Lexical Preactivation in Basic Linguistic Phrases

Lexical Preactivation in Basic Linguistic Phrases Joseph Fruchter1*, Tal Linzen1*, Masha Westerlund1,2, and Alec Marantz1,2 Abstract ■ Many previous studies have shown that predictable words are read faster and lead to reduced neural activation, consistent with a model of reading in which words are activated in advance of being encountered. The nature of such preactivation, however, has typically been studied indirectly through its subsequent effect on word recognition. Here, we use magnetoencephalography to study the dynamics of prediction within serially presented adjective–noun phrases, beginning at the point at which the predictive information is first available to the reader. Using corpus transitional probability to estimate the predictability of a noun, INTRODUCTION Top–down predictive processing is one of the fundamental principles of brain function (Bar, 2007). Using prior knowledge and contextual information, higherorder areas communicate expectations to lower areas, which then compare the received input to the predicted input. Language processing is no exception to this rule. For example, listeners move their eyes to items that are predictable from context, before the item itself has been named (Altmann & Kamide, 1999; Kamide, Altmann, & Haywood, 2003). Likewise, predictable words are read more quickly (Ehrlich & Rayner, 1981) and elicit reduced neural signals, most commonly observed as a reduction in the N400 ERP component (Kutas & Hillyard, 1984). Predictability effects have been taken as support for the notion that likely upcoming words are at least partly preactivated in advance of being encountered (Kutas & Federmeier, 2000), although some have argued for an alternative explanation in which these effects stem from the increased ease of integrating predictable words into the preceding context (Brown & Hagoort, 1993; Norris, 1986; see Lau, Phillips, & Poeppel, 2008, for a review of the different interpretations of the N400 response). Recent empirical support for the preactivation account comes from an experiment showing a modulation in N400 effects for the English indefinite articles a and an, based on whether the context licensed a predic- 1 New York University, 2New York University Abu Dhabi *Joseph Fruchter and Tal Linzen contributed equally to this work. © 2015 Massachusetts Institute of Technology we found an increase in activity in the left middle temporal gyrus in response to the presentation of highly predictive adjectives (i.e., adjectives that license a strong noun prediction). Moreover, we found that adjective predictivity and expected noun frequency interacted, such that the response to the highly predictive adjectives (e.g., stainless) was modulated by the frequency of the expected noun (steel ). These results likely reflect preactivation of nouns in highly predictive contexts. The fact that the preactivation process was modulated by the frequency of the predicted item is argued to provide support for a frequency-sensitive lexicon. ■ tion for a noun that agreed with the article (e.g., The day was breezy so the boy went outside to fly a [kite]/ *an [airplane]; DeLong, Urbach, & Kutas, 2005). Because both indefinite articles should be equally easy to integrate into the semantic context of the sentence, the only plausible explanation for the N400 effect in this case is that participants were preactivating the representation of the upcoming noun. Similar responses have been reported for semantically vacuous agreement features in other European languages ( Van Berkum, Brown, Zwitserlood, Kooijman, & Hagoort, 2005; Wicha, Moreno, & Kutas, 2003). A separate line of evidence for lexical preactivation comes from predictability effects in early sensory responses (Kim & Lai, 2012; Dikker & Pylkkänen, 2011; Dikker, Rabagliati, & Pylkkänen, 2009). For example, Dikker and Pylkkänen (2011) presented participants with pictures, followed by words that either did or did not match the presented image. Some of the pictures were predictive of a specific word (e.g., a picture of an apple), and some were not, instead denoting a larger semantic category from which a single predictable word could not be isolated (e.g., a picture of a shopping bag full of groceries). When the strong prediction for the word apple generated by the presentation of a picture of an apple was violated, there was an increase in the M100, a magnetoencephalography (MEG) signal generated in visual cortex around 100 msec after stimulus presentation. Importantly, the contexts that did not afford a specific lexical prediction did not elicit a similar violation response. These findings point to a top–down modulation of visual cortex activity by lexical expectations generated in language regions (Dikker et al., 2009). In summary, Journal of Cognitive Neuroscience 27:10, pp. 1912–1935 doi:10.1162/jocn_a_00822 predictability effects for semantically vacuous words, on the one hand, and top–down modulatory effects in sensory cortex, on the other hand, both provide support for the lexical preactivation account of predictability effects. A popular model for these top–down predictability effects proposes that prediction arises directly from the organization of neurons in the cortex. Predictive coding theories propose that cortical regions contain two types of neuron populations: “expectation” neurons, which encode the representations, and “surprise” neurons, which encode the mismatch between the predicted representation and the bottom–up input (Friston, 2005). The predictive coding model makes the prediction that anticipatory processing of a stimulus should elicit neural activity in some of the same regions that are active when that same stimulus is perceived. This prediction has been increasingly supported by recent evidence (Kok, Failing, & de Lange, 2014; Egner, Monti, & Summerfield, 2010). For example, seeing an image of a face and anticipating seeing a face both activate the fusiform face area (Egner et al., 2010). One might therefore expect that anticipation of a lexical stimulus would be detectable in the same regions involved in lexical processing more generally. Indeed, a follow-up analysis of the picture–noun data set described above showed that MEG activity in temporal and occipital cortex was increased in the presence of a specific prediction (picture of an apple) before the presentation of the word (Dikker & Pylkkänen, 2013). Predictive coding models make the additional prediction that lexical preactivation in predictive contexts will reflect the identity of the individual item being predicted. This has been demonstrated for predictive processing in earlier sensory areas. Kok, Jehee, and de Lange (2012), for example, showed that the representation of a predicted element was “sharpened,” in that it became easier to decode its identity from activity in visual cortex. More recently, Kok et al. (2014) showed that patterns of activity in early visual cortex evoked by expected (but not presented) stimuli had similar feature specificity to those evoked by the stimulus itself. Design The goal of the current experiment is to find direct evidence for the preactivation of particular lexical items. One challenge in tackling this question lies in the fact that it is difficult to pinpoint the exact moment at which a linguistic prediction is generated. Prediction has typically been studied by varying the predictability of the last word of a sentence (Kutas & Hillyard, 1984, and many others). An issue with using this paradigm for studying preactivation is that a prediction for the last word of a sentence likely arises gradually as more and more information about the sentence is accumulated, rather than being generated in its entirety immediately before the last word of the sentence. For example, in the sentence he loosened the tie around his neck, much of the information that enables a reader to predict the word neck is likely to already be available after the word tie. This complication makes it difficult to temporally isolate a preactivation signal. We departed from this classic paradigm in two respects. First, following Bemis and Pylkkänen (2011), we used simple adjective–noun phrases, such as stainless steel, as opposed to full sentences. Participants read the phrases while their neural activity was recorded using MEG. To ensure that they were fully engaged with the materials, they made a lexical decision on the second word of the phrase (the noun). While reading isolated phrases is undoubtedly less natural than reading full sentences, this paradigm has several advantages for our purposes. First, the sources of information used to generate the predictions are more limited, giving us better control over the nature of the prediction signal. For example, syntactic structure is kept constant across all stimuli, minimizing variation in syntactic predictions, which can affect neural responses (Linzen, Marantz, & Pylkkänen, 2013). Second, because each of the items is much shorter, we can include significantly more items than in a sentential paradigm. Finally, and most crucially, this paradigm allowed us to achieve precise control over the point in time at which a prediction can be generated: in the phrase stainless steel, a specific lexical prediction can be generated immediately at stainless. Using MEG, we can then measure neural activity before the word steel, giving us a direct measure of a prediction signal, rather than an indirect error response. A second way in which our paradigm differs from classic N400 experiments is in the way predictability was operationalized. Most studies of predictability have estimated it using the cloze procedure (Taylor, 1953), a pretest in which native speakers read a sentence with a missing final word and are instructed to fill in the blank. The cloze probability of a word is defined as the proportion of participants who completed the sentence using that word. For example, if 97% of participants complete the sentence he loosened the tie around his… with the word neck, the cloze probability of the word neck would be .97. By contrast, here we operationalized predictability using corpus transitional probability (TP), that is, the probability of encountering a second word w2 given that a first word w1 has been encountered (P(w2|w1)), estimated from frequencies in the Corpus of Contemporary American English (Davies, 2009). For stainless steel, for example, TP is calculated as the number of times stainless steel appeared in the corpus divided by the total number of times stainless appeared in the corpus (McDonald & Shillcock, 2003). One advantage of TP over cloze probability is that, whereas cloze probability is always bounded by the number of respondents (with 100 participants, the lowest possible cloze probability is .01), TP does not have this limitation, which makes it possible Fruchter et al. 1913 to study differences between items with fairly low predictability, for example, TP = .05, and very low predictability, for example, TP = .005 (Smith & Levy, 2013). To recapitulate, our paradigm allowed us to characterize the nature of the signal generated by the expectation of a noun. We quantified the degree to which an adjective evokes an expectation for a noun using corpus TP from the adjective to the noun. Anatomical ROI As mentioned above, predictive coding models suggest that expectation of a stimulus involves neural activity in the same region that processes that stimulus when it is encountered. Preactivation of a specific lexical item is therefore likely to occur in the areas that are involved in lexical access more generally. Consequently, we focused our analysis on the left middle temporal gyrus (MTG), a cortical region thought to play a central role in lexical access (Friederici, 2012; Hickok & Poeppel, 2007; Rodd, Davis, & Johnsrude, 2005; Indefrey & Levelt, 2004; Binder et al., 1997). This area has also recently been implicated in generating expectations from linguistic stimuli and matching them against perceptual stimuli (Francken, Kok, Hagoort, & De Lange, 2015). Word Frequency The preactivation of a word is likely to involve