg et al., 1996). This may be because these decision- Interpret multivariate results in the light of the
making mechanisms are implemented within wide- large existing univariate ERP and MEG literatures
spread fronto-parietal cortices (Brosnan et al., 2020), Our first take-home message is that it is critical to inter-
and MEG is relatively insensitive to these sources pret any spatial similarity MEG/EEG results in the light of
(Ahlfors et al., 2010; see Study 2 for discussion). the large existing univariate ERP and MEG literatures.
This is because, as we show clearly in the present
Better together: synthesising the univariate and study, any multivariate measures estimated using Pear-
multivariate results of the MEG and EEG datasets son’s r will covary with the magnitude of well-known
between 600–1000ms event-related components like the N400 and P600.
Once again, therefore, the results of these different Perhaps because of the belief that newer multivariate
approaches complement each other. We again suggest measures detect “representational information” that is
that this full set of findings can be understood by functionally distinct from the “cognitive process”
appealing to the principles of predictive coding. In the detected by event-related responses (see further
case of anomalous inputs, however, the predictive below), some multivariate MEG/EEG language studies
coding algorithm would fail to converge between 300– make only minimal reference to the ERP/ERF literature.
500 ms, leading to two downstream consequences However, a failure to fully consider this literature not
within the later 600–1000 ms time window. only misses the opportunity to capitalise on a large
The first consequence is that higher levels of rep- body of existing knowledge, but can also lead to false
resentation will fail to produce accurate top-down pre- inferences about the functional significance of multi-
dictions that switch off lower-level prediction error. variate effects. For example, we know from the ERP lit-
This can explain why the anomalous bottom-up input erature that the N400 event-related component is
was reprocessed in this later 600–1000 ms time highly sensitive to numerous lexical and semantic vari-
window (see Wang et al., 2023 for discussion), as ables (e.g. word frequency: Rugg, 1990; Van Petten &
detected by the univariate and multivariate MEG Kutas, 1990; word concreteness or semantic richness:
findings. Amsel, 2011; Holcomb et al., 1999; Kounios & Holcomb,
The second consequence is that the prior context will 1994; Rabovsky et al., 2012). In addition, during sentence
continue to generate inaccurate top-down predictions comprehension, the N400 evoked by each incoming
based on the preceding verb. For example, in the anom- word is highly sensitive to its relationship with the
alous sentence, “ … he greeted the *wires … ”, the brain prior context, including its lexical predictability
would continue to generate predictions of <animate > (DeLong et al., 2005; Federmeier et al., 2007; Kutas & Hill-
features, while in the anomalous sentence, “ … he yard, 1984), its plausibility (Kuperberg et al., 2020;
clamped the *guests … ”, the brain would continue to Nieuwland et al., 2020), its semantic overlap with a incor-
generative predictions of <inanimate > features. This rectly predicted word (e.g. Federmeier & Kutas, 1999),
can explain why the EEG spatial similarity effect and its semantic association with prior words in the
reflected the internal semantic similarity structure of local context (e.g. Van Petten et al., 1997). All these
these incorrect predictions, rather than that of the event-related effects are likely to be detected by multi-
observed anomalous bottom-up anomalous inputs. As variate measures that use Pearson’s r. Therefore, when-
discussed above, the simultaneous activation of conflict- ever one sees a multivariate effect between 300–
ing top-down and bottom-up information within this 500 ms following stimulus onset, it is important to con-
time window would inform the evidence accumulation sider the possibility that it is driven by the effects of
process reflected by the P600 ERP component. these variables on the well-known N400 component.
As discussed in detail in the Discussion section of
Study 1, there is no simple way of capturing multivariate
General implications
effects that are not driven by univariate effects. We
Our findings have several broad implications that go therefore recommend that, when using Pearson corre-
beyond the specific set of results reported in this lations or Pearson distance measures as a similarity
study. In this final section, we summarise these metric, the best way to determine which multivariate
22 L. WANG AND G. R. KUPERBERG

effects are and are not driven by univariate effects is to related analysis that we learned that Plausibility but
carry out both a multivariate and univariate analysis on not Animacy modulated the amplitude of neural activity
the same dataset (see Haxby et al., 2001 and Aly & produced between 300–500 ms. This, in turn, provided
Turk-Browne, 2015, for similar recommendations using important constraints on our theoretical interpretation.
fMRI), ideally using analogous methods (as in the
present study), and qualitatively comparing the spatial Do not assume that MEG and EEG always reflect
similarity and evoked response time series. precisely the same underlying neural sources and
mechanisms
Abandon the assumption of a one-to-one mapping Our third take-home message is that we cannot assume
between analytic approach and functional that MEG and EEG results will always converge. In the
interpretation present study, this lack of convergence was particularly
Our second take-home message is that we should clear between 600–1000 ms (but, because of com-
abandon any assumption of a simple one-to-one ponent overlap in EEG but not MEG, it can also impact
mapping between analytic approach (univariate versus the earlier 300–500 ms time window, see Discussion in
multivariate) and functional interpretation (“cognitive Study 2; and see also Wang et al., 2023, Supplementary
process” versus “representational content”). This distinc- Materials, Discussion Section 2).
tion was first proposed in the fMRI literature (Krieges- Between 600–1000 ms, MEG and EEG differed in both
korte, 2011) where it was assumed that “processing” the univariate and multivariate effects in the anomalous
equated to the engagement of metabolically demand- condition, providing evidence that the two techniques
ing cognitive resources that operated upon “represen- were differentially sensitive to distinct neural sources
tational information”, which remained static over time. that were activated in this late time window. This in
While this assumption may sometimes be valid for turn, provided important theoretical constrains on our
interpreting the sluggish hemodynamic response interpretation of the prolonged response produced by
(Jimura & Poldrack, 2012), any clear-cut distinction linguistic errors. We suggested that MEG and EEG may
between “process” and “representational information” be differentially sensitive to the reprocessing of the
falls apart when we come to interpret MEG/EEG activity bottom-up input and the re-activation of prior predic-
produced during real-time language comprehension in tions respectively. If future studies confirm that EEG
which each new piece of information is rapidly and MEG are indeed differentially sensitive to feedback
encoded at multiple levels of linguistic representation, versus feedforward activity within this later time
and so, at each time point following stimulus onset, window, then this would provide essential information
the activation (encoding) of this information corre- that constrains the interpretation of future investi-
sponds to its processing. For example, as should be gations using both techniques to study language
clear from our discussion of the N400 component, comprehension.
within a predictive coding framework, the difficulty of
semantic “processing” (or “retrieval” or “access”)
between 300–500 ms corresponds directly to the Conclusion
amount of unpredicted semantic representational infor-
In conclusion, our findings suggest that a comprehen-
mation that is being encoded (see Kuperberg, 2016;
sive understanding of the neural and computational
Nour Eddine, 2021; Nour Eddine et al., 2023).
