Stimulus-independent neural coding of event semantics: Evidence from cross-sentence fMRI decoding

Multivariate neuroimaging studies indicate that the brain represents word and object concepts in a format that readily generalises across stimuli. Here we investigated whether this was true for neural representations of events described using sentences. Participants viewed sentences describing four events in different ways. Multivariate classifiers were trained to discriminate the four events using a subset of sentences, allowing us to test generalisation to novel sentences. We found that neural patterns in a left-lateralised network of frontal, temporal and parietal regions discriminated events in a way that generalised successfully over changes in the syntactic and lexical properties of the sentences used to describe them. In contrast, decoding in visual areas was sentence-specific and failed to generalise to novel sentences. In the reverse analysis, we tested for decoding of syntactic and lexical form, independent of the event being described. Regions displaying this coding were limited and largely fell outside the canonical semantic network. Our results indicate that a distributed neural network represents the meaning of event sentences in a way that is robust to changes in their structure and form. They suggest that the semantic system disregards the surface properties of stimuli in order to represent their underlying conceptual significance.

2018; Frankland and Greene, 2020). Such models can successfully predict activation patterns to new sentences that describe different events. Activation in left temporal and parietal cortices appears to be critical in supporting such predictions.
Cross-language studies have begun to use similar techniques to investigate whether neural representations of sentence meaning generalise across changes in linguistic form. Hu et al. (2019) trained a multivariate classifier to discriminate between coherent and incoherent sentence pairs presented to bilinguals in Chinese or Japanese script. Patterns in left parietal cortex and inferior prefrontal cortex showed successful cross-language classification, suggesting that a similar neural representation was engaged by both languages. However, in this study the classifier was not required to discriminate the content of the sentences, only whether they described a coherent event. In contrast, Yang et al. (2017) developed a feature-based semantic model to code the content of 60 distinct English sentences. This model was trained with fMRI data from English speakers to form a predictive model of sentence activation patterns across the whole brain. Once trained, the model could successfully predict neural responses to Portuguese translations of the sentences in a new group of Portuguese speakers. This study demonstrated that the areas of the brain activated by particular events are broadly similar even across different individuals who speak different languages. However, it was not designed to identify which specific brain regions code event information in a stimulus-independent way. In the present study we aimed to do this in a multivariate fMRI study that held constant the language used to describe events, but systematically varied the linguistic forms of the descriptions.
We investigated neural representations of four distinct events, systematically varying the syntactic structures and lexical items used to describe each event. We aimed to identify brain regions whose activity patterns discriminated between different events and then to assess whether this discrimination was robust to changes in the syntactic and lexical properties of the sentences. We predicted that the regions of the left-lateralised canonical semantic network, previously linked with supramodal representation of words and objects, would show stimulus-independent coding of event meanings. We then reversed our analysis, testing for brain regions that could successfully discriminate between syntactic structures, while generalising over the semantics of the events. Finally, we tested for regions that discriminated between different lexical items used to describe the same events. Together, these analysis allowed to distinguish between brain regions that process the conceptual content of sentences and those involved in processing their form.

Method
Participants: 26 native English speakers took part in the study (20 female; mean age 22.48, range 18-35 years). All were classified as right-handed using the Edinburgh Handedness Inventory (Oldfield, 1971) and none reported dyslexia or any history of neurological illness.
All provided informed consent.
Stimuli: We probed neural representations of four coherent events designed to cover a broad range of conceptual knowledge, including living and non-living agents (cow, student; computer, lorry), motion and static verbs (driving, jumping; processing, considering) and concrete and abstract patients (fence, bridge; files, problems). Each event was presented to participants in four different sentence forms that varied in their syntactic structure and lexical items (see Table 1 for examples). Syntactic structure was varied by describing the event with either an active or passive sentence. Lexical items were varied by substituting semantically-similar words to describe each agent, patient and action. Thus, in total there were 16 event sentences used in our multivariate analyses. We also created 16 anomalous sentences that used the same lexical items and syntactic structures as the event sentences but did not describe a coherent event. These were created by reshuffling the constituents of the event sentences (e.g., "the document was considered by the lorry"). To validate this manipulation, 18 participants, who did not take part in the main study, were asked to rate the meaningfulness of each sentence on a five-point scale. Event sentences received significantly higher ratings than anomalous sentences (Meaningful M = 4.56, SD = 0.32; Anomalous M = 1.53, SD = 0.57; t(30) = 18.6, p < 0.001). The pupil pondered the issue.