some of the same processes that are engaged when the word is accessed in other circumstances. One of the most reliable predictors of ease of lexical access is word frequency. Frequent words are processed faster in lexical decision experiments ( Whaley, 1978; Rubenstein, Garfield, & Millikan, 1970) and during natural reading (Inhoff & Rayner, 1986). EEG experiments have found that the amplitude of the N400 is reduced for frequent words ( Van Petten & Kutas, 1990; Smith & Halgren, 1987). Importantly, within the MEG literature, frequency effects have been found in the left MTG during the time window of the M350, the evoked response thought to be associated with lexical access (Solomyak & Marantz, 2010; Embick, Hackl, Schaeffer, Kelepir, & Marantz, 2001). This body of evidence leads us to predict that the preactivation of an infrequent word should be more effortful than the preactivation of a frequent one. Concretely, when an adjective is predictive of a specific noun, we expect the frequency of the expected noun to modulate MTG activity before the presentation of the noun. For example, upon recognition of the adjective stainless, we expect participants to preactivate the linguistic representation of the likely continuation steel; we therefore expect to see concomitant effects of the frequency of steel associated with this preactivation in the MTG. More generally, we expect to see an interaction in the MTG between adjective predictivity and the frequency of the expected noun continuation, such that as adjective predictivity in1914 Journal of Cognitive Neuroscience creases, we are more likely to observe effects of the frequency of the most likely noun continuation. This should occur in the time window subsequent to recognition of the adjective, but before the presentation of the noun. After the presentation of the noun, we expect that frequency effects will be reduced for the more predictable items, consistent with EEG experiments that have shown that N400 frequency effects are only significant for words that appear earlier in a sentence ( Van Petten & Kutas, 1990) or that are less predictable from context (Dambacher, Kliegl, Hofmann, & Jacobs, 2006). METHODS Materials We first define the variables that we calculated for each phrase and then describe how the phrases were selected (partially based on those variables). Lexical Variables We illustrate the calculation of the lexical variables using the phrase economic reform. The most likely continuation of economic is not reform, but growth. In this case, we say that the expected noun is growth, and the presented noun is reform. We define the following variables: • Adjective frequency: freq(economic) • Adjective predictivity: TP from the adjective to its most likely noun continuation: P( growth|economic) • Expected noun frequency: the frequency of the adjective’s most likely noun continuation: freq( growth) • Presented noun frequency: the frequency of the noun that was actually presented: freq(reform) • Presented noun predictability: TP from the adjective to the presented noun: P(reform|economic) Focusing on a single expected noun is clearly a simplification; most adjectives license more than one prediction. After reading the adjective economic, for example, participants may well predict both growth and reform. These continuations would likely be preactivated in proportion to their conditional probability (Smith & Levy, 2013; DeLong et al., 2005): Following recognition of economic, the noun growth may be activated to a greater extent than reform.1 To capture this intuition, we defined a generalization of expected noun frequency that we term weighted expected noun frequency. This variable is a weighted average of the frequencies of the adjective’s continuations, where the weights are given by the TPs of the continuations. We only considered noun continuations within phrases that met our minimum frequency requirement (i.e., 50 tokens in the corpus, corresponding to a probability of 1 in 8 million). Consequently, some of the conditional probability mass for each adjective was not assigned to any noun; we assigned this probability to a generic noun that had the average frequency of Volume 27, Number 10 all nouns in the corpus. As an illustration, in the case of economic (assuming that there are only two suprathreshold noun continuations), the calculation of this variable would be given by WFðeconomicÞ ¼ PðreformjeconomicÞ  freqðreformÞ þ Pð growthjeconomicÞ  freqð growthÞ þ ð1 − PðreformjeconomicÞ − Pð growthjeconomicÞÞ  avgNounFreq The shape of frequency effects has long been known to be approximately logarithmic ( Whaley, 1978), and there is increasing evidence that this is the case for predictability effects as well (Smith & Levy, 2013). We therefore log-transformed all frequency and predictability variables before entering them into our statistical models. Selection Criteria A set of 474 adjective–noun phrases was obtained as follows. We first selected all sequences of two words from the Corpus of Contemporary American English (Davies, 2009) that satisfied the following conditions: (1) The first word was tagged as an adjective at least 90% of the time, according to the automatic partof-speech tagging included with the corpus. (2) The second word was tagged as a noun at least 90% of the time. (3) The sequence had a frequency of at least 50 tokens in the corpus out of ∼400 million tokens in the corpus, corresponding to a probability of approximately 1 in 8 million. (4) The length of both words was between three and nine characters. (5) All nouns had an accuracy of at least 75% in the lexical decision data in the English Lexicon Project (this criterion was implemented to ensure that participants were likely to be familiar with the words). Many phrases contained the same adjectives or nouns as other phrases in the selection (e.g., high table and high chair, or black chair and high chair). Whenever this was the case, we only kept the phrase in which the noun was most predictable. Because phrases with highly predictable nouns are relatively rare, this procedure maximized our coverage of the predictability range. More specifically, we first grouped the phrases by noun (e.g., black chair and high chair) and excluded all but the most predictable items; we then grouped the remaining phrases by adjective (e.g., high table and high chair) and again excluded all but the most predictable items. This process yielded a candidate set of phrases, each composed of a unique adjective and unique noun. A side effect of this procedure was that many phrases, particularly towards the lower end of the TP range, contained nouns that were not the most predictable ones given the adjective. For example, the phrase economic reform, which has TP = .01, was included even though a phrase with the same adjective, economic growth, had higher TP (.05). This was done because growth occurs in rapid growth, which has even higher TP (.08). In our final set, the noun was the most expected continuation of the adjective in 51% of the phrases (242 of 474); in the top quartile of adjective predictivity (i.e., adjectives that had a noun continuation with TP > .10), this proportion was 77% (90 of 117). Because the phrase used when the most predicted noun was not available typically had the second highest TP among all phrases that included the adjective, the order of magnitude of the TP of the selected phrase was usually similar to that of the phrase that was excluded (median ratio of highest TP to selected TP: 2.63). Given the candidate set of phrases, we excluded items that were clearly part of a longer phrase (e.g., congestive heart, which always appears in the context congestive heart failure) and items that are usually capitalized, which tend to be names of places or works of art (e.g., Purple Haze). Finally, we asked six undergraduate students to rate the phrases for familiarity and excluded phrases that five of six raters rated as unfamiliar (e.g., logistic regression). The final list contained 474 phrases, which are listed in the Appendix, along with their associated TP values. After sorting by TP, the noun in every other phrase was replaced with a pronounceable nonword (e.g., academic dusporate), obtained using Wuggy (Keuleers & Brysbaert, 2010). Table 1 lists the descriptive statistics for key stimulus variables. Noun frequency and noun predictability (TP between adjective and noun) were correlated (r = .34), Table 1. Descriptive Statistics for Stimulus Variables Adjective Frequency (log) Noun Frequency (log) Phrase Frequency (log) Adjective Predictivity (log) Noun Predictability (log) Min 5.05 5.91 3.93 −5.36 −7.9 Max 12.88 12.68 10.77 −0.17 −0.17 Median 8.59 9.77 4.98 −3.02 −3.49 Mean 8.69 9.73 5.23 −2.91 −3.45 SD 1.43 1.27 1.05 1.00 1.37 Fruchter et al. 1915 as were adjective frequency and noun predictability (r = −.71). Adjective frequency was also correlated with adjective predictivity (r = −.59). None of the variables were strongly correlated with adjective length or noun length (all r < .3). The high correlation between adjective frequency and the two predictability measures is due to the fact that these quantities are mathematically related: logðTPÞ ¼ logðfreqðphraseÞ=freqðadjÞÞ ¼ logðfreqðphraseÞÞ− logðfreqðadjÞÞ As mentioned above, we only selected phrases that appeared with a frequency of at least 1 per 8 million to eliminate implausible or ungrammatical phrases. This entails that log(TP) and log(freq(adj)) must sum to at least 3.9, and therefore, a phrase cannot simultaneously have log(TP) = −6 and log(freq(adj )) = 7. Note that because log(freq( phrase)) is always positive, many combinations of values for log(TP) and log(freq(adj )) would still be impossible even if the frequency threshold for phrases were lifted. Participants Sixteen participants (nine women) from New York City participated in the experiment. All participants provided informed consent and were paid for their participation. Participants ranged in age from 19 to 45 years (median = 25.5 years). All participants were right-handed (assessed using the Edinburgh Handedness Inventory; Oldfield, 1971) and were native speakers of English with normal or corrected-to-normal vision. Procedure The experiment was conducted in the KIT/NYU facility at New York University. Before recording, the head shape of each participant was digitized to allow source localization and coregistration with structural MRIs. We also digitized three fiducial points (the nasion and the left and right preauricular points) and the position of five coils, placed around the participant’s face. Once the participant was situated in the magnetically shielded room for the experiment, the position of these coils was localized with respect to the MEG sensors, allowing us to assess the position of the participant’s head for source reconstruction. Data were recorded continuously with a 157-channel axial gradiometer (Kanazawa Institute of Technology, Kanazawa, Japan). Structural MRIs were obtained for 15 of the 16 participants; the MEG data from one participant were thus eliminated from analysis because of failure to obtain a structural MRI. Before the experiment, participants were not given any indication of the goal of the experiment or the properties of the materials. The exact instructions were as follows: “You will read two letter strings on the screen, one at a 1916 Journal of Cognitive Neuroscience time. If the second string is a real English word, respond with your index finger. If it is not, respond with your middle finger.” Each participant saw all 474 items. The order of presentation was randomized for each participant individually. The assignment of items to conditions was fixed across participants; in other words, the same nouns were replaced with nonwords for all participants (see Materials for details). A given adjective was always presented with the same noun or nonword; for example, stainless was followed by steel, and uncharted was followed by the nonword cothenent (which replaced the predicted continuation territory) for all participants. Stimuli were presented using the Psychophysics Toolbox (Brainard, 1997; Pelli, 1997) and