basis of language processing can only be achieved
We therefore suggest that when interpreting EEG/
through the combined use of univariate and multivariate
MEG data, we should begin with the assumption that
methods, along with different neuroimaging techniques.
both univariate and multivariate measures can tell us
Our results illustrate the importance of interpreting the
whether and when the brain encodes information
results of these different tools together to provide
within new bottom-up input, and that the critical ques-
deeper insights into the complex mechanisms under-
tion that we should be asking is whether encoding (or,
lying language comprehension.
equivalently accessing or retrieving) this information is,
or is not, accompanied by an overall increase in neural
signal. As we show in the present study, the answer to Notes
this question can only be provided by carrying out
1. In the MEG literature, the N400 is sometimes referred to
both types of analyses. Our spatial similarity results
as the M400 or N400m.
alone were unable to tell us which of the effects of 2. As noted under Methods, there were small differences
Plausibility or Animacy were accompanied by changes between the animate and inanimate nouns in their
in response magnitude. It was only through the event- length and frequency. Effects of word length can
 LANGUAGE, COGNITION AND NEUROSCIENCE 23

influence earlier ERP components (Hauk & Pulvermuller, However, because in the present investigation, Study 1
2004; Osterhout et al., 1997; Wydell et al., 2003), and showed that the amplitude of the MEG N400 was just
differences in frequency can modulate the N400 (Dam- as large to the anomalous animate as to the anomalous
bacher et al., 2006; Payne et al., 2015; Van Petten & inanimate nouns, we think that a component overlap
Kutas, 1990). These differences, however, are unlikely explanation for the attenuation of the ERP N400 to the
to explain the absence of a univariate effect of animate anomalous nouns is more likely.
Animacy between 300-500ms. 8. In the earlier 300-500ms time window, the smaller EEG
3. This may include (a) spontaneous ongoing neural spatial similarity values for the anomalous animate
activity, (b) activity stemming from other physiological (versus anomalous inanimate) nouns can be explained
sources, and (c) noise from the environment or from by the smaller evoked response to these animate con-
our measurement instruments (Luck, 2014b). We also tinuations, as a result of an early start of the P600 com-
note that while spatial similarity and event-related ana- ponent, as discussed above.
lyses are sensitive to many of the same sources of 9. In predictive coding, “prediction error” is computed as
noise, there are some differences: In an event-related an inherent component of the inference algorithm –
analysis, spontaneous neural activity that is not the process of inferring semantic features from a
phase-locked to an event of interest is always con- word’s linguistic form. This differs from how prediction
sidered “noise”. In a spatial similarity analysis, error has been simulated in other computational
however, non-phase-locked activity can, under certain models of the N400, where the error is computed
circumstances, contribute to increases in spatial simi- outside the model’s architecture (Fitz & Chang, 2019;
larity. For example, if the phase value of a waveform Rabovsky & McRae, 2014). It also differs from compu-
measured at channel 1 varies across trials, then the tational models that have simulated the N400 as the
event-related response at this channel will always be total activity (Cheyette & Plaut, 2017; Laszlo & Plaut,
small. However, if the relative differences in amplitude 2012) or the total change in activity produced by a
across the three channels (e.g. 1 < 2 < 3) are consistent single set of units (Brouwer et al., 2017; Rabovsky
across trials, then this will give rise to an increase in et al., 2018; see Nour Eddine et al., 2022 for discussion
spatial similarity. and review). We also note that, although in the
4. This is because without cross validation, estimates of the present study “prediction error” was produced by a lin-
true Euclidean distance become increasingly more inac- guistic “error” (an anomalous input), a linguistic error is
curate (larger) with increases in noise. As discussed not necessary to produce prediction error (or an N400)
above, this is also true of Pearson’s distance estimates. within a predictive coding framework. According to
However, the Pearson’s distance between two entirely this framework, prediction error is produced at the
random patterns has a maximal value of 1 (or, equiva- lexico-semantic level whenever incoming lexico-seman-
lently, a Pearson’s r value of 0), regardless of noise tic information cannot be explained by prior lexico-
level. In contrast, the Euclidean distance is non-negative semantic predictions. This will result in a larger N400
and unbounded. Therefore, without cross-validation, the to unexpected but plausible words, compared to
noise inflation of Euclidean distance is greater than that expected words. We have recently argued that, in
of the Pearson’s distance. Therefore, without cross-vali- the case of anomalous inputs, prediction error is
dation, Euclidean distance values are far less interpret- additionally generated at a higher event level of rep-
able than Pearson’s distance values (see Guggenmos resentation because the newly inferred anomalous
et al., 2018; Walther et al., 2016 for discussion). event cannot be explained by still higher-level predic-
5. We also carried out supplementary analyses to check tions based on longer-term real-world knowledge
that the event-related and spatial similarity effects (Wang et al., 2023).
were similar between the two EEG datasets. For both 10. Note, however, that unlike the P300, which is thought to
types of analyses, the pattern of findings across the index the accumulation of raw evidence to make an
two datasets was indeed qualitatively similar. initial decision, we have argued that the P600 reflects
6. Activity is also produced by tangentially orientated the accumulation of evidence that is gathered after
dipoles within the walls of a single sulcus. However, the initial disruption in comprehension. In this sense,
these dipoles frequently face in opposite directions, can- this component may bear more resemblance to later
celing each other out, and so they are not detected at positivities within the P300 family of ERP components
the head surface by either MEG or EEG. (Boldt & Yeung, 2015; Desender et al., 2019; Desender
7. We note that this interpretation is slightly different from et al., 2021; Murphy et al., 2015; Steinhauser & Yeung,
the one that we offered in our previous ERP study which 2010), which are also thought to track the brain’s confi-
used a similar design but a different set of stimuli (Pac- dence in relation to a previous choice (Boldt & Yeung,
zynski & Kuperberg, 2011, Experiment 1). In that study, 2015; Desender et al., 2019; Murphy et al., 2015; see
we also reported a smaller negativity to animate than Desender et al., 2021 for a review).
inanimate anomalies in the N400 time window, and we
also attributed this difference to the fact that the
animate but not the inanimate anomalies violated
Acknowledgments
expectations based on the animacy hierarchy. In that
study, we suggested that this led to reduced semantic We thank Trevor Brothers and Victoria Sharpe for helpful dis-
processing of the anomalous animate (versus anomalous cussion. We also thank Jeff Stibel for his support of Drs. Kuper-
inanimate) nouns on the N400 component itself. berg and Wang.
24 L. WANG AND G. R. KUPERBERG

Disclosure statement in Psychology, 8, 1327. https://doi.org/10.3389/fpsyg.2017.