Passive form
The problem was considered by the student.
The issue was pondered by the pupil.
Procedure: Participants completed six runs of scanning. In each run, they were presented with each of the 32 sentences once and were asked to decide whether the sentence was meaningful or not. Manual responses were made using the left and right hands, with the mapping of these to response options counterbalanced over participants. Each trial began with a fixation cross presented for 500ms. This was followed by the written sentence, presented in the centre of the screen for 4s. Trials were separated by a jittered interstimulus interval of between 4s and 8s (mean 6s). The order of the sentences was fully randomised for each run and for each participant.
Image acquisition and preprocessing: Images were acquired on a 3T Siemens Prisma scanner using a 32-channel head coil. We employed a whole-brain multi-echo acquisition protocol, in which data was simultaneously acquired at three echo times (13ms, 31ms and 48ms). Data from the three echo series were weighted and combined and the resulting time-series denoised using independent components analysis (ICA). This protocol reduces the influence of motion and other artefacts and, importantly, improves signal quality in regions that typically suffer from susceptibility artefacts, such as the ventral anterior temporal lobes . The TR was 1.7s and images consisted of 46 slices with an 80 x 80 matrix and isotropic voxel size of 3mm. Multiband acceleration with a factor of 2 was used and the flip angle was 73°. Six runs of 195 volumes (331.5s) were acquired. A high-resolution T1-weighted structural image was also acquired for each participant using an MP-RAGE sequence with 0.8mm isotropic voxels, TR = 2.62s, TE = 4.5ms.
Images were preprocessed using the TE-Dependent Analysis Toolbox 0.0.7 (Tedana) (DuPre et al., 2019) and SPM12. Motion parameters were estimated using the first echo series, prior to slice-timing correction (as recommended by Power et al., 2017). Slice-timing correction was then applied, images were interpolated to the space of the first volume using the previously obtained motion estimates, and Tedana was used to optimally combine the three echo series into a single time-series. Tedana's denoising algorithms were also applied to the data. These use ICA to partition the data and then to classify each component as either BOLD-related or noise-related, based on its pattern of signal decay over increasing TEs . Components classified as noise were discarded. Following denoising, SPM was used to unwarp images with a B0 fieldmap, coregister them to the anatomical scans and normalise them to MNI space with DARTEL (Ashburner, 2007).
For initial univariate analyses, images were smoothed with a kernel of 8mm FWHM.
Data were treated with a high-pass filter with a cut-off of 128s and the six runs were analysed using a single general linear model. For each run, one regressor modelled presentation of event sentences and another presentation of anomalous sentences. Covariates consisted of six motion parameters and their first-order derivatives.
Multi-voxel pattern analysis: For multivariate analysis, smoothing of 4mm FWHM was applied to the normalised images. Each run was analysed with a separate general linear model which included a separate regressor for each of the 32 sentences. T-maps were generated for each event sentence, which were then submitted to decoding analyses using CoSMoMVPA (Oosterhof et al., 2016). Our main analyses tested for effects across the whole of the cortex using a spherical searchlight with a radius of four voxels. We also tested for decoding effects in five anatomical regions of interest described later. We performed five different analyses that investigated the neural representation of different aspects of sentence information.
Analysis 1: Sentence-specific decoding: The first analysis tested the ability of brain regions to discriminate between different events without requiring generalisation to different descriptions. We selected a set of four sentences that described the four events (with the constraint that they all had the same syntactic structure). We trained the model to discriminate between the four sentences and tested it on new instances of the same sentences (see Figure 1A). This process was repeated for all possible combinations of four sentences representing the four events (32 iterations) and the results averaged to give an overall accuracy map for each participant. Data were partitioned into train and test sets with leaveone-run-out cross-validation: the classifier was trained on data from five runs and tested on the remaining run and this process was repeated until each run had served as the test set.
Analyses 2 and 3: Generalisation of event identity across sentence forms: The second and third analyses tested for brain regions that code event information in a way that generalises to new descriptions of the events. Analysis 2 tested generalisation to new lexical forms and Analysis 3 generalisation to new syntactic forms. In Analysis 2, the classifier was trained each time with data from eight sentences. In this set of eight sentences, a single set of lexical forms was used to describe each event. The classifier was then tested on the remaining eight sentences, which described the same four events using different lexical items (see Figure 2A). Thus, above-chance decoding could only be achieved where brain regions coded information about the events in a way that generalised to novel sentences. The process was repeated until all possible combinations of lexical items had served as the training set (2 4 = 16 iterations). Analysis 3 proceeded in a similar fashion, except that train and test sets differed in the syntactic forms used to describe each event (see Figure 2C).
Analysis 4: Decoding syntactic structure: Here we reversed the logic of the previous analyses and asked whether any brain regions coded the syntactic structure of the sentences, independent of the events being described. The model was trained to discriminate between active and passive sentences. Trials were divided such that eight sentences describing two events were used as the training set and eight sentences describing the remaining two events were used as the test set (see Figure 3A). Thus, above-chance classification could only be achieved if patterns of brain activity discriminated between active and passive sentences in a way that generalised across different events. Decoding was repeated until all possible pairs of events had served as the training set (C(4,2) = 6 iterations).
Analysis 5: Decoding lexical items: Finally, we wanted to test whether any regions coded for the lexical items used in the sentences, independent of events. Here it was not possible to test for generalisation across events as each event was associated with a different set of lexical items. Instead, we treated each event separately and trained the classifier to discriminate between the two sets of lexical forms associated with it (see Figure 3C). The classifier was then tested on new instances of the same sentences, with leave-one-run-out cross-validation used to ensure independence of training and testing data. This process was repeated for each of the four events and the results averaged.
In all analyses, classification was performed using a support vector machine (LIBSVM) with the regularisation parameter C set to 1. To determine whether classification was better than expected by chance, permutation tests were performed using the method introduced by Stelzer et al. (2013). For each participant, we trained and tested the classifier repeatedly on data in which the class labels had been randomly permuted. This process was repeated 100 times (divided equally between all iterations of the training stimuli) to provide an accuracy distribution for each participant under the null hypothesis. Following this, a Monte Carlo approach was taken to generate a null accuracy distribution at the group level. Specifically, from each participant's null distribution, we randomly selected one accuracy map for each training iteration and averaged these to give a group mean. This process was repeated 10,000 times to generate a distribution of the expected group accuracy under the null hypothesis. In ROI analyses, the position of the observed group accuracy in this null distribution was used to determine a p-value (e.g., if the observed accuracy was greater than 99% of values in the null distribution, this would represent a p-value of 0.01). For searchlight analyses, observed and null accuracy maps were entered into CoSMoMVPA's Monte Carlo cluster statistics function, which returned a statistical map corrected for multiple comparisons using thresholdfree cluster enhancement (Smith and Nichols, 2009). These maps were thresholded at corrected p < 0.05.
To aid interpretation of results, unthresholded decoding accuracy maps were submitted to Neurosynth and correlated with its meta-analytic maps (Rubin et al., 2017).
Neurosynth's maps index the strength of association between terms used in the neuroimaging literature and areas of reported activation. By correlating these with our results, we were able to determine which terms are most commonly associated with regions where we observed strong decoding. To visualise the results, terms relating to anatomical structures were removed and the 20 most correlated terms were extracted and plotted as a word cloud.
Regions of interest: ROI analysis focused on left-hemisphere anatomical regions, selected based on their involvement in semantic processing. These are shown in Figure 4A. Four of the five ROIs were defined using probability distribution maps from the Harvard-Oxford brain atlas (Makris et al., 2006), including all voxels with a >30% probability of falling within the The final ROI covered the angular gyrus (AG) and included voxels with a >30% probability of falling within this region in the LPBA40 atlas (Shattuck et al., 2008). A different atlas was used in this case because the AG region defined in the Harvard-Oxford atlas is small and does not include parts of the inferior parietal cortex typically implicated in semantic processing. The 30% inclusion threshold we used to define ROIs is consistent with our previous work (Hoffman, 2019;Hoffman and Tamm, 2020).