projected onto a screen approximately 50 cm away from the participant. They were presented in white 30-point Courier font on a gray background. The structure of each trial was as follows. First, a fixation cross was presented in the center of the screen for 300 msec, followed by a blank screen presented for 300 msec. The adjective was then presented for 300 msec, followed again by a blank screen presented for 300 msec. Finally, the noun (or nonword) was presented for 300 msec, and participants responded to the latter stimulus by pressing a button. Preprocessing The preprocessing and analysis of the MEG data closely followed the procedures of Solomyak and Marantz (2009, 2010). Environmental noise was removed from the data by regressing signals recorded from three orthogonally oriented magnetometers, placed approximately 20 cm away from the recording array, against the recorded data, using the continuously adjusted least squares method (Adachi, Shimogawara, Higuchi, Haruta, & Ochiai, 2001). The data were then low-pass filtered at 40 Hz, resampled to 250 Hz to facilitate analysis, and high-pass filtered at 0.1 Hz. MEG channels in which there was no signal or excessive amounts of noise were interpolated from neighboring channels or rejected (at most three per participant). Trials in which at least one channel showed a peak-to-peak amplitude exceeding 4000 fT were rejected, as these amplitude values are likely to reflect blinks and noise artifacts (the number of rejected trials ranged from 39 to 112, mean = 77.1, median = 77; the minimum number of trials analyzed for a given participant was 362). None of the participants were excluded because of excessive trial rejections. The MNE software package (Martinos Center MGH, Boston, MA) was used to estimate neuroelectric current strength based on the recorded magnetic field strengths using minimum l2 norm estimation (Dale & Sereno, 1993; Hämäläinen, Hari, Ilmoniemi, Knuutila, & Lounasmaa, 1993). Current sources were modeled as three orthogonal dipoles spaced approximately 5 mm apart across the cortical surface (Dale et al., 2000), yielding approximately 2500 potential electrical sources per hemisphere. The Volume 27, Number 10 participants’ cortical surfaces were reconstructed based on their structural MRIs using Freesurfer (Martinos Center). The neuromagnetic data were coregistered with the structural MRIs using MNE by first aligning the fiducial points and then using an Iterative Closest Point algorithm to minimize the difference between the scalp and the points defining the head shape of each participant. The forward solution was calculated for each source using a single-layer boundary element model based on the inner skull boundary. Noise covariance estimates were obtained from a 200-msec baseline period before the presentation of each adjective. Using the grand average of all trials across conditions (i.e., both Word and Nonword trials), the inverse solution was computed to determine the most likely distribution of neural activity. We utilized a free orientation analysis, in which the source orientations were unconstrained with respect to the cortical surface. The resulting source estimates were signed with a positive sign indicating an upward directionality and a negative sign indicating a downward directionality in the coordinate space defined by the head. The estimated activation was normalized into a test statistic by dividing the estimates by their predicted standard error given the noise covariance, yielding signed dynamic statistical parametric maps (dSPMs; Dale et al., 2000). The signal-to-noise ratio parameter, which controls the regularization of the estimates, was set to 1. Regions and Time Windows of Interest Main Analysis ROIs were defined anatomically using the cortical parcellation performed by FreeSurfer on the basis of the Desikan–Killiany gyral atlas (Desikan et al., 2006). We selected the left MTG anatomical ROI (Figure 3A later in the paper) for the purposes of our main analysis. In this and other temporal lobe labels, the Desikan–Killiany atlas includes the gyrus along with the banks of the surrounding sulci. We defined three time windows of interest: adjective lexical access, presented noun lexical access, and preactivation of the expected noun (Figure 3B). The time windows were defined on the basis of the peaks of the M350 evoked response. This component has been argued to be associated with lexical access (Pylkkänen & Marantz, 2003) and has demonstrated sensitivity to lexical variables such as frequency (Embick et al., 2001). Lexical access of the adjective was assessed in a time window starting 100 msec before the peak of the M350 response evoked by the presentation of the adjective and ending 100 msec after the peak (concretely, 242–442 msec after adjective presentation). Likewise, lexical access of the presented noun was assessed in a 200-msec time window centered around the peak of the M350 response to the noun, which was slightly earlier than the response to the adjective (197–397 msec post-noun onset). Finally, we made the simplifying assumption that ef- fects of lexical preactivation would be most evident after lexical access of the adjective was complete. Consequently, the preactivation time window started at the end of the adjective lexical access time window. To avoid including activity evoked by the presented noun, this time window ended at the presentation of the noun (concretely, the time window extended from 442 to 600 msec post-adjective onset) and was therefore somewhat shorter than the two other time windows. All three time windows are illustrated in Figure 3B. Figure 3B also shows the average left MTG response to the adjective for a four-way split of adjective predictivity as well as the response to the noun for a median split of presented noun predictability. Time-varying correlations (Figure 3C–E) were generated using sliding 50-msec windows centered at [25, 75, …, 575] msec post-adjective onset or post-noun onset. Supplemental Analyses We conducted post hoc analyses for two additional anatomical ROIs, which roughly corresponded to significant peaks of activity in the evoked response to the adjective at 300–400 msec (Figure 1A): the left lateral OFC (LOFC; Figure 4A later in the paper) and left superior temporal gyrus (STG) anatomical ROIs. Finally, because previous investigations of predictability showed early effects at the M100 ( Dikker & Pylkkänen, 2011; Dikker et al., 2009), we conducted an analysis of early predictability effects in this study, in which we examined the time window 100–200 msec post-noun onset in the MTG; we also examined 10-msec windows centered around the two candidate M100 peaks in the left cuneus anatomical ROI, corresponding to the location of visual cortex within the occipital lobe. Exploratory Analysis of Left Hemisphere Language Regions In addition to the main confirmatory ROI analysis, we conducted an exploratory analysis of a broad language network, covering most of the sources located within lateral cortical regions in the left frontal and left temporal lobes. Specifically, we pooled the sources within the following anatomical regions in the left hemisphere, as specified in the Desikan et al. (2006) parcellation: STG, transverse temporal gyrus, banks of the STS, MTG, inferior temporal gyrus, temporal pole, fusiform gyrus, insula,2 inferior frontal gyrus (pars triangularis, pars opercularis, and pars orbitalis), and LOFC. The analysis was conducted for the average activity in a given source over sliding 100-msec windows centered at [50, 150, …, 550] msec post-adjective onset or post-noun onset. The resulting t maps for the variables of interest are shown in Figures 1B–E and 2B–E; see the next section for details on how those t values were obtained. The figures for the evoked responses to the adjective (Figure 1A) and the noun (Figure 2A) were generated by first morphing Fruchter et al. 1917 Figure 1. Neural response to the adjective in left hemisphere language areas. (A) The grandaveraged evoked response to the adjective (in dSPM units). (B) The effect of adjective frequency. (C) The effect of adjective predictivity. (D) The effect of expected noun frequency for the items in the top half of adjective predictivity (controlling for adjective frequency and adjective predictivity). (E) The interaction between adjective predictivity and expected noun frequency (controlling for adjective frequency). In B–E, the t values represent the results of a second-level t test of the within-subject β-coefficients (described more fully in the Methods section). For all images, red and yellow indicate positively signed values, and blue indicates negatively signed values. Figure 2. Neural response to the noun in left hemisphere language areas. (A) The grandaveraged evoked response to the noun (in dSPM units). (B) The effect of presented noun frequency. (C) The effect of presented noun predictability. (D) The effect of presented noun frequency for the items in the bottom half of presented noun predictability (controlling for presented noun predictability). (E) The interaction between presented noun predictability and presented noun frequency. In B–E, the t values represent the results of a second-level t test of the within-subject β-coefficients (described more fully in the Methods section). For all images, red and yellow indicate positively signed values, and blue indicates negatively signed values. 1918 Journal of Cognitive Neuroscience Volume 27, Number 10 the grand-averaged activity (in dSPM units) for each participant into the neuroanatomical space of the average brain, followed by averaging across all participants; unlike the t maps, the evoked responses were calculated for all cortical sources within the left hemisphere. Statistical Methodology Behavioral Analysis After excluding the Nonword trials, we performed a logarithmic transformation of the RTs for the Word trials, following standard practice. For each participant, we excluded trials for which the log-transformed RTs were more than 2.5 standard deviations away from the participant’s mean and trials in which the RT was less than 100 msec or more than 5000 msec. We used the lme4 package in R (Bates, Maechler, & Bolker, 2013) to fit a linear mixed-effects model with crossed random effects for participants and items. Traditional repeated-measures designs account for “random” differences across participants that are irrelevant to the experimental manipulation and therefore enable generalization of results beyond the specific group of participants used in the experiment. Just like participants, linguistic materials may also differ from one another in many ways that are irrelevant to the experimental manipulation. Mixed-effects models with crossed random effects extend the logic of repeated measures to participants and items simultaneously and enable generalization beyond both the sample of participants and the sample of items used in the experiment (Baayen, Davidson, & Bates, 2008). This model was used to predict log-transformed RTs from presented noun predictability and presented noun frequency. We used a maximal random effects structure: for items, only a random intercept, and for participants, random slopes for presented noun predictability and presented noun frequency and their interaction as well as a random intercept. Predictors were centered before being entered into the model. The reported p values are derived from likelihood ratio tests in stepwise regression. MEG ROI Analyses Linear mixed-effects models were fitted to the average activity in an ROI over a time window of interest, following rejection of trials with activity at least 4 standard deviations away from the mean across all trials and all participants. The linguistic variables (e.g., frequency) were entered into the models as fixed effects. A maximal random effects structure was used, with random intercepts for participants and items, as well as a random slope for the particular linguistic variable being tested in the model. Predictors were centered before being entered into the models. To obtain the p values for the main effects as well as the significance