01327
No potential conflict of interest was reported by the author(s). Brouwer, H., Crocker, M. W., Venhuizen, N. J., & Hoeks, J. C. J.
(2017). A neurocomputational model of the N400 and the
P600 in language processing. Cognitive Science, 41, 1318–
Funding 1352. https://doi.org/10.1111/cogs.12461
This work was supported by National Institutes of Health: Brysbaert, M., & New, B. (2009). Moving beyond Kučera and
[Grant Number R01HD082527]. Francis: A critical evaluation of current word frequency norms
and the introduction of a new and improved word frequency
measure for American English. Behavior Research Methods,
41(4), 977–990. https://doi.org/10.3758/BRM.41.4.977
References
Buzsáki, G., Anastassiou, C. A., & Koch, C. (2012). The origin of
Ahlfors, S. P., Han, J., Belliveau, J. W., & Hämäläinen, M. S. (2010). extracellular fields and currents — EEG, ECoG, LFP and
Sensitivity of MEG and EEG to source orientation. Brain spikes. Nature Reviews Neuroscience, 13(6), 407–420.
Topography, 23(3), 227–232. https://doi.org/10.1007/ https://doi.org/10.1038/nrn3241
s10548-010-0154-x Carlson, T., Tovar, D. A., Alink, A., & Kriegeskorte, N. (2013).
Ahlfors, S. P., Han, J., Lin, F. H., Witzel, T., Belliveau, J. W., Representational dynamics of object vision: The first 1000
Hämäläinen, M. S., & Halgren, E. (2010). Cancellation of ms. Journal of Vision, 13, https://doi.org/10.1167/13.10.1
EEG and MEG signals generated by extended and distribu- Chao, L. L., Haxby, J. V., & Martin, A. (1999). Attribute-based
ted sources. Human Brain Mapping, 31(1), 140–149. neural substrates in temporal cortex for perceiving and
Aly, M., & Turk-Browne, N. B. (2015). Attention stabilizes rep- knowing about objects. Nature Neuroscience, 2(10), 913–
resentations in the human hippocampus. Cerebral Cortex, 919. http://www.ncbi.nlm.nih.gov/entrez/query.fcgi?cmd=
26(2), 783–796. https://doi.org/10.1093/cercor/bhv041 Retrieve&db=PubMed&dopt=Citation&list_uids=10491613
Amsel, B. D. (2011). Tracking real-time neural activation of con- http://www.nature.com/neuro/journal/v2/n10/pdf/nn1099_
ceptual knowledge using single-trial event-related poten- 913.pdf. https://doi.org/10.1038/13217
tials. Neuropsychologia, 49(5), 970–983. https://doi.org/10. Cheyette, S. J., & Plaut, D. C. (2017). Modeling the N400 ERP
1016/j.neuropsychologia.2011.01.003 component as transient semantic over-activation within a
Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed- neural network model of word comprehension. Cognition,
effects modeling with crossed random effects for subjects 162, 153–166. https://doi.org/10.1016/j.cognition.2016.10.
and items. Journal of Memory and Language, 59(4), 390– 016
412. https://doi.org/10.1016/j.jml.2007.12.005 Cichy, R. M., & Pantazis, D. (2017). Multivariate pattern analysis
Bell, A. J., & Sejnowski, T. J. (1997). The “independent com- of MEG and EEG: A comparison of representational structure
ponents” of natural scenes are edge filters. Vision Research, in time and space. Neuroimage, 158, 441–454. https://doi.
37(23), 3327–3338. https://doi.org/10.1016/S0042-6989 org/10.1016/j.neuroimage.2017.07.023
(97)00121-1 Cichy, R. M., Pantazis, D., & Oliva, A. (2014). Resolving human
Boldt, A., & Yeung, N. (2015). Shared neural markers of decision object recognition in space and time. Nature Neuroscience,
confidence and error detection. The Journal of Neuroscience, 17(3), 455–462. https://doi.org/10.1038/nn.3635
35(8), 3478–3484. https://doi.org/10.1523/JNEUROSCI.0797- Clark, H. H. (1973). The language-as-fixed-effect fallacy: A cri-
14.2015 tique of language statistics in psychological research.
Bornkessel-Schlesewsky, I., & Schlesewsky, M. (2019). Toward a Journal of Verbal Learning and Verbal Behavior, 12(4), 335–
neurobiologically plausible model of language-related, 359. https://doi.org/10.1016/S0022-5371(73)80014-3
negative event-related potentials. Frontiers in Psychology, Coulson, S., King, J. W., & Kutas, M. (1998). Expect the unex-
10, 298. https://doi.org/10.3389/fpsyg.2019.00298 pected: Event-related brain response to morphosyntactic
Brosnan, M. B., Sabaroedin, K., Silk, T., Genc, S., Newman, D. P., violations. Language and Cognitive Processes, 13(1), 21–58.
Loughnane, G. M., Fornito, A., O’Connell, R. G., & Bellgrove, https://doi.org/10.1080/016909698386582
M. A. (2020). Evidence accumulation during perceptual Cuffin, B. N., & Cohen, D. (1979). Comparison of the
decisions in humans varies as a function of dorsal frontopar- magnetoencephalogram and electroencephalogram.
ietal organization. Nature Human Behaviour, 4(8), 844–855. Electroencephalography and Clinical Neurophysiology, 47(2),
https://doi.org/10.1038/s41562-020-0863-4 132–146. https://doi.org/10.1016/0013-4694(79)90215-3
Brothers, T., Wlotko, E. W., Warnke, L., & Kuperberg, G. R. (2020). Dambacher, M., Kliegl, R., Hofmann, M., & Jacobs, A. M. (2006).
Going the extra mile: Effects of discourse context on two late Frequency and predictability effects on event-related poten-
positivities during language comprehension. Neurobiology tials during reading. Brain Research, 1084(1), 89–103. https://
of Language, 1(1), 135–160. https://doi.org/10.1162/nol_a_ doi.org/10.1016/j.brainres.2006.02.010
00006 DeLong, K. A., Urbach, T. P., & Kutas, M. (2005). Probabilistic
Brothers, T., Zeitlin, M., Choi Perrachione, A., Choi, C., & word pre-activation during language comprehension
Kuperberg, G. (2022). Domain-general conflict monitoring inferred from electrical brain activity. Nature Neuroscience,
predicts neural and behavioral indices of linguistic error pro- 8(8), 1117–1121. https://doi.org/10.1038/nn1504
cessing during reading comprehension. Journal of Desender, K., Murphy, P., Boldt, A., Verguts, T., & Yeung, N.