Results
Behavioural performance: We first checked the accuracy of the meaningfulness judgements in each imaging run to ensure that participants were responding attentively. Runs in which the participant responded correctly on fewer than 60% of trials were excluded from further analysis. A total of five runs from three participants were excluded; all participants had at least three valid runs for analysis. One additional run was excluded because it was terminated early due to a scanner fault. Participants were highly accurate in responding to event sentences (M = 95.5%; SD = 4.1%) and to anomalous sentences (M = 91.3%; SD = 6.1%), though more accurate for the event sentences (t(25) = 3.02, p = 0.006). They were slower to respond to anomalous sentences (M = 1940ms; SD = 390ms) than to event sentences (M = 1811ms; SD = 388ms; t(25) = 5.19, p < 0.001). We also compared performance for event sentences that used an active vs. passive syntactic structure. Participants were faster to AG. This suggests that semantic judgements about anomalous sentences required greater engagement of the semantic system, consistent with longer reaction times on these trials.
Greater activation for event sentences was found in ventromedial prefrontal and posterior cingulate cortices, consistent with previous data implicating these regions specifically in the processing of coherent events (Lerner et al., 2011).
Searchlight analyses: Analysis 1 tested whether activity patterns could discriminate between specific sentences describing the four events, without requiring generalisation to new descriptions of the events. Figure 1B shows decoding accuracies in areas where decoding significantly exceeded chance performance. Above-chance decoding was observed across the cortex and bilaterally, with the highest accuracies found in primary and early visual cortices.
Other areas showing high decoding accuracy included left pMTG, regions along the length of the superior temporal gyrus, the lateral prefrontal cortex, particularly IFG, and the dorsomedial frontal lobe, close to the pre-supplementary motor area. In the right hemisphere, decoding performance was generally poorer but exceeded chance levels in many areas. The Neurosynth word cloud ( Figure 1C) indicates that regions displaying high decoding accuracies were associated with terms relating to visual processing and also to language and semantic processes. Supplementary Figure 2 splits the results of this analysis into classifiers trained with active vs. passive sentences. Significant decoding was found in similar regions for both sentence types, though decoding was generally more successful when the analysis was restricted to passive sentences. This might be indicative of stronger neural signals elicited on these trials, which participants took around 250ms longer to process. In summary, many regions of the brain showed distinct patterns of activation for specific sentences that described different events. However, this analysis does not allow us to determine whether these effects are driven by differences in the conceptual content of the sentences or by differences in their surface orthographic, phonological, lexical and syntactic properties.