of the stepwise regressions and inter- action effects, likelihood ratio tests were employed for the relevant nested linear mixed-effects models. In many cases, multiple regression was used to address correlations between stimulus variables; for example, the interaction between adjective predictivity and expected noun frequency was tested in a regression model that included adjective frequency as well (i.e., adjective frequency was controlled for). We analyzed the adjective and the noun time windows separately. All of the trials were included in the analyses of the adjective time window, and only Word trials were included in the analyses of the noun time window (i.e., Nonword trials were excluded). Exploratory MEG Analysis Because it was not computationally feasible to conduct the full mixed-effects analysis for each source individually, we employed a summary statistic approach (Holmes & Friston, 1998), as follows. For a given participant’s data, at each source and during each time window, we computed the β-coefficient from a linear regression model predicting the source activity (in dSPM units) as a function of the linguistic variable of interest (e.g., frequency). We then morphed the resulting maps of the β-coefficients from each participant’s neuroanatomical space to the space of the average brain using seven iterative smooth steps. Because the a priori selection of the MTG anatomical ROI allowed us to establish the significance of the effects in the main analysis, we did not correct the resulting t maps for multiple comparisons across sources. Consequently, the t maps should not be used to determine the significance of the effects, but rather to verify their general spatial distribution. RESULTS Behavioral Results Accuracy ranged from 95% to 99% (median = 97.9%). Mean RTs ranged from 491 to 1057 msec (median = 735 msec). Both presented noun predictability and presented noun frequency were inversely correlated with RTs (predictability: β = −.012; frequency: β = −.02). Both variables were significant in stepwise regressions (predictability: p = .007; frequency: p = .002). The interaction between the variables did not reach significance (β = −.0004, p = .88). Adding trial number to the regression model revealed that participants became significantly slower over the course of the experiment (β = .0002, p = .002). This effect did not interact with either predictability or frequency (predictability: p = .6; frequency: p = .8). There is therefore no evidence that participants modified their prediction strategy over the course of the experiment. Finally, a logistic mixed-effects model showed that accuracy was marginally higher on Fruchter et al. 1919 high frequency nouns (β = .3, p = .09) and did not vary based on predictability. Discrepancy Trials To see whether the discrepancy between the predicted noun and the presented noun affected RTs, we conducted a separate analysis restricted to those Word trials in which the presented noun was not the most predictable continuation for the given adjective (“discrepancy trials”). The number of trials included in this analysis was roughly a quarter of the total trials in the experiment, because we eliminated all Nonword trials as well as Word trials in which the presented noun was the most predictable continuation for the adjective; thus, the statistical power for this particular analysis is lower than other analyses in this study. As in the complete set of Word trials, higher predictability and presented noun frequency were associated with shorter RTs (frequency: β = −.019, p = .004; predictability: β = −.02, p = .01), and there was no interaction between the two ( p = .93). There was a marginal positive effect of adjective predictivity on RTs (β = .016, p = .05), indicating that recognition of a noun is slowed down by the presence of a conflicting prediction for a different noun. MEG Results: Evoked Response The grand-averaged evoked responses to the adjective and the noun are shown from lateral perspectives in Figures 1A and 2A, respectively. There is widespread negative activity (shown in blue) within the left temporal lobe, particularly at 300–400 msec post-adjective onset, as well as a significant patch of positive activity (shown in red and yellow) within the left inferior frontal cortex, mostly overlapping with the LOFC ROI, at 300–400 msec post-adjective onset. Figure 3B displays the time course of average activity within the left MTG, for the adjective and noun time windows. The second negative peak in the adjective time window (i.e., the M350 response to the adjective) occurs at a latency of roughly 350 msec post-adjective onset, whereas the second negative peak in the noun time window (i.e., the M350 response to the noun) occurs at a latency of roughly 300 msec post-noun onset. MTG ROI Analysis same region and time window (t = −3.74, p = .0007).4 Because of the high correlation between adjective frequency and adjective predictivity (r = −.59), we assessed the significance of each variable in a stepwise regression with the other variable. Using this procedure, only adjective predictivity remained significant (χ 2 = 5.96, p = .01). In the presented noun lexical access time window, after exclusion of the Nonword trials, higher presented noun frequency was associated with weaker left MTG activity (t = 2.04, p = .05). Higher presented noun predictability led to significantly weaker activity in the left MTG in the same time window (t = 4.66, p = .0002). In a stepwise regression, only presented noun predictability remained significant (χ 2 = 12.30, p = .0005). In summary, adjectives that license a relatively strong prediction evoked increased left MTG activity in the adjective lexical access time window; less predictable nouns evoked increased activity in the presented noun lexical access time window. Interaction Effects In the preactivation time window, adjective predictivity and expected noun frequency interacted in the left MTG (χ 2 = 5.73, p = .02, controlling for adjective frequency; see Figure 3D). A median split on adjective predictivity (Figure 3E) indicated that this interaction was driven by the fact that higher expected noun frequency led to significantly weaker activity, but only for the items in the top half of adjective predictivity (t = 2.43, p = .02, controlling for both adjective frequency and adjective predictivity). Adding weighted expected noun frequency to the model increased the fit somewhat, although the difference did not reach significance (χ 2 = 2.29, p = .13). The items in the bottom half of adjective predictivity showed no effect of expected noun frequency (t = −1.08, p = .28, controlling for adjective frequency and adjective predictivity). In the presented noun lexical access time window, there was a significant interaction between presented noun predictability and presented noun frequency in the left MTG (χ 2 = 6.51, p = .01; see Figure 3D). A median split on presented noun predictability (Figure 3E) indicated that this interaction was driven by significantly weaker activity in response to higher-frequency presented nouns, but only for the items in the bottom half of presented noun predictability (t = 3.05, p = .006, controlling for presented noun predictability). The items in the top half of presented noun predictability showed no effect of presented noun frequency (t = −0.94, p = .35, controlling for presented noun predictability). Main Effects In the adjective lexical access time window, higher adjective frequency was associated with weaker activity in the left MTG (t = 2.94, p = .004; Figure 3C).3 Higher adjective predictivity was associated with stronger activity in the 1920 Journal of Cognitive Neuroscience Discrepancy Trials We again conducted a separate analysis of Word trials in which the presented noun was not the most expected Volume 27, Number 10 Figure 3. Left MTG ROI analysis. (A) ROI: The MTG ROI, displayed in green on the average brain, from lateral and ventral perspectives. (B) Average activity: On the left, the average response in the MTG to the adjective for high (top 10%: blue line), mid-high (top 10–50%: solid black line), mid-low (bottom 10–50%: dotted black line), and low (bottom 10%: red line) adjective predictivity. On the right, the average response in the MTG to the noun for the top half (solid black line) and bottom half (dotted black line) of presented noun predictability. Horizontal lines with arrows indicate the time windows of interest for the ROI analyses. (C) Main effects: On the left, the effects of adjective predictivity (blue) and adjective frequency (red) during the adjective time window. On the right, the effects of presented noun predictability (blue) and presented noun frequency (red) during the noun time window. (D) Interaction effects: On the left, the interaction between adjective predictivity and expected noun frequency (controlling for adjective frequency) during the adjective time window. On the right, the interaction between presented noun predictability and presented noun frequency during the noun time window. (E) Binned analyses: On the left, the effect of expected noun frequency (controlling for adjective frequency and adjective predictivity) during the adjective time window for the top half (solid red line) and bottom half (dotted red line) of adjective predictivity. On the right, the effect of presented noun frequency (controlling for presented noun predictability) during the noun time window for the top half (solid red line) and bottom half (dotted red line) of presented noun predictability. The dotted black lines in C–E represent the level of correlation needed to reach statistical significance at p = .05 (uncorrected). Fruchter et al. 1921 one. In the presented noun lexical access time window, there was no effect of adjective predictivity (t = −0.95, p = .35, controlling for presented noun predictability). We repeated the analysis in the later time window 300– 500 msec, which we selected post hoc to more accurately capture the peak of the presented noun predictability effect. In this time window, higher adjective predictivity was associated with greater MTG activity (t = −2.00, p = .05, controlling for presented noun predictability); this effect did not reach significance; however, when presented noun frequency was included in the model as well (t = −1.60, p = .11). There is therefore some neural evidence for an opposing effect of a violated strong prediction, relative to the effect of presented noun predictability. Removal of High Valence Items Because our phrase selection process was automatic, our final set of materials included some phrases with high valence (e.g., rectal exam). To rule out the possibility that some of our effects were due to the presence of these high valence phrases, we manually eliminated 18 phrases (denoted with asterisks in the Appendix) that we judged to contain a high valence adjective or noun and subsequently repeated our primary MTG analyses without these items. In the adjective lexical access time window, adjective frequency (t = 2.71, p = .008) and adjective predictivity (t = −3.66, p = .001) remained significant. In the preactivation time window, there was a significant interaction of adjective predictivity and expected noun frequency (χ 2 = 5.97, p = .01, controlling for adjective frequency); this interaction was driven by a significant effect of expected noun frequency for the items in the top half of adjective predictivity (t = 2.78, p = .01, controlling for adjective frequency and adjective predictivity). In the presented noun lexical access time window, presented noun frequency was below significance (t = 1.70, p = .10), although presented noun predictability remained highly significant (t = 4.15, p = .0006). The interaction of presented noun frequency and presented noun predictability remained significant (χ 2 = 4.37, p = .04), driven by a significant effect of presented noun frequency for the items in the bottom half of presented noun predictability (t = 2.81, p = .01, controlling for presented noun predictability). In summary, all of the effects in the MTG survived the removal of the high valence items, with