Experimental Psychology: General, 151(7), 1502–1519. (2019). A post-decisional neural marker of confidence pre-
https://doi.org/10.1037/xge0001130 dicts information-seeking in decision-making. The Journal
Brouwer, H., & Crocker, M. W. (2017). On the proper treatment of Neuroscience, 39(17), 3309–3319. https://doi.org/10.
of the N400 and P600 in language comprehension. Frontiers 1523/JNEUROSCI.2620-18.2019
 LANGUAGE, COGNITION AND NEUROSCIENCE 25

Desender, K., Ridderinkhof, K. R., & Murphy, P. (2021). categories in Alzheimer’s disease. Brain and Language, 57
Understanding neural signals of post-decisional perform- (2), 254–279. https://doi.org/10.1006/brln.1997.1752
ance monitoring: An integrative review. Elife, 10, e67556. Gramfort, A., Luessi, M., Larson, E., Engemann, D. A., Strohmeier,
https://doi.org/10.7554/eLife.67556 D., Brodbeck, C., Parkkonen, L., & Hämäläinen, M. S. (2014).
Devereux, B. J., Clarke, A., Marouchos, A., & Tyler, L. K. (2013). MNE software for processing MEG and EEG data.
Representational similarity analysis reveals commonalities Neuroimage, 86, 446–460. https://doi.org/10.1016/j.neuro
and differences in the semantic processing of words and image.2013.10.027
objects. The Journal of Neuroscience, 33(48), 18906–18916. Grynszpan, F., & Geselowitz, D. B. (1973). Model studies of the
https://doi.org/10.1523/JNEUROSCI.3809-13.2013 magnetocardiogram. Biophysical Journal, 13(9), 911–925.
Devlin, J. T., Gonnerman, L. M., Andersen, E. S., & Seidenberg, M. http://www.ncbi.nlm.nih.gov/pmc/articles/PMC1484372/pdf/
S. (1998). Category-specific semantic deficits in focal and biophysj00707-0056.pdf. https://doi.org/10.1016/S0006-349
widespread brain damage: A computational account. 5(73)86034-5
Journal of Cognitive Neuroscience, 10(1), 77–94. https:// Guggenmos, M., Sterzer, P., & Cichy, R. M. (2018). Multivariate
www.ncbi.nlm.nih.gov/pubmed/9526084. https://doi.org/ pattern analysis for MEG: A comparison of dissimilarity
10.1162/089892998563798 measures. Neuroimage, 173, 434–447. https://doi.org/10.
Diedrichsen, J., Ridgway, G. R., Friston, K. J., & Wiestler, T. (2011). 1016/j.neuroimage.2018.02.044
Comparing the similarity and spatial structure of neural Hagoort, P. (2013). Muc (memory, unification, control) and
representations: A pattern-component model. Neuroimage, beyond. Frontiers in Psychology, 4, 416. https://doi.org/10.
55(4), 1665–1678. https://doi.org/10.1016/j.neuroimage. 3389/fpsyg.2013.00416
2011.01.044 Halgren, E., Dhond, R. P., Christensen, N., Van Petten, C.,
Dikker, S., Assaneo, M. F., Gwilliams, L., Wang, L., & Kösem, A. Marinkovic, K., Lewine, J. D., & Dale, A. M. (2002). N400-like
(2020). Magnetoencephalography and language. magnetoencephalography responses modulated by seman-
Neuroimaging Clinics of North America, 30(2), 229–238. tic context, word frequency, and lexical class in sentences.
https://doi.org/10.1016/j.nic.2020.01.004 Neuroimage, 17(3), 1101–1116. https://doi.org/10.1006/
Fairhall, S. L., & Caramazza, A. (2013). Brain regions that rep- nimg.2002.1268
resent amodal conceptual knowledge. Journal of Hämäläinen, M. S., Hari, R., Ilmoniemi, R. J., Knuutila, J. E. T., &
Neuroscience, 33(25), 10552–10558. https://doi.org/10.1523/ Lounasmaa, O. V. (1993). Magnetoencephalography—
JNEUROSCI.0051-13.2013 theory, instrumentation, and applications to noninvasive
Federmeier, K. D., & Kutas, M. (1999). A rose by any other name: studies of the working human brain. Reviews of Modern
Long-term memory structure and sentence processing. Physics, 65(2), 413–497. https://doi.org/10.1103/RevMod
Journal of Memory and Language, 41(4), 469–495. https:// Phys.65.413
doi.org/10.1006/jmla.1999.2660 Hauk, O., & Pulvermuller, F. (2004). Effects of word length and
Federmeier, K. D., & Laszlo, S. (2009). Psychology of learning frequency on the human event-related potential. Clinical
and motivation. Psychology of Learning and Motivation, 51, Neurophysiology, 115(5), 1090–1103. https://doi.org/10.
1–44. https://doi.org/10.1016/S0079-7421(09)51001-8 1016/j.clinph.2003.12.020
Federmeier, K. D., Wlotko, E. W., De Ochoa-Dewald, E., & Kutas, Haxby, J. V., Gobbini, M. I., Furey, M. L., Ishai, A., Schouten, J. L.,
M. (2007). Multiple effects of sentential constraint on word & Pietrini, P. (2001). Distributed and overlapping represen-
processing. Brain Research, 1146, 75–84. https://doi.org/10. tations of faces and objects in ventral temporal cortex.
1016/j.brainres.2006.06.101 Science, 293(5539), 2425–2430. https://doi.org/10.1126/
Fitz, H., & Chang, F. (2019). Language ERPs reflect learning science.1063736
through prediction error propagation. Cognitive Psychology, Hebart, M. N., & Baker, C. I. (2018). Deconstructing multivariate
111, 15–52. https://doi.org/10.1016/j.cogpsych.2019.03.002 decoding for the study of brain function. Neuroimage, 180,
Friston, K. J. (2005). A theory of cortical responses. Philosophical 4–18. https://doi.org/10.1016/j.neuroimage.2017.08.005
Transactions of the Royal Society B: Biological Sciences, 360 Helenius, P., Salmelin, R., Service, E., & Connolly, J. (1998).
(1456), 815–836. https://doi.org/10.1098/rstb.2005.1622 Distinct time courses of word and context comprehension
Garrard, P., Lambon-Ralph, M. A., Hodges, J. R., & Patterson, K. in the left temporal cortex. Brain, 121(6), 1133–1142.