Figure 1: Analysis 1 (A) Training and test stimuli for one iteration of the analysis. (B) Decoding accuracy map, relative to chance level. Decoding accuracy is thresholded at cluster-corrected p< 0.05 and maps were smoothed at 5mm FWHM for display purposes. (C) Terms most correlated with accuracy map in Neurosynth.
In Analyses 2 and 3, we trained classifiers on one set of sentences and tested their ability to classify a new set of sentences with different lexical or syntactic forms. Areas of above-chance decoding for each analysis are shown in Figure 2B and 2D; the conjunction map in Figure 2E shows regions that were significantly above-chance in both analyses (for peak activation co-ordinates, see Supplementary Materials). Both analyses identified a similar set of brain regions. Strongest decoding was observed in left AG and pMTG in both cases, with additional clusters present in the left ATL and IFG. Decoding was also successful in the righthemisphere homologues of most of these regions. The posterior cingulate and ventromedial prefrontal cortices also showed evidence of event classification that generalised across stimuli.
Compared with Analysis 1, decoding in early visual regions was much weaker. Indeed, Figure   2F indicates that the Neurosynth terms most associated with accuracy in Analyses 2 and 3 refer to semantic processing, theory of mind and the default mode network but not to perception. Thus, when required to generalise event information to novel sentences with different structure and form, decoding in early perceptual regions was much reduced while it remained robust in supramodal semantic regions.
The final two analyses, shown in Figure 3, tested for regions whose activity discriminated between different syntactic and lexical forms, rather than between different events. In Analysis 4, we investigated which regions were able to discriminate between active and passive sentence structures. High levels of decoding were observed in early visual regions and decoding also exceeded chance levels in the intraparietal sulcus bilaterally, left posterior IFG (BA44) and the left precentral gyrus. These regions show little overlap with the eventcoding regions found in Analyses 2 and 3. Finally, Analysis 5 tested whether any regions could discriminate between sentences that used different lexical items to describe the same event.
Above-chance decoding in this analysis was only found in occipital cortices, extending into left occipitotemporal cortex.

Figure 2: Analyses 2 and 3 (A) Training and test stimuli for one iteration of Analysis 2. (B) Decoding accuracy map for Analysis 2, relative to chance level. (C) Training and test stimuli for one iteration of Analysis 3. (D) Decoding accuracy map for Analysis 3, relative to chance level. (E) Conjunction map showing regions that significantly exceeded chance level in both analyses. (F) Terms most correlated with mean accuracy map in Neurosynth.
Decoding accuracy is thresholded at clustercorrected p< 0.05 and maps were smoothed at 5mm FWHM for display purposes.