the exception of the presented noun frequency effect, which dipped below the significance threshold. Post hoc ROI Analyses STG Given the significant patch of negative activity in the left STG at 300–400 msec (Figures 1A and 2A) as well as the role of this region in the language network (e.g., Friederici, 1922 Journal of Cognitive Neuroscience 2012), we conducted a post hoc analysis of activity in this region. In the adjective lexical access time window, there were significant effects of adjective frequency (t = 3.59, p = .002) and adjective predictivity (t = −2.22, p = .04). Both effects were in the same direction as those found in the MTG; the frequency effect was slightly stronger, and the predictivity effect was slightly weaker than in the MTG. In the preactivation time window, there was no significant interaction between adjective predictivity and expected noun frequency (χ 2 = 1.93, p = .16, controlling for adjective frequency). Furthermore, there was no significant effect of expected noun frequency for the items in the top half of adjective predictivity (t = 1.08, p = .29, controlling for adjective frequency and adjective predictivity). In the presented noun lexical access time window, there was no effect of presented noun frequency (t = 0.28, p = .78), and the effect of presented noun predictability was marginally significant (t = 1.88, p = .07). There was no interaction of presented noun frequency and presented noun predictability (χ 2 = 0.19, p = .67) and no effect of presented noun frequency for the items in the bottom half of presented noun predictability (t = 0.80, p = .43, controlling for presented noun predictability). In summary, most of the effects in this region are either similar to or weaker than the effects in the left MTG, supporting the selection of the latter region as our central ROI. LOFC Within the frontal lobe, the region that is traditionally associated with language processing is the left inferior frontal gyrus (IFG), which includes Broca’s area (e.g., Friederici, 2012). However, the patch of positive evoked activity in the frontal lobe at 300–400 msec post-adjective onset (Figure 1A) did not localize to the IFG and instead overlaps almost entirely with the left LOFC (see Figure 4A). We therefore report a post hoc analysis of the activity in that anatomical region rather than the IFG. The evoked response in the LOFC (Figure 4B) shows a prominent positive peak at roughly 350–400 msec following word presentation; correspondingly, the direction of the presented noun predictability effect is such that activity is more negative for the high predictability condition. Thus, the sign of the correlation with presented noun predictability is negative, as opposed to the effects in the MTG and STG, which were positive; the latter point follows from the fact that a negative correlation with a positive peak indicates a weakening of activity, whereas a negative correlation with a negative peak indicates a strengthening of activity. In the adjective lexical access time window, higher adjective frequency was associated with weaker LOFC activity (t = −2.56, p = .01; Figure 4C), but there was no Volume 27, Number 10 Figure 4. Left LOFC post hoc ROI analysis. (A) ROI: The LOFC ROI, displayed in green on the average brain, from lateral and ventral perspectives. (B) Average activity: On the left, the average response in the LOFC to the adjective for the top half (solid black line) and bottom half (dotted black line) of adjective predictivity. On the right, the average response in the LOFC to the noun for the top half (solid black line) and bottom half (dotted black line) of presented noun predictability. (C) Main effects: On the left, the effects of adjective predictivity (blue) and adjective frequency (red) during the adjective time window. On the right, the effects of presented noun predictability (blue) and presented noun frequency (red) during the noun time window. (D) Interaction effects: On the left, the interaction between adjective predictivity and expected noun frequency (controlling for adjective frequency) during the adjective time window. On the right, the interaction between presented noun predictability and presented noun frequency during the noun time window. (E) Binned analyses: On the left, the effect of expected noun frequency (controlling for adjective frequency and adjective predictivity) during the adjective time window for the top half (solid red line) and bottom half (dotted red line) of adjective predictivity. On the right, the effect of presented noun frequency (controlling for presented noun predictability) during the noun time window for the top half (solid red line) and bottom half (dotted red line) of presented noun predictability. The dotted black lines in C–E represent the level of correlation needed to reach statistical significance at p = .05 (uncorrected). Fruchter et al. 1923 main effect of adjective predictivity (t = 0.72, p = .48). In the preactivation time window, there was a significant interaction of adjective predictivity and expected noun frequency (χ 2 = 7.19, p = .007, controlling for adjective frequency; Figure 4D); this interaction was driven by the fact that higher expected noun frequency led to weaker activity, for the items in the top half of adjective predictivity (t = −3.21, p = .003, controlling for adjective frequency and adjective predictivity; Figure 4E). Adding weighted expected noun frequency to the model marginally improved the fit (χ 2 = 3.37, p = .07). In the presented noun lexical access time window, higher presented noun frequency was associated with marginally weaker LOFC activity (t = −1.96, p = .06; Figure 4C), and higher presented noun predictability was associated with significantly weaker activity (t = −2.98, p = .008). There was no interaction of presented noun frequency and presented noun predictability (χ 2 = 0.004, p = .95; Figure 4D), and there was no significant effect of presented noun frequency for the items in the bottom half of presented noun predictability (t = −1.31, p = .19, controlling for presented noun predictability; Figure 4E). In summary, most of the variables had similar effects on LOFC activity as they did on MTG activity (though with opposite signs, as discussed earlier), with the exception of the main effect of predictivity in the adjective lexical access time window and the interaction between presented noun frequency and predictability in the presented noun lexical access time window, which were found in the MTG but not the LOFC. Early Predictability Effects To determine whether there was an early effect of presented noun predictability in the MTG, we analyzed the time window 100–200 msec post-noun onset. This time window indeed showed a significant effect of presented noun predictability (t = 3.29, p = .002). Given the possibility of spillover from the earlier effect of adjective frequency in this region, we also ran a stepwise regression with adjective frequency and presented noun frequency; in this model, presented noun predictability was no longer significant (χ 2 = 1.22, p = .27). However, it is difficult to interpret the latter fact in light of the high correlation of adjective frequency and presented noun predictability (r = −.71), which would serve to reduce the effects of each variable when present in the same model. In summary, there is somewhat inconclusive evidence for early predictability effects in the MTG after noun presentation. The evoked activity in the left cuneus, roughly overlapping with the location of visual cortex in the occipital lobe, showed a negative peak at 76 msec, followed by a positive peak at 136 msec post-noun onset. Following the M100 analysis in Dikker and Pylkkänen (2011), we analyzed 10-msec windows centered around both peaks.5 1924 Journal of Cognitive Neuroscience In addition to noun predictability, we tested for an effect of noun length to validate our selection of visual ROI (under the assumption that early visual processing should be sensitive to visual form properties, such as word length). For the time window around the earlier peak, there was no effect of presented noun predictability (t = 0.53, p = .60, for the time window 71–81 msec post-noun onset), but higher presented noun length was associated with stronger activity (t = −2.52, p = .02). For the time window around the later peak, higher presented noun predictability was associated with marginally weaker activity (t = −1.87, p = .06, for the time window 131–141 msec post-noun onset), and higher presented noun length was associated with stronger activity (t = 2.74, p = .01). In a stepwise regression with presented noun length, the effect of presented noun predictability dipped further below significance (χ 2 = 1.71, p = .19). The evidence in our data for visual predictability effects is therefore inconclusive, although suggestive. Exploratory Analysis of Language Areas Figures 1 and 2 display the results of the exploratory analysis of language areas for the response to the presentation of the adjective and noun, respectively. Many of the patterns observed within the ROI results are visible in the present analysis: (i) there are positive effects of adjective frequency (Figure 1B) and presented noun frequency (Figure 2B) within the mid-anterior temporal lobe and corresponding negative effects in the LOFC, both peaking at 400–500 msec after the presentation of the each word; (ii) the effects of adjective predictivity (Figure 1C) and presented noun predictability (Figure 2C) have opposite directionalities, and in particular, there is a negative effect of adjective predictivity and a positive effect of presented noun predictability, peaking at 400–500 msec in the temporal lobe; (iii) the preactivation effect—the effect of expected noun frequency in response to the highly predictive adjectives at 500–600 msec post-adjective onset (Figure 1D)—displays a strikingly similar spatial distribution to both the earlier effect of adjective frequency at 400–500 msec post-adjective onset as well as the later effect of presented noun frequency at 400–500 msec post-noun onset; and finally, (iv) the preactivation effect peaks at 500–600 msec, although the effects of adjective frequency and adjective predictivity are no longer visible at that latency, suggesting that these effects are distinct from each other. DISCUSSION This study set out to characterize the neural signal corresponding to lexical preactivation. MEG activity was recorded while participants performed a lexical decision task on the second word of visually presented adjective– noun phrases (e.g., stainless steel ). The behavioral results Volume 27, Number 10 showed that predictable and frequent nouns were recognized faster, replicating previous results (Fischler & Bloom, 1979; Whaley, 1978; Rubenstein et al., 1970, and many others). Neurally, lexical preactivation manifested in increased activity: During the adjective time window, left MTG activity was greater for predictive adjectives (e.g., stainless, which is predictive of steel ). Later, during the noun time window, left MTG activity was significantly reduced for predictable nouns (e.g., steel, which is a predictable continuation of stainless). These results are consistent with the previously observed association of the left MTG with lexical access (Friederici, 2012; Hickok & Poeppel, 2007) as well as with the predictive coding hypothesis, according to which predictive processing modulates the same region implicated in bottom–up processing of a stimulus (Egner et al., 2010; Friston, 2005). Moreover, these results indicate that the nature of this predictive processing is such that neural activity is increased at the point at which a specific prediction is generated, whereas activity is reduced at the point at which the prediction is verified; this corroborates the findings of Dikker and Pylkkänen (2013), in which activity in a left middle temporal ROI (among other regions) was increased during the generation of a specific lexical prediction based on a presented picture and reduced when such predictions were satisfied by the presentation of the expected word. In this study, the left