(2001). Prototypicality, distinctiveness, and intercorrelation: https://doi.org/10.1093/brain/121.6.1133
Analyses of the semantic attributes of living and nonliving Hinton, G. E., McClelland, J. L., & Rumelhart, D. E. (1986).
concepts. Cognitive Neuropsychology, 18(2), 125–174. Distributed representations. In D. E. Rumelhart, J. L.
https://doi.org/10.1080/02643290125857 McClelland, & PDP Research Group (Eds.), Parallel distributed
Garrido, L., Vaziri-Pashkam, M., Nakayama, K., & Wilmer, J. processing: Explorations in the microstructure of cognition,
(2013). The consequences of subtracting the mean pattern Vol. 1: Foundations (pp. 77–109). MIT Press.
in fMRI multivariate correlation analyses. Frontiers in Hirshorn, E. A., Li, Y., Ward, M. J., Richardson, R. M., Fiez, J. A., &
Neuroscience, 7, 174. https://doi.org/10.3389/fnins.2013. Ghuman, A. S. (2016). Decoding and disrupting left midfusi-
00174 form gyrus activity during word reading. Proceedings of the
Geisler, C. D., & Gerstein, G. L. (1961). The surface EEG in relation National Academy of Sciences, 113(29), 8162–8167. https://
to its sources. Electroencephalography and Clinical doi.org/10.1073/pnas.1604126113
Neurophysiology, 13(6), 927–934. https://doi.org/10.1016/ Holcomb, P. J., Kounios, J., Anderson, J. E., & West, W. C. (1999).
0013-4694(61)90199-7 Dual-coding, context-availability, and concreteness effects
Gonnerman, L. M., Andersen, E. S., Devlin, J. T., Kempler, D., & in sentence comprehension: An electrophysiological investi-
Seidenberg, M. S. (1997). Double dissociation of semantic gation. Journal of Experimental Psychology: Learning,
26 L. WANG AND G. R. KUPERBERG

Memory, and Cognition, 25(3), 721–742. https://doi.org/10. Kuperberg, G. R. (2016). Separate streams or probabilistic infer-
1037/0278-7393.25.3.721 ence? What the N400 can tell US about the comprehension
Huth, A. G., de Heer, W. A., Griffiths, T. L., Theunissen, F. E., & of events. Language, Cognition and Neuroscience, 31(5), 602–
Gallant, J. L. (2016). Natural speech reveals the semantic 616. https://doi.org/10.1080/23273798.2015.1130233
maps that tile human cerebral cortex. Nature, 532(7600), Kuperberg, G. R., Alexander, E., & Brothers, T. (under review).
453–458. https://doi.org/10.1038/nature17637 The posterior P600 does not reflect error correction: An
Ihara, A., Hayakawa, T., Wei, Q., Munetsuna, S., & Fujimaki, N. information seeking account of linguistic error processing.
(2007). Lexical access and selection of contextually Kuperberg, G. R., Brothers, T., & Wlotko, E. (2020). A tale of
appropriate meaning for ambiguous words. Neuroimage, two positivities and the N400: Distinct neural signatures
38(3), 576–588. https://doi.org/10.1016/j.neuroimage.2007. are evoked by confirmed and violated predictions at
07.047 different levels of representation. Journal of Cognitive
Jimura, K., & Poldrack, R. A. (2012). Analyses of regional-average Neuroscience, 32(1), 12–35. https://doi.org/10.1162/jocn_a_
activation and multivoxel pattern information tell comp- 01465
lementary stories. Neuropsychologia, 50(4), 544–552. Kuperberg, G. R., Kreher, D. A., Sitnikova, T., Caplan, D. N., &
https://doi.org/10.1016/j.neuropsychologia.2011.11.007 Holcomb, P. J. (2007). The role of Animacy and thematic
Jung, T. P., Makeig, S., Humphries, C., Lee, T. W., McKeown, M. J., relationships in processing active English sentences:
Iragui, V., & Sejnowski, T. J. (2000). Removing electroence- Evidence from event-related potentials. Brain and
phalographic artifacts by blind source separation. Language, 100(3), 223–237. https://doi.org/10.1016/j.bandl.
Psychophysiology, 37(2), 163–178. https://doi.org/10.1111/ 2005.12.006
1469-8986.3720163 Kuperberg, G. R., Sitnikova, T., Caplan, D., & Holcomb, P. J.
Kappenman, E. S., & Luck, S. J. (2012). ERP components: The ups (2003). Electrophysiological distinctions in processing con-
and downs of brainwave recordings. In S. J. Luck, & E. S. ceptual relationships within simple sentences. Cognitive
Kappenman (Eds.), The Oxford handbook of event-related Brain Research, 17(1), 117–129. https://doi.org/10.1016/
potential components (pp. 3–30). Oxford University Press. S0926-6410(03)00086-7
Karimi-Rouzbahani, H., Ramezani, F., Woolgar, A., Rich, A., & Kutas, M., & Federmeier, K. D. (2011). Thirty years and counting:
Ghodrati, M. (2021). Perceptual difficulty modulates the Finding meaning in the N400 component of the event-
direction of information flow in familiar face recognition. related brain potential (ERP). Annual Review of Psychology,
Neuroimage, 233, 117896. https://doi.org/10.1016/j. 62(1), 621–647. https://doi.org/10.1146/annurev.psych.
neuroimage.2021.117896 093008.131123
Kemp, C., & Tenenbaum, J. B. (2008). Structured models of Kutas, M., & Hillyard, S. A. (1980). Reading senseless sentences:
semantic cognition. Behavioral and Brain Sciences, 31(6), Brain potentials reflect semantic incongruity. Science, 207
717–718. https://doi.org/10.1017/S0140525X08005931 (4427), 203–205. https://doi.org/10.1126/science.7350657
Khaligh-Razavi, S. M., Cichy, R. M., Pantazis, D., & Oliva, A. Kutas, M., & Hillyard, S. A. (1984). Brain potentials during
(2018). Tracking the spatiotemporal neural dynamics of reading reflect word expectancy and semantic association.
real-world object size and Animacy in the human brain. Nature, 307(5947), 161–163. https://doi.org/10.1038/
Journal of Cognitive Neuroscience, 30(11), 1559–1576. 307161a0
https://doi.org/10.1162/jocn_a_01290 Lambon-Ralph, M. A., Jefferies, E., Patterson, K., & Rogers, T. T.