Figure 3: Analyses 4 and 5 (A) Training and test stimuli for one iteration of Analysis 4. (B) Decoding accuracy map for Analysis 4, relative to chance level. (C) Training and test stimuli for one iteration of Analysis 5. (D) Decoding accuracy map for Analysis 5, relative to chance level. (E) Terms most correlated with Analysis 4 accuracy map in Neurosynth. (F) Terms most correlated with Analysis 5 accuracy map in Neurosynth.
Decoding accuracy is thresholded at cluster-corrected p< 0.05 and maps were smoothed at 5mm FWHM for display purposes. Figure 4B. When the classifiers were trained to discriminate between events (Analyses 1 to 3), all ROIs showed better performance than expected by chance (with the exception of vATL, which was not above chance in Analysis 1). In contrast, no regions performed better than chance when decoding syntactic structure of the sentences, independent of their content (Analysis 4). In

Regions of interest: Decoding performance in anatomical ROIs is shown in
Analysis 5, which tested ability to decode the different lexical forms used to describe an event, only pMTG showed classification significantly better than chance (though power to detect decoding effects was lower in this analysis because fewer trials were used in each iteration of training). Thus the general pattern is that regions within the semantic network encode the conceptual content of sentences but not their surface characteristics.
For each ROI, we directly compared decoding accuracy between analyses on a pairwise basis. For each pairwise comparison, we used the Stelzer et al. (2013) permutation method to generate a distribution of differences under the null hypothesis, and the position of the observed difference in this distribution was used to determine its p-value. The results of these pairwise comparisons were FDR-corrected for multiple comparisons and significant differences are shown in Figure 4C. All regions except IFG showed some significant differences between analyses. In three regions -pMTG, aSTG and AG -the syntax classifier (A4) performed more poorly than one or more event-based classifiers (A1-A3). This is consistent with the findings of the searchlight analyses, which showed that patterns in the canonical semantic network were able to decode event identity but not syntactic structure.
There were also differences between the event-based classifiers. Classification in pMTG was more successful in Analysis 1, where trained and tested sentences were the same, compared with Analyses 2 and 3, where generalisation to novel sentences was required. Interestingly, the reverse was true for vATL and AG: these regions showed better classification when required to generalise to new sentence forms.

Figure 4: Region of interest analyses (A) Location of anatomical regions of interest. (B)
Decoding accuracy for each ROI in each analysis. * indicates accuracy significantly greater than chance (onetailed p < 0.05, corrected for multiple comparisons using the false discovery rate approach). (C) Results of pairwise comparison of analyses in each ROI. Circles indicate cases where the difference between analyses exceeded that expected by chance (two-tailed p < 0.05, corrected for multiple comparisons using the false discovery rate approach). A1 = Analysis 1 etc.