MTG displayed a significant interaction between adjective predictivity and expected noun frequency in what we termed the preactivation time window (∼450–600 msec post-adjective onset). This interaction was driven by a significant effect of expected noun frequency for predictive adjectives (e.g., stainless, which is predictive of steel ), but not for less predictive adjectives (e.g., important, which is not predictive of any particular noun). Later, in what we termed the presented noun lexical access time window (∼200–400 msec postnoun onset), there was a significant interaction between presented noun frequency and presented noun predictability, driven by a significant effect of presented noun frequency for less predictable nouns only (e.g., clue in important clue). These results suggest that participants not only preactivated the likely continuations for predictive adjectives before presentation of any noun, but that this preactivation was sensitive to the frequency of the expected noun. By contrast, in the case of less predictive adjectives, participants waited until the presentation of the noun to access the appropriate lexical representation. This evidence for preactivation argues against the strong form of the integration theory of predictability effects, according to which predictable words are easier to process not because any prediction has taken place before they are read but solely because they are easier to integrate into an existing semantic representation (Norris, 1986). It is still possible that some of the effects of predictability can be attributed to greater ease of integration, but our results, in conjunction with form prediction effects (Dikker & Pylkkänen, 2011; DeLong et al., 2005), suggest that ease of integration cannot be the whole story (see also Smith & Levy, 2013). The fact that a word’s frequency could modulate neural activity before its presentation raises the question of how to understand such an effect within the explanatory frameworks used to understand word frequency effects more generally. Rational models of reading (Smith & Levy, 2008, 2013; Norris, 2006) emphasize the influence of predictability on word recognition. Within such a framework, readers optimize their behavior on the basis of their estimates for the likelihood of upcoming words. In the absence of context, word frequency is taken as the baseline expectation for encountering a word. Given the presentation of a high-frequency (or highly predictable) word, a reader might require less perceptual evidence to decide on its identity (Norris, 2006) or less processing time because of prior preparation (Smith & Levy, 2008). One important implication of this approach to word recognition is that unconditional word frequency should be irrelevant when a word is highly predictable from context. Consistent with this hypothesis as well as with prior EEG findings (Dambacher et al., 2006; Van Petten & Kutas, 1990), we found that the response in the MTG to a presented noun was only modulated by its frequency when the noun was a less predictable continuation of the preceding adjective. In summary, our main effects of predictability, as well as the reduction in frequency effects for predictable nouns, largely validate the rational models’ emphasis on predictability as the central determinant of reading behavior. However, although the pervasive effects of predictability indeed suggest that the rational models are on the right track, a strict interpretation of such a model proposes that frequency effects should be entirely accounted for by predictability (Smith & Levy, 2008), a position that is inconsistent with our finding of frequency effects for anticipated nouns before their presentation. Instead, these results suggest that frequency effects are associated with the very process of lexical access itself. Although it is not immediately obvious how to account for this phenomenon within the framework of the rational models, several prominent models within the psycholinguistic literature crucially predict this phenomenon. For example, according to Morton’s (1969) Logogen model, word frequency determines the resting level of activation for a lexical item; similarly, according to Forster’s (1976) Serial Search model, the lexicon is composed of frequency-ordered bins. Thus, our experiment can be seen as providing some new evidence for a long-held view within the psycholinguistic literature, in which frequency effects arise because of the architecture of the lexicon. The language production literature suggests another intriguing interpretation of the effect of expected noun frequency. A family of recent models argues that lexical prediction employs some of the same mechanisms as Fruchter et al. 1925 language production (Dell & Chang, 2014; Federmeier, 2007; Pickering & Garrod, 2007). For example, older adults with high verbal fluency scores show stronger prediction effects than those with lower fluency scores, suggesting that predicted words may be actively generated using the language production system (Federmeier, McLennan, Ochoa, & Kutas, 2002). Frequency effects in production have been extensively documented (Strijkers, Costa, & Thierry, 2010; Kittredge, Dell, Verkuilen, & Schwartz, 2008; Jescheniak & Levelt, 1994; Oldfield & Wingfield, 1965). These effects seem to arise both at the form level and at the semantic level (Kittredge et al., 2008); information at both of these levels needs to be accessed to generate form predictions (Dikker & Pylkkänen, 2011) and semantic feature predictions (Federmeier & Kutas, 1999). The expected noun frequency effects in our experiment may therefore reflect the retrieval of the predicted concept, the retrieval of the orthographic form associated with it, or both. Our main analysis focused on the left MTG, based on research implicating it in semantic and lexical access (Friederici, 2012). Exploratory analysis of left hemisphere language areas showed that effects generally localized to an anterior section of the temporal lobe (cf. Lau, Weber, Gramfort, Hämäläinen, & Kuperberg, 2014, who reported a similar location for the effects of lexical–semantic prediction), as well as to a portion of the inferior frontal lobe. Post hoc ROI analyses also confirmed effects of the variables of interest in regions outside the MTG: a temporal region, the STG, and a prefrontal region, the LOFC. The STG is standardly assumed to be part of the language network; the effects in that region were qualitatively similar to, though weaker than, the effects we found in the MTG. The prefrontal effects are consistent with the role of the pFC in anticipatory processing (Dikker & Pylkkänen, 2013; Bar, 2007). Somewhat unexpectedly, the prefrontal effects localized to LOFC rather than to the IFG, which is the prefrontal region more commonly associated with the language network. It is possible that the spatial distribution of the prefrontal effects may be due to a source localization error. However, a recent MEG study, using the same source space analysis methodology as this study, found effects of a semantic variable at ∼300–500 msec in the LOFC (Fruchter & Marantz, 2015); moreover, a different analysis technique has localized prefrontal lexical prediction effects to ventromedial pFC (Dikker & Pylkkänen, 2013), which is closer to LOFC than to IFG. Multimodal recordings might be able to shed light on the precise localization of this effect. Finally, post hoc analysis revealed equivocal evidence for early (∼100– 200 msec) presented noun predictability effects within the MTG and the left cuneus, consistent with previous findings of lexical predictability effects in early sensory responses (Kim & Lai, 2012; Dikker & Pylkkänen, 2011), although the early effects reported here did not reach statistical significance when control variables were included in the regression models. 1926 Journal of Cognitive Neuroscience There are further aspects of our study that remain open as future avenues of investigation. We quantified the predictability of the second word in a phrase via corpus TP. It is an open question how closely this corpus measure would relate to an empirically derived cloze probability measure, the traditional stand-in for predictability (see Smith & Levy, 2011, for a comparison of sentential cloze probabilities with corpus measures of predictability). In particular, any interpretation of a TP effect confounds prediction based on raw co-occurrence statistics with prediction based on semantics and world knowledge (Frisson, Rayner, & Pickering, 2005). Although not easily distinguishable in this study, the potentially independent effects of these two sources of information could be investigated in a future study. In this study, participants performed a lexical decision after each phrase. This task, although a useful tool to ensure that participants are paying attention to the materials, may have engaged conscious prediction strategies that are not recruited during naturalistic language comprehension (Neely, 1991). A conscious prediction strategy, developed over the course of the experiment, would likely manifest as an increased effect of predictability in later trials compared to earlier ones; such an effect was not observed. Nevertheless, it is worth investigating whether the effects reported here would generalize to a more ecologically valid paradigm, such as a passive reading task. Finally, the primary index of preactivation in our study was the frequency of the most likely noun continuation. Clearly, there is reason to suspect that readers might predict more than a single possible continuation. A preliminary step toward addressing this possibility was taken in this study; the weighted average of the frequencies of possible continuations was shown to slightly improve the model fit relative to the frequency of the single most likely continuation, although this difference did not reach statistical significance. The latter point provides some tentative evidence in favor of a richer conceptualization of lexical preactivation. Hopefully, future work will serve to further characterize the nature of such preactivation, particularly the extent to which possible continuations are preactivated in proportion to their conditional probability given the preceding context. Conclusion This study used MEG to probe the neural signals that correspond to the generation of a lexical prediction, using minimal adjective–noun phrases such as stainless steel. We observed an increase in activity in the left MTG in response to the presentation of more highly predictive adjectives (e.g., stainless). Later, though still before the presentation of the noun, neural activity was modulated by the frequency of the predicted noun (steel ). Correspondingly, when the noun was later presented, predictable nouns elicited weaker neural activity than unpredictable ones. Volume 27, Number 10 APPENDIX. List of Stimuli and Associated TPs (before Logarithmic Transformation) APPENDIX. (continued) Adjective Noun TP Adjective Noun TP residual limb .221 unsalted butter .845 vast majority .217 stainless steel .824 bilingual education .214 barbed wire .793 septic tank .211 umbilical cord .663 everyday life .208 iced tea .575 rectal exam* .206 soapy water .539 cellular phone .205 renewable energy .523 anaerobic digestion .203 pubic hair* .521 hallowed ground .200 undivided attention .453 marital status .199 untimely death* .423 slippery slope .199 concerted effort .421 oncoming traffic .197 immune system .395 catalytic converter .193 salivary gland .389 salutary effect .184 airtight container .378 habitable zone .184 soy sauce .372 bearded man .182 rheumatic fever .355 digestive tract .179 runny nose .354 foreign policy .168 watchful eye .346 illicit drug .165 mental health .311 auditory canal .162 sour cream .307 deviant behavior .161 uncharted territory .307 unholy alliance .159 cervical cancer* .306 uncanny ability .152 taxable income .303 allergic reaction .152 cloudless sky .302 husky voice .152 ballistic missile .294 wooded area .151 unborn child .289 breakneck pace .151 prickly pear .278 pivotal role .150 thankless job .271 angular momentum .150 soluble