Kounios, J., & Holcomb, P. J. (1994). Concreteness effects in (2017). The neural and computational bases of semantic
semantic processing: ERP evidence supporting dual-coding cognition. Nature Reviews Neuroscience, 18(1), 42–55.
theory. Journal of Experimental Psychology: Learning, https://doi.org/10.1038/nrn.2016.150
Memory, and Cognition, 20(4), 804–823. https://doi.org/10. LaRocque, K. F., Smith, M. E., Carr, V. A., Witthoft, N., Grill-
1037/0278-7393.20.4.804 Spector, K., & Wagner, A. D. (2013). Global similarity and
Kriegeskorte, N. (2011). Pattern-information analysis: From pattern separation in the human medial temporal lobe
stimulus decoding to computational-model testing. predict subsequent memory. The Journal of Neuroscience,
Neuroimage, 56(2), 411–421. https://doi.org/10.1016/j. 33(13), 5466–5474. https://doi.org/10.1523/JNEUROSCI.
neuroimage.2011.01.061 4293-12.2013
Kriegeskorte, N., & Bandettini, P. (2007). Analyzing for infor- Laszlo, S., & Plaut, D. C. (2012). A neurally plausible parallel dis-
mation, not activation, to exploit high-resolution fMRI. tributed processing model of event-related potential word
Neuroimage, 38(4), 649–662. https://doi.org/10.1016/j. reading data. Brain and Language, 120(3), 271–281. https://
neuroimage.2007.02.022 doi.org/10.1016/j.bandl.2011.09.001
Kriegeskorte, N., Mur, M., & Bandettini, P. (2008). How does Lau, E. F., Gramfort, A., Hämäläinen, M. S., & Kuperberg, G. R.
nature program neuron types? Frontiers in Neuroscience, 2 (2013). Automatic semantic facilitation in anterior temporal
(1), 4. https://doi.org/10.3389/neuro.01.016.2008 cortex revealed through multimodal neuroimaging. The
Kriegeskorte, N., Mur, M., Ruff, D. A., Kiani, R., Bodurka, J., Esteky, Journal of Neuroscience, 33(43), 17174–17181. https://doi.
H., Tanaka, K., & Bandettini, P. A. (2008). Matching categori- org/10.1523/JNEUROSCI.1018-13.2013
cal object representations in inferior temporal cortex of man Lau, E. F., Phillips, C., & Poeppel, D. (2008). A cortical network for
and monkey. Neuron, 60(6), 1126–1141. https://doi.org/10. semantics: (De)constructing the N400. Nature Reviews
1016/j.neuron.2008.10.043 Neuroscience, 9(12), 920–933. https://doi.org/10.1038/
Kuperberg, G. R. (2007). Neural mechanisms of language nrn2532
comprehension: Challenges to syntax. Brain Research, 1146, Liuzzi, A. G., Aglinskas, A., & Fairhall, S. L. (2020). General and
23–49. https://doi.org/10.1016/j.brainres.2006.12.063 feature-based semantic representations in the semantic
 LANGUAGE, COGNITION AND NEUROSCIENCE 27

network. Scientific Reports, 10(1), 8931. https://doi.org/10. Nour Eddine, S., Brothers, T., Wang, L., Spratling, M., &
1038/s41598-020-65906-0 Kuperberg, G. R. (2023). A predictive coding model of the
Luck, S. J. (2014a). Chapter 8: Baseline correction, averaging, N400. bioRxiv, 04.
and time-frequency analysis. In S.J. Luck (Ed.), An introduc- Nunez, P. L. (1990). Localization of brain activity with electroen-
tion to the event-related potential technique (2 ed., pp. 249– cephalography. In S. Sato (Ed.), Magnetoencephalography
282). MIT Press. (advances in neurology, Vol. 54) (Vol. 54, pp. 39–65). Raven
Luck, S. J. (2014b). Chapter 6: Artifact rejection and correction. Press.
In S.J. Luck (Ed.), An introduction to the event-related potential Oostenveld, R., Fries, P., Maris, E., & Schoffelen, J.-M. (2011).
technique (2nd ed, pp. 185–218). MIT Press. FieldTrip: Open source software for advanced analysis of
Maess, B., Herrmann, C. S., Hahne, A., Nakamura, A., & Friederici, MEG, EEG, and invasive electrophysiological data.
A. D. (2006). Localizing the distributed language network Computational Intelligence and Neuroscience, 201, 1.
responsible for the N400 measured by MEG during auditory https://doi.org/10.1155/2011/156869
sentence processing. Brain Research, 1096(1), 163–172. Op de Beeck, H. P., Baker, C. I., DiCarlo, J. J., & Kanwisher, N. G.
https://doi.org/10.1016/j.brainres.2006.04.037 (2006). Discrimination training alters object representations
Maris, E., & Oostenveld, R. (2007). Nonparametric statistical in human extrastriate cortex. The Journal of Neuroscience, 26
testing of EEG- and MEG-data. Journal of Neuroscience (50), 13025–13036. https://doi.org/10.1523/JNEUROSCI.
Methods, 164(1), 177–190. https://doi.org/10.1016/j.jneu 2481-06.2006
meth.2007.03.024 Osterhout, L., Bersick, M., & McKinnon, R. (1997). Brain poten-
Martin, A., & Chao, L. L. (2001). Semantic memory and the brain: tials elicited by words: Word length and frequency predict
Structure and processes. Current Opinion in Neurobiology, 11 the latency of an early negativity. Biological Psychology, 46
(2), 194–201. http://www.ncbi.nlm.nih.gov/entrez/query. (2), 143–168. http://www.ncbi.nlm.nih.gov/entrez/query.
fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids= fcgi?cmd=Retrieve&db=PubMed&dopt=Citation&list_uids=
11301239. https://doi.org/10.1016/S0959-4388(00)00196-3 9288411. https://doi.org/10.1016/S0301-0511(97)05250-2
McRae, K., de Sa, V. R., & Seidenberg, M. S. (1997). On the nature Osterhout, L., Kim, A., & Kuperberg, G. R. (2012). The neurobiol-
and scope of featural representations of word meaning. ogy of sentence comprehension. In M. Spivey, M. Joannisse,
Journal of Experimental Psychology: General, 126(2), 99–130. & K. McRae (Eds.), The Cambridge handbook of psycholinguis-
https://doi.org/10.1037/0096-3445.126.2.99 tics (pp. 365–389). Cambridge University Press.
Mikolov, T., Chen, K., Corrado, G. S., & Dean, J. (2013). Efficient Paczynski, M., & Kuperberg, G. R. (2011). Electrophysiological
estimation of word representations in vector space. 1st evidence for use of the animacy hierarchy, but not thematic
International Conference on Learning Representations role assignment, during verb-argument processing.
(ICLR), workshop track proceedings, Scottsdale, Arizona. Language and Cognitive Processes, 26(9), 1402–1456.
Moss, H. E., Tyler, L. K., Durrant-Peatfield, M., & Bunn, E. M. https://doi.org/10.1080/01690965.2011.580143
(1998). ‘Two eyes of a see-through’: Impaired and intact Paczynski, M., & Kuperberg, G. R. (2012). Multiple influences of
semantic knowledge in a case of selective deficit for living semantic memory on sentence processing: Distinct effects
things. Neurocase, 4(4-5), 291–310. https://doi.org/10.1080/ of semantic relatedness on violations of real-world event/
13554799808410629 state knowledge and animacy selection restrictions.