Discussion
Previous studies have found that a network of left-lateralised semantic processing regions represent word and object concepts in a supramodal fashion, independently of stimulus form. Here we investigated whether such stimulus independence is also a feature of the neural coding of event semantics. We trained a classifier to discriminate between sentences describing four distinct events and tested its ability to decode the same events from new sentences with different syntactic structure or lexical constituents. Neural patterns in a network of semantic processing regions, predominately in the left hemisphere, showed event coding that generalised over both syntactic and lexical changes to the stimuli. In contrast, early visual regions were able to decode specific event sentences but could not generalise to new sentences. Decoding of syntactic structure and lexical form, independent of event content, was limited and occurred primarily outside the canonical semantic network.
Our results indicate that a distributed network of left-dominant regions code information about the conceptual significance of event sentences in a way that is robust to changes in their surface forms.
Searchlight analyses revealed strongest evidence for stimulus-independent event representations in left pMTG and AG. In addition to being considered core parts of the semantic network, both of these regions have been specifically implicated in processing of event knowledge. pMTG shows increased engagement for processing verbs relative to nouns and is also involved in the neural representation of tools and actions (Bedny et al., 2008;Caspers et al., 2010;Peelen et al., 2012;Ishibashi et al., 2016). AG, in addition, has been implicated in the processing of meaningful stimuli that are extended in time, such as sentences, narratives and action sequences (Humphries et al., 2006;Lerner et al., 2011;Humphreys and Lambon Ralph, 2017;Branzi et al., 2020). This has led to claims that this region is essential for combinatorial semantic processes, in which individual word meanings are integrated into a coherent global representation (Humphries et al., 2006;Bemis and Pylkkänen, 2012;Price et al., 2015). Beyond these two core event processing regions, however, a more extended network of areas were able to decode events in a cross-sentence fashion, including portions of the ATLs, left IFG and parts of the posterior cingulate and ventromedial prefrontal cortex.
All of these areas have been identified as belonging to the brain's core semantic network (Binder et al., 2009;Binder and Desai, 2011;Lambon Ralph et al., 2017). In addition to overlapping considerably with the default mode network, these areas tend to be the most physically distant and functionally distinct from regions serving primary sensory and motor processing (Margulies et al., 2016). This suggests that they are particularly well-suited to serving supramodal and internally-directed cognitive processes that abstract away from perceptual input. Our results support this general view and indicate that activation patterns throughout this network can distinguish between different conceptual states in a stimulusindependent manner.
The network of regions observed here converges with other studies applying multivariate decoding techniques to sentences (Frankland and Greene, 2015;Anderson et al., 2017;Pereira et al., 2018) and to language at the discourse level (Wehbe et al., 2014;Huth et al., 2016). These also report successful decoding of sentence meaning in left inferior parietal, lateral temporal and inferior prefrontal voxels. Like previous studies, we also found that the network coding meanings was left-lateralised, though there was evidence for above-chance decoding in right-hemisphere homologues as well. However, the present study diverges from some previous work in identifying in left ventral ATL (inferior temporal and fusiform gyri) as an additional area involved in the coding of sentence meaning. This area often suffers from susceptibility artefacts and associated signal dropout, which our use of multi-echo fMRI may have alleviated (Visser et al., 2010;Kundu et al., 2017).
The ventral ATL is a key site for semantic representation and, along with the adjacent perirhinal cortex, is thought to be involved in representing conceptual relationships amongst objects and words (Rogers et al., 2004;Clarke and Tyler, 2015;Lambon Ralph et al., 2017;Bruffaerts et al., 2019). Accounts have emphasised the role of this region in understanding single words and objects but have not proposed a specific role in the representation of events.
Indeed, some researchers propose a division of labour within the semantic system whereby the ATL codes taxonomic semantic structure while the temporoparietal cortex is represents thematic associations between concepts (Schwartz et al., 2011;Mirman et al., 2017). On this view, ATL should primarily be involved in coding the properties of individual objects or words while the temporoparietal cortex plays a greater role in coding the semantics of events, due to its sensitivity to thematic information. Our data are agnostic on this issue. We found that patterns in ventral ATL do discriminate between different event descriptions. However, the events we used contained different agents and patients and it is possible that ventral ATL voxels coded the properties of these constituents (e.g., discriminating between cows and lorries), rather an integrated event-level representation. Future studies could disentangle these possibilities by using sentences that recombine the same constituents into different event structures (e.g., Frankland and Greene, 2015;Frankland and Greene, 2020).
A functional division within the semantic system has also been proposed between representational regions that code meaning and control regions that regulate and shape semantic activation based on current goals (Badre and Wagner, 2007;Jefferies, 2013;Hoffman et al., 2018). Following this distinction, the Controlled Semantic Cognition theory proposes that IFG and pMTG are involved in semantic control processes rather than representation . In the present study, we observed robust decoding of events in both of these areas, as have other semantic decoding studies at the word and sentence levels (Devereux et al., 2013;Fairhall and Caramazza, 2013;Anderson et al., 2017;Pereira et al., 2018). If these regions are important for regulating semantic activation rather than storing conceptual representations, why can their neural patterns reliably distinguish between different event concepts? There are a number of possible answers to this. Different events may vary, for example, in the demands they place on control processes and this systematic variation in engagement might be sufficient to allow classification. Alternatively, different events might require different control processes with distinct neural signatures. Better understanding of this issue will require more precise formulation of the mechanisms and computations involved in semantic control. Many computational models of semantic representation exist and these make specific predictions about the representational structures involved (e.g., Rogers et al., 2004;O'Connor et al., 2009;Taylor et al., 2012;Schapiro et al., 2013). Conversely, there has been little formal modelling of semantic control processes (though for recent exceptions, see Hoffman et al., 2018;Jackson et al., 2019). The lack of established theoretical models makes it difficult to make precise predictions about how neural patterns in control regions should vary and how these areas might be distinguished from those that represent knowledge. This remains a key challenge for future work to address.
While we found that brain regions throughout the canonical semantic network were able to decode event semantics independent of syntactic structure, coding of syntactic information was much more limited. None of our semantic ROIs were able to discriminate between active and passive sentences in an event-independent fashion; instead, this form