fiber .268 abdominal pain .144 martial law .262 elective office .144 powdered sugar .260 daunting task .143 eminent domain .245 incurable disease .141 high school .240 facial nerve .140 toothy grin .239 saline solution .140 magnetic field .233 empirical evidence .138 leaded glass .232 impartial spectator .137 crude oil .224 floppy disk .135 jobless rate .223 lethal injection* .134 Fruchter et al. 1927 APPENDIX. (continued) APPENDIX. (continued) Adjective Noun TP negligent homicide* .131 undue burden .086 crusty bread .131 postwar period .084 electoral college .127 unsolved murder* .083 outer space .124 cerebral cortex .082 virtual reality .118 nonprofit group .082 lifeless body* .116 honorary doctorate .082 unskilled labor .115 cubic foot .081 royal family .115 radial velocity .080 boneless pork .112 aerobic fitness .080 fictional world .110 disabled list .079 humid air .109 schematic diagram .079 custodial parent .108 keen interest .078 stormy weather .107 imminent danger .077 evasive action .104 glacial ice .075 indecent exposure .101 sane person .075 sore throat .099 teenage girl .074 intrinsic value .099 periodic table .074 unwed mother .098 outspoken critic .073 tropical storm .098 minor league .073 bald head .098 lunar surface .073 timely manner .098 naval base .072 coercive power .097 traumatic event .072 artistic director .096 unanimous decision .071 unmarked car .096 brisk business .071 leaky roof .095 utopian vision .071 hind leg .095 natural gas .071 domestic violence* .095 nominal fee .070 frontal lobe .094 volcanic activity .070 populous country .093 bridal gown .069 infinite number .092 molten lava .069 vicious cycle .091 rightful owner .069 sluggish economy .091 vaginal dryness* .069 thorny issue .090 violent crime* .067 exclusive interview .089 literal sense .067 nasal cavity .089 wide variety .066 offensive line .089 ultimate goal .065 bacterial infection .087 oily skin .065 rapid growth .087 private sector .065 doctoral degree .086 perpetual motion .064 1928 Journal of Cognitive Neuroscience Adjective Noun TP Volume 27, Number 10 APPENDIX. (continued) Adjective APPENDIX. (continued) Noun TP Adjective Noun TP lanky frame .064 stony silence .045 fertile soil .064 spectral type .045 untreated sewage .063 discreet distance .044 sexual abuse* .063 viable option .044 immortal soul .061 indoor plumbing .043 regular basis .059 stylistic analysis .043 notable exception .057 irregular heartbeat .043 skeletal muscle .056 romantic comedy .043 sole purpose .055 canned food .042 gay marriage .055 upper lip .042 prolific writer .055 covert operation .042 pungent odor .055 dismal failure .041 rugged terrain .055 rosy picture .040 surgical procedure .054 unfair advantage .040 miniature golf .054 inaugural ball .040 explosive device .054 universal coverage .040 unequal treatment .053 awkward position .040 radiant heat .052 patriotic duty .039 spinal column .052 sheer size .039 enormous amount .051 stellar evolution .038 tentative agreement .051 unwanted pregnancy* .038 impending doom .051 generic term .037 khaki shirt .051 ample room .037 raw material .050 decisive victory .036 popular culture .049 homeless shelter .035 speedy trial .049 earthly paradise .035 candid camera .049 sizable chunk .035 stressful situation .049 sensual pleasure .035 bad news .048 deaf ear .034 temperate climate .048 muscular strength .034 adverse impact .048 planetary scientist .033 funny thing .046 coarse meal .033 electric mixer .046 sweet potato .033 petite woman .046 normative sample .033 sensory input .046 factual knowledge .032 receptive audience .045 judicial activism .032 modernist art .045 lifelong friend .031 primal scene .045 geometric pattern .031 ethnic identity .045 lively debate .031 Fruchter et al. 1929 APPENDIX. (continued) Adjective APPENDIX. (continued) Noun TP Adjective Noun TP barren landscape .031 reliable source .022 festive mood .031 thermal expansion .022 optimal level .031 broad daylight .022 orbital debris .030 tragic accident .022 liberal democracy .030 upcoming book .022 damp cloth .030 durable peace .021 optical illusion .029 dental hygiene .021 extra money .029 casual observer .021 ancestral homeland .029 annual budget .021 shallow dish .028 dirty laundry .021 monthly payment .028 mere fact .021 fiscal crisis .028 static pressure .021 insane asylum .027 blind date .021 nervous breakdown .027 fuzzy logic .021 polite applause .027 spiritual leader .021 biblical text .026 vibrant color .021 clinical practice .026 nice guy .020 genetic diversity .026 wild card .020 serious problem .025 cardiac output .020 socialist realism .025 logical extension .020 strict liability .025 identical twin .020 integral component .025 slim chance .020 senior editor .025 free agent .019 exact location .025 incoming freshman .019 creamy texture .025 fresh lemon .019 preschool teacher .024 cautious optimism .019 main reason .024 lucrative contract .019 glossy magazine .024 potent symbol .019 rigorous training .024 weekly newspaper .019 dumb luck .024 solar radiation .018 digital video .024 endless stream .018 dietary intake .024 rational choice .018 vivid memory .024 portable radio .018 sleepy town .024 automatic pilot .018 polar cap .024 ambitious project .018 soft tissue .024 eternal damnation* .018 linear model .023 civic pride .018 drunken driver .023 cruel joke .018 fatal flaw* .023 tall grass .017 1930 Journal of Cognitive Neuroscience Volume 27, Number 10 APPENDIX. (continued) Adjective APPENDIX. (continued) Noun TP Adjective Noun TP academic success .017 financial planner .013 exotic dancer .017 organic carbon .013 positive attitude .017 atomic physics .013 cultural heritage .017 gentle breeze .013 genuine concern .017 valuable lesson .013 ripe tomato .017 conscious awareness .013 good idea .016 aesthetic quality .012 humble opinion .016 distant cousin .012 frequent flier .016 explicit reference .012 fair game .016 dense foliage .012 urgent message .016 suitable habitat .012 full moon .016 tiny fraction .012 seasonal flu .016 economic reform .011 fierce battle .016 terrible tragedy .011 silent auction .016 religious belief .011 symbolic capital .016 retail industry .011 stable condition .016 faithful servant .011 absolute certainty .016 coastal region .011 uneasy truce .016 nuclear arsenal .011 informal survey .016 healthy diet .011 indirect discourse .015 vague notion .011 jealous rage .015 ethical dilemma .011 racial equality .015 urban renewal .011 selective breeding .015 neutral hydrogen .011 immediate aftermath .015 practical advice .011 precious commodity .015 bitter pill .011 black pepper .015 stiff neck .010 joint statement .015 eastern seaboard .010 bare chest .014 dominant theme .010 guilty plea .014 original sin .010 useful tool .014 dull knife .010 secular humanism .014 permanent residence .010 extreme poverty .014 innocent victim .010 inner self .014 moral theology .010 physical therapy .014 risky strategy .009 spatial scale .014 personal trainer .009 nasty stuff .014 anonymous donor .009 constant reminder .013 divine creation .009 steep incline .013 blue cheese .009 Fruchter et al. 1931 APPENDIX. (continued) Adjective APPENDIX. (continued) Noun TP Adjective Noun TP pure vanilla .009 recent poll .005 remote corner .009 critical acclaim .005 low profile .009 efficient method .005 defensive posture .009 athletic shoe .005 narrow path .009 angry mob .005 mild recession .009 quick trip .005 abstract concept .009 dangerous precedent .005 empty stomach .009 crucial aspect .005 temporary relief .008 dramatic reduction .005 northern border .008 rare occasion .004 brilliant career .008 similar vein .004 deadly virus* .008 common stock .004 rough patch .008 creative genius .004 bright sunlight .008 entire universe .004 accurate diagnosis .008 sad song .004 slight movement .008 illegal gambling .004 corporate ladder .008 military campaign .004 severe drought .007 active volcano .003 ongoing dialogue .007 proper burial .003 honest broker .007 strong supporter .003 native tongue .007 musical notation .003 yellow squash .007 huge crowd .003 visual imagery .007 perfect timing .003 negative publicity .007 strange sensation .003 elderly gentleman .007 sick bay .003 regional stability .007 massive influx .003 southern accent .007 rural county .003 modest proposal .007 thin sheet .003 emotional intensity .006 current crop .002 previous page .006 easy prey .002 safe passage .006 powerful engine .002 medical marijuana .006 famous phrase .002 large pot .006 major obstacle .002 cheap plastic .006 expensive jewelry .002 heavy saucepan .006 quiet dignity .002 legal pad .006 tough stance .002 apparent suicide* .006 local chapter .002 basic premise .005 fine mist .002 formal complaint .005 beautiful scenery .002 1932 Journal of Cognitive Neuroscience Volume 27, Number 10 APPENDIX. (continued) Adjective REFERENCES Noun TP modern reader .002 political rhetoric .002 young adulthood .001 dead giveaway* .001 small intestine .001 possible scenario .001 important clue .000 *Denotes high valence items, which were removed for the supplemental MTG analysis. Acknowledgments This material is based upon work supported by the National Science Foundation under grant BCS-0843969 and by the NYU Abu Dhabi Research Council under grant G1001 from the NYUAD Institute, New York University Abu Dhabi. We would like to thank three anonymous reviewers for their informative feedback on this manuscript. Reprint requests should be sent to Joseph Fruchter, Department of Psychology, New York University, 2nd Floor, 6 Washington Place, New York, NY 10003, or via e-mail: [email protected]. Notes 1. Alternatively, for an individual trial, one might regard a participant as predicting only a single possible noun with a probability equal to that noun’s TP; on the aggregate, however, we would nevertheless observe effects proportional to the relevant conditional probabilities. 2. Because there was no definition for the insula region within the parcellation, we obtained the region via an alternate parcellation for the average brain, which was then morphed back to each participant’s neuroanatomical space. 3. Because the M350 response is negatively signed, a positive correlation indicates a weakening of activity, and a negative correlation indicates a strengthening of activity. 4. Despite the highly significant adjective predictivity effect in the main analysis (see also Figure 3C), a median split analysis failed to show a comparably robust separation between the items in the top half and bottom half of adjective predictivity. This discrepancy indicated that the continuous regression using linear mixed-effects models was a more sensitive measure of the adjective predictivity effect. Consequently, we decided to split the data into the top 10% (blue line), top 10–50% (solid black line), bottom 10–50% (dotted black line), and bottom 10% (red line) of adjective predictivity (Figure 3B), which confirmed our hypothesis. In particular, the continuous regression models are a better fit to the data than the median split, because the predictivity effect is more significant at the higher and lower ranges of adjective predictivity values, relative to the items in the middle of the distribution. 