Mumford, D. (1992). On the computational architecture of the Journal of Memory and Language, 67(4), 426–448. https://
neocortex. Biological Cybernetics, 66(3), 241–251. https://doi. doi.org/10.1016/j.jml.2012.07.003
org/10.1007/BF00198477 Payne, B. R., Lee, C. L., & Federmeier, K. D. (2015). Revisiting
Murphy, P. R., Robertson, I. H., Harty, S., & O’Connell, R. G. the incremental effects of context on word processing:
(2015). Neural evidence accumulation persists after choice Evidence from single-word event-related brain potentials.
to inform metacognitive judgments. Elife, 4, e11946. Psychophysiology, 52(11), 1456–1469. https://doi.org/10.
https://doi.org/10.7554/eLife.11946 1111/psyp.12515
Nieuwland, M. S., Barr, D. J., Bartolozzi, F., Busch-Moreno, S., Perrin, F., Pernier, J., Bertrand, O., & Echallier, J. F. (1989).
Darley, E., Donaldson, D. I., Ferguson, H. J., Fu, X., Spherical splines for scalp potential and current density
Heyselaar, E., Huettig, F., & Matthew Husband, E. (2020). mapping. Electroencephalography and Clinical
Dissociable effects of prediction and integration during Neurophysiology, 72(2), 184–187. https://doi.org/10.1016/
language comprehension: Evidence from a large-scale 0013-4694(89)90180-6
study using brain potentials. Philosophical Transactions of Proklova, D., Kaiser, D., & Peelen, M. V. (2016). Disentangling
the Royal Society B: Biological Sciences, 375(1791), representations of object shape and object category in
20180522. https://doi.org/10.1098/rstb.2018.0522 human visual cortex: The animate-inanimate distinction.
Nili, H., Wingfield, C., Walther, A., Su, L., Marslen-Wilson, W., & Journal of Cognitive Neuroscience, 28(5), 680–692. https://
Kriegeskorte, N. (2014). A toolbox for representational simi- doi.org/10.1162/jocn_a_00924
larity analysis. PLoS Computational Biology, 10(4), e1003553. Rabovsky, M., Hansen, S. S., & McClelland, J. L. (2018). Modelling
https://doi.org/10.1371/journal.pcbi.1003553 the N400 brain potential as change in a probabilistic rep-
Nour Eddine, S. (2021). Divide and Concur: A predictive coding resentation of meaning. Nature Human Behaviour, 2(9),
account of the N400 ERP component [Doctoral dissertation]. 693–705. https://doi.org/10.1038/s41562-018-0406-4
Tufts University. Rabovsky, M., & McRae, K. (2014). Simulating the N400 ERP
Nour Eddine, S., Brothers, T., & Kuperberg, G. R. (2022). The component as semantic network error: Insights from a
N400 in silico: A review of computational models. In K. feature-based connectionist attractor model of word
Federmeier (Ed.), Psychology of learning and motivation meaning. Cognition, 132(1), 68–89. https://doi.org/10.1016/
(Vol. 76, pp. 123–206). Academic Press. j.cognition.2014.03.010
28 L. WANG AND G. R. KUPERBERG

Rabovsky, M., Sommer, W., & Abdel Rahman, R. (2012). The time Swaab, T. Y., Ledoux, K., Camblin, C. C., & Boudewyn, M. A.
course of semantic richness effects in visual word recog- (2012). Language-related ERP components. In S. J. Luck, &
nition. Frontiers in Human Neuroscience, 6, 11. https://doi. E. S. Kappenman (Eds.), The Oxford handbook of event-
org/10.3389/fnhum.2012.00011 related potential components (pp. 397–439). Oxford
Randall, B., Moss, H. E., Rodd, J. M., Greer, M., & Tyler, L. K. University Press.
(2004). Distinctiveness and correlation in conceptual struc- Szewczyk, J. M., & Schriefers, H. (2013). Prediction in language
ture: Behavioral and computational studies. Journal of comprehension beyond specific words: An ERP study on
Experimental Psychology: Learning, Memory, and Cognition, sentence comprehension in Polish. Journal of Memory and
30(2), 393–406. https://doi.org/10.1037/0278-7393.30.2.393 Language, 68(4), 297–314. https://doi.org/10.1016/j.jml.
Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the 2012.12.002
visual cortex: A functional interpretation of some extra-clas- Tarkiainen, A., Helenius, P., Hansen, P. C., Cornelissen, P. L., &
sical receptive-field effects. Nature Neuroscience, 2(1), 79–87. Salmelin, R. (1999). Dynamics of letter string perception in
https://doi.org/10.1038/4580 the human occipitotemporal cortex. Brain, 122(11), 2119–
Rogers, T. T., & McClelland, J. L. (2008). Précis of semantic cog- 2132. https://doi.org/10.1093/brain/122.11.2119
nition: A parallel distributed processing approach. Taylor, K. I., Devereux, B. J., & Tyler, L. K. (2011). Conceptual
Behavioral and Brain Sciences, 31(6), 689–714. https://doi. structure: Towards an integrated neuro-cognitive account.
org/10.1017/S0140525X0800589X Language and Cognitive Processes, 26(9), 1368–1401.
Rugg, M. D. (1990). Event-related brain potentials dissociate rep- https://doi.org/10.1080/01690965.2011.568227
etition effects of high-and low-frequency words. Memory & Twomey, D. M., Murphy, P. R., Kelly, S. P., & O’Connell, R. G.
Cognition, 18(4), 367–379. https://doi.org/10.3758/BF03197 (2015). The classic P300 encodes a build-to-threshold
126 decision variable. European Journal of Neuroscience, 42(1),
Sassenhagen, J., & Fiebach, C. J. (2019). Finding the P3 in the 1636–1643. https://doi.org/10.1111/ejn.12936
P600: Decoding shared neural mechanisms of responses to Tyler, L. K., & Moss, H. E. (2001). Towards a distributed account
syntactic violations and oddball targets. Neuroimage, 200, of conceptual knowledge. Trends in Cognitive Sciences, 5(6),
425–436. https://doi.org/10.1016/j.neuroimage.2019.06.048 244–252. http://www.ncbi.nlm.nih.gov/pubmed/11390295.