5. The latency of the M100 peak in Dikker and Pylkkänen’s (2011) study was 97 msec, which is roughly midway between the two peaks observed here; we thus decided to analyze both peaks in the present data. It should be noted, however, that Dikker and Pylkkänen (2011) performed a sensor space analysis, which may yield results that are not comparable to the results of the present source space analysis. Adachi, Y., Shimogawara, M., Higuchi, M., Haruta, Y., & Ochiai, M. (2001). Reduction of non-periodic environmental magnetic noise in MEG measurement by continuously adjusted least squares method. IEEE Transactions on Applied Superconductivity, 11, 669–672. Altmann, G. T. M., & Kamide, Y. (1999). Incremental interpretation at verbs: Restricting the domain of subsequent reference. Cognition, 73, 247–264. Baayen, R. H., Davidson, D. J., & Bates, D. M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. Journal of Memory and Language, 59, 390–412. Bar, M. (2007). The proactive brain: Using analogies and associations to generate predictions. Trends in Cognitive Sciences, 11, 280–289. Bates, D., Maechler, M., & Bolker, B. (2013). lme4: Linear mixed-effects models using S4 classes. R package version 0.999999-2. http://CRAN.R-project.org/package=lme4. Bemis, D. K., & Pylkkänen, L. (2011). Simple composition: A magnetoencephalography investigation into the comprehension of minimal linguistic phrases. Journal of Neuroscience, 31, 2801–2814. Binder, J. R., Frost, J. A., Hammeke, T. A., Cox, R. W., Rao, S. M., & Prieto, T. (1997). Human brain language areas identified by functional magnetic resonance imaging. Journal of Neuroscience, 17, 353–362. Brainard, D. H. (1997). The Psychophysics Toolbox. Spatial Vision, 10, 433–436. Brown, C., & Hagoort, P. (1993). The processing nature of the N400: Evidence from masked priming. Journal of Cognitive Neuroscience, 5, 34–44. Dale, A. M., Liu, A. K., Fischl, B. R., Buckner, R. L., Belliveau, J. W., Lewine, J. D., et al. (2000). Dynamic statistical parametric mapping: Combining fMRI and MEG for high resolution imaging of cortical activity. Neuron, 26, 55–67. Dale, A. M., & Sereno, M. I. (1993). Improved localization of cortical activity by combining EEG and MEG with MRI cortical surface reconstruction: A linear approach. Journal of Cognitive Neuroscience, 5, 162–176. Dambacher, M., Kliegl, R., Hofmann, M., & Jacobs, A. M. (2006). Frequency and predictability effects on event-related potentials during reading. Brain Research, 1084, 89–103. Davies, M. (2009). The 385+ million word Corpus of Contemporary American English (1990–2008+): Design, architecture, and linguistic insights. International Journal of Corpus Linguistics, 14, 159–190. Dell, G. S., & Chang, F. (2014). The P-chain: Relating sentence production and its disorders to comprehension and acquisition. Philosophical Transactions of the Royal Society, Series B, Biological Sciences, 369, 20120394. DeLong, K., Urbach, T., & Kutas, M. (2005). Probabilistic word preactivation during language comprehension inferred from electrical brain activity. Nature Neuroscience, 8, 1117–1121. Desikan, R. S., Ségonne, F., Fischl, B., Quinn, B. T., Dickerson, B. C., Blacker, D., et al. (2006). An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral based regions of interest. Neuroimage, 31, 968–980. Dikker, S., & Pylkkänen, L. (2011). Before the N400: Effects of lexical-semantic violations in visual cortex. Brain and Language, 118, 23–28. Dikker, S., & Pylkkänen, L. (2013). Predicting language: MEG evidence for lexical preactivation. Brain and Language, 127, 55–64. Dikker, S., Rabagliati, H., & Pylkkänen, L. (2009). Sensitivity to syntax in visual cortex. Cognition, 110, 293–321. Egner, T., Monti, J. M., & Summerfield, C. (2010). Expectation and surprise determine neural population responses in the Fruchter et al. 1933 ventral visual stream. Journal of Neuroscience, 30, 16601–16608. Ehrlich, S. F., & Rayner, K. (1981). Contextual effects on word perception and eye movements during reading. Journal of Verbal Learning and Verbal Behavior, 20, 641–655. Embick, D., Hackl, M., Schaeffer, J., Kelepir, M., & Marantz, A. (2001). A magnetoencephalographic component whose latency reflects lexical frequency. Cognitive Brain Research, 10, 345–348. Federmeier, K. D. (2007). Thinking ahead: The role and roots of prediction in language comprehension. Psychophysiology, 44, 491–505. Federmeier, K. D., & Kutas, M. (1999). A rose by any other name: Long-term memory structure and sentence processing. Journal of Memory and Language, 41, 469–495. Federmeier, K. D., McLennan, D. B., Ochoa, E., & Kutas, M. (2002). The impact of semantic memory organization and sentence context information on spoken language processing by younger and older adults: An ERP study. Psychophysiology, 39, 133–146. Fischler, I., & Bloom, P. A. (1979). Automatic and attentional processes in the effects of sentence contexts on word recognition. Journal of Verbal Learning and Verbal Behavior, 18, 1–20. Forster, K. I. (1976). Accessing the mental lexicon. In R. J. Wales & E. C. T. Walker (Eds.), New approaches to language mechanisms (pp. 257–287). Amsterdam: North-Holland. Francken, J. C., Kok, P., Hagoort, P., & De Lange, F. P. (2015). The behavioral and neural effects of language on motion perception. Journal of Cognitive Neuroscience, 27, 175–184. Friederici, A. D. (2012). The cortical language circuit: From auditory perception to sentence comprehension. Trends in Cognitive Sciences, 16, 262–268. Frisson, S., Rayner, K., & Pickering, M. J. (2005). Effects of contextual predictability and transitional probability on eye movements during reading. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31, 862–877. Friston, K. (2005). A theory of cortical responses. Philosophical Transactions of the Royal Society, Series B, 360, 815–836. Fruchter, J., & Marantz, A. (2015). Decomposition, lookup, and recombination: MEG evidence for the full decomposition model of complex visual word recognition. Brain and Language, 143, 81–96. Hämäläinen, M., Hari, R., Ilmoniemi, R. J., Knuutila, J., & Lounasmaa, O. V. (1993). Magnetoencephalography— Theory, instrumentation, and applications to noninvasive studies of the working human brain. Review of Modern Physics, 65, 413–497. Hickok, G., & Poeppel, D. (2007). The cortical organization of speech processing. Nature Reviews Neuroscience, 8, 393–402. Holmes, A. P., & Friston, K. J. (1998). Generalisability, random effects and population inference. Neuroimage, 7, S754. Indefrey, P., & Levelt, W. J. M. (2004). The spatial and temporal signatures of word production components. Cognition, 92, 101–144. Inhoff, A. W., & Rayner, K. (1986). Parafoveal word processing during eye fixations in reading: Effects of word frequency. Perception & Psychophysics, 40, 431–439. Jescheniak, J. D., & Levelt, W. J. (1994). Word frequency effects in speech production: Retrieval of syntactic information and of phonological form. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 824. Kamide, Y., Altmann, G., & Haywood, S. L. (2003). The timecourse of prediction in incremental sentence processing: Evidence from anticipatory eye movements. Journal of Memory and Language, 49, 133–156. 1934 Journal of Cognitive Neuroscience Keuleers, E., & Brysbaert, M. (2010). Wuggy: A multilingual pseudoword generator. Behavior Research Methods, 42, 627–633. Kim, A., & Lai, V. (2012). Rapid interactions between lexical semantic and word form analysis during word recognition in context: Evidence from ERPs. Journal of Cognitive Neuroscience, 24, 1104–1112. Kittredge, A. K., Dell, G. S., Verkuilen, J., & Schwartz, M. F. (2008). Where is the effect of frequency in word production? Insights from aphasic picture-naming errors. Cognitive Neuropsychology, 25, 463–492. Kok, P., Failing, M. F., & de Lange, F. P. (2014). Prior expectations evoke stimulus templates in the primary visual cortex. Journal of Cognitive Neuroscience, 26, 1546–1554. Kok, P., Jehee, J. F., & de Lange, F. P. (2012). Less is more: Expectation sharpens representations in the primary visual cortex. Neuron, 75, 265–270. Kutas, M., & Federmeier, K. D. (2000). Electrophysiology reveals semantic memory use in language comprehension. Trends in Cognitive Sciences, 4, 463–470. Kutas, M., & Hillyard, S. A. (1984). Brain potentials during reading reflect word expectancy and semantic association. Nature, 307, 161–163. Lau, E. F., Phillips, C., & Poeppel, D. (2008). A cortical network for semantics: (De)constructing the N400. Nature Reviews Neuroscience, 9, 920–933. Lau, E. F., Weber, K., Gramfort, A., Hämäläinen, M. S., & Kuperberg, G. R. (2014). Spatiotemporal signatures of lexical-semantic prediction. Cerebral Cortex. Advance online publication. doi:10.1093/cercor/bhu219. Linzen, T., Marantz, A., & Pylkkänen, L. (2013). Syntactic context effects in visual word recognition: An MEG study. The Mental Lexicon, 8, 117–139. McDonald, S. A., & Shillcock, R. C. (2003). Low-level predictive inference in reading: The influence of transitional probabilities on eye movements. Vision Research, 43, 1735–1751. Morton, J. (1969). Interaction of information in word recognition. Psychological Review, 76, 165–178. Neely, J. H. (1991). Semantic priming effects in visual word recognition: A selective review of current findings and theories. In D. Besner & G. W. Humphreys (Eds.), Basic processes in reading: Visual word recognition (pp. 264–336). Hillsdale, NJ: Erlbaum. Norris, D. (1986). Word recognition: Context effects without priming. Cognition, 22, 93–136. Norris, D. (2006). The Bayesian reader: Explaining word recognition as an optimal Bayesian decision process. Psychological Review, 113, 327–357. Oldfield, R. C. (1971). The assessment and analysis of handedness: The Edinburgh inventory. Neuropsychologia, 9, 97–113. Oldfield, R. C., & Wingfield, A. (1965). Response latencies in naming objects. Quarterly Journal of Experimental Psychology, 17, 273–281. Pelli, D. G. (1997). The VideoToolbox software for visual psychophysics: Transforming numbers into movies. Spatial Vision, 10, 437–442. Pickering, M. J., & Garrod, S. (2007). Do people use language production to make predictions during comprehension?. Trends in Cognitive Sciences, 11, 105–110. Pylkkänen, L., & Marantz, A. (2003). Tracking the time course of word recognition with MEG. Trends in Cognitive Sciences, 7, 187–189. Rodd, J. M., Davis, M. H., & Johnsrude, I. S. (2005). The neural mechanisms of speech comprehension: fMRI studies of semantic ambiguity. Cerebral Cortex, 15, 1261–1269. Volume 27, Number 10 Rubenstein, H., Garfield, L., & Millikan, J. A. (1970). Homographic entries in the internal lexicon. Journal of Verbal Learning and Verbal Behavior, 9, 487–494. Smith, M. E., & Halgren, E. (1987). Event-related potentials during lexical decision: Effects of repetition, word frequency, pronounceability, and concreteness. Electroencephalography and Clinical Neurophysiology Supplement, 40, 417–421. Smith, N. J., & Levy, R. (2008). Optimal processing times in reading: A formal model and empirical investigation. In B. C. Love, K. McRae, & V. M. Sloutsky (Eds.), Proceedings of the Thirtieth Annual Conference of the Cognitive Science Society (pp. 595–600). Austin, TX: Cognitive Science Society. Smith, N. J., & Levy, R. (2011). Cloze but no cigar: The complex relationship between cloze, corpus, and subjective probabilities in language processing. In L. Carlson, C. Hölscher, & T. Shipley (Eds.), Proceedings of the 33rd Annual Conference of the Cognitive Science Society (pp. 1637–1642). Austin, TX: Cognitive Science Society. Smith, N. J., & Levy, R. (2013). The effect of word predictability on reading time is logarithmic. Cognition, 128, 302–319. Solomyak, O., & Marantz, A. (2009). Lexical access in early stages of visual word processing: A single-trial correlational MEG study of heteronym recognition. Brain and Language, 108, 191–196. Solomyak, O., & Marantz, A. (2010). Evidence for early morphological decomposition in visual word recognition. Journal of Cognitive Neuroscience, 22, 2042–2057. Strijkers, K., Costa, A., & Thierry, G. (2010). Tracking lexical access in speech production: Electrophysiological correlates of word frequency and cognate effects. Cerebral Cortex, 20, 912–928. Taylor, W. L. (1953). “Cloze procedure”: A new tool for measuring readability. Journalism Quarterly, 30, 415–433. Van Berkum, J. J. A., Brown, C. M., Zwitserlood, P., Kooijman, V., & Hagoort, P. (2005). Anticipating upcoming words in discourse: Evidence from ERPs and reading times. Journal of Experimental Psychology: Learning, Memory, and Cognition, 31, 443–467. Van Petten, C., & Kutas, M. (1990). Interactions between sentence context and word frequency in event-related brain potentials. Memory & Cognition, 18, 380–393. Whaley, C. P. (1978). Word-nonword classification time. Journal of Verbal Learning and Verbal Behavior, 17, 143–154. Wicha, N. Y., Moreno, E. M., & Kutas, M. (2003). Expecting gender: An event related brain potential study on the role of grammatical gender in comprehending a line drawing with a written sentence in Spanish. Cortex, 39, 483–508. Fruchter et al. 1935