Sassenhagen, J., Schlesewsky, M., & Bornkessel-Schlesewsky, I. https://doi.org/10.1016/S1364-6613(00)01651-X
(2014). The P600-as-P3 hypothesis revisited: Single-trial ana- Uusitalo, M. A., & Ilmoniemi, R. J. (1997). Signal-space projec-
lyses reveal that the late EEG positivity following linguistically tion method for separating MEG or EEG into components.
deviant material is reaction time aligned. Brain and Language, Medical & Biological Engineering & Computing, 35(2), 135–
137, 29–39. https://doi.org/10.1016/j.bandl.2014.07.010 140. https://doi.org/10.1007/BF02534144
Sha, L., Haxby, J. V., Abdi, H., Guntupalli, J. S., Oosterhof, N. N., Van Berkum, J. J. A. (2009). The neuropragmatics of
Halchenko, Y. O., & Connolly, A. C. (2015). The animacy con- ‘simple’ utterance comprehension: An ERP review. In U.
tinuum in the human ventral vision pathway. Journal of Sauerland, & K. Yatsushiro (Eds.), Semantics and pragmatics:
Cognitive Neuroscience, 27(4), 665–678. https://doi.org/10. From experiment to theory (pp. 276–316). Palgrave
1162/jocn_a_00733 Macmillan.
Siedenberg, R., Goodin, D. S., Aminoff, M. J., Rowley, H. A., & van de Meerendonk, N., Kolk, H. H. J., Chwilla, D. J., & Vissers,
Roberts, T. P. L. (1996). Comparison of late components in C. T. W. M. (2009). Monitoring in language perception.
simultaneously recorded event-related electrical potentials Language and Linguistics Compass, 3(5), 1211–1224.
and event-related magnetic fields. Electroencephalography https://doi.org/10.1111/j.1749-818X.2009.00163.x
and Clinical Neurophysiology, 99(2), 191–197. https://doi. Van Petten, C., & Kutas, M. (1990). Interactions between sen-
org/10.1016/0013-4694(96)95215-3 tence context and word frequencyinevent-related brainpo-
Silverstein, M. (1976). Hierarchy of features and ergativity. In tentials. Memory & Cognition, 18(4), 380–393. https://doi.
R. M. W. Dixon (Ed.), Grammatical categories in Australian org/10.3758/BF03197127
languages (pp. 112–171). Australian Institute of Aboriginal Van Petten, C., & Luka, B. J. (2012). Prediction during language
Studies. comprehension: Benefits, costs, and ERP components.
Smith, N. J., & Kutas, M. (2015). Regression-based estimation of International Journal of Psychophysiology, 83(2), 176–190.
ERP waveforms: I. The rERP framework. Psychophysiology, 52 https://doi.org/10.1016/j.ijpsycho.2011.09.015
(2), 157–168. https://doi.org/10.1111/psyp.12317 Van Petten, C., Weckerly, J., McIsaac, H. K., & Kutas, M. (1997).
Soltani, M., & Knight, R. T. (2000). Neural origins of the P300. Working memory capacity dissociates lexical and sentential
Critical Reviews™ in Neurobiology, 14(3-4), 26. https://doi. context effects. Psychological Science, 8(3), 238–242. https://
org/10.1615/CritRevNeurobiol.v14.i3-4.20 doi.org/10.1111/j.1467-9280.1997.tb00418.x
Spratling, M. W. (2016). Predictive coding as a model of cogni- Walther, A., Nili, H., Ejaz, N., Alink, A., Kriegeskorte, N., &
tion. Cognitive Processing, 17(3), 279–305. https://doi.org/10. Diedrichsen, J. (2016). Reliability of dissimilarity measures
1007/s10339-016-0765-6 for multi-voxel pattern analysis. Neuroimage, 137, 188–200.
Steinhauser, M., & Yeung, N. (2010). Decision processes in https://doi.org/10.1016/j.neuroimage.2015.12.012
human performance monitoring. The Journal of Wang, L., Hagoort, P., & Jensen, O. (2018). Language predic-
Neuroscience, 30(46), 15643–15653. https://doi.org/10. tion is reflected by coupling between frontal gamma and
1523/JNEUROSCI.1899-10.2010 posterior alpha oscillations. Journal of Cognitive
Stokes, M. G., Wolff, M. J., & Spaak, E. (2015). Decoding rich Neuroscience, 30(3), 432–447. https://doi.org/10.1162/
spatial information with high temporal resolution. Trends jocn_a_01190
in Cognitive Sciences, 19(11), 636–638. https://doi.org/10. Wang, L., Kuperberg, G., & Jensen, O. (2018). Specific lexico-
1016/j.tics.2015.08.016 semantic predictions are associated with unique spatial
 LANGUAGE, COGNITION AND NEUROSCIENCE 29

and temporal patterns of neural activity. Elife, 7, e39061. Woolnough, O., Donos, C., Rollo, P. S., Forseth, K. J., Lakretz, Y.,
https://doi.org/10.7554/eLife.39061 Crone, N. E., Fischer-Baum, S., Dehaene, S., & Tandon, N.
Wang, L., Schoot, L., Brothers, T., Alexander, E., Warnke, L., Kim, (2021). Spatiotemporal dynamics of orthographic and
M., Khan, S., Hämäläinen, M., & Kuperberg, G. R. (2023). lexical processing in the ventral visual pathway. Nature
Predictive coding across the left fronto-temporal hierarchy Human Behaviour, 5(3), 389–398. https://doi.org/10.1038/
during language comprehension. Cerebral Cortex, 4478– s41562-020-00982-w
4497. https://doi.org/10.1093/cercor/bhac356 Wydell, T. N., Vuorinen, T., Helenius, P., & Salmelin, R. (2003).
Wang, L., Wlotko, E., Alexander, E. J., Schoot, L., Kim, M., Neural correlates of letter-string length and lexicality
Warnke, L., & Kuperberg, G. R. (2020). Neural evidence for during reading in a regular orthography. Journal of
the prediction of Animacy features during language com- Cognitive Neuroscience, 15(7), 1052–1062. https://doi.org/
prehension: Evidence from MEG and EEG Representational 10.1162/089892903770007434
Similarity Analysis. The Journal of Neuroscience, 40(16), Xiang, M., & Kuperberg, G. (2015). Reversing expectations
3278–3291. https://doi.org/10.1523/JNEUROSCI.1733-19. during discourse comprehension. Language, Cognition and
2020 Neuroscience, 30(6), 648–672. https://doi.org/10.1080/
Warrington, E. K., & McCarthy, R. (1987). Categories of knowl- 23273798.2014.995679
edge. Brain, 110(5), 1273–1296. https://doi.org/10.1093/ Zannino, G. D., Perri, R., Pasqualetti, P., Caltagirone, C., &
brain/110.5.1273 Carlesimo, G. A. (2006). Analysis of the semantic represen-
Warrington, E. K., & Shallice, T. (1984). Category specific seman- tations of living and nonliving concepts: A normative
tic impairments. Brain, 107(3), 829–853. https://doi.org/10. study. Cognitive Neuropsychology, 23(4), 515–540. https://
1093/brain/107.3.829 doi.org/10.1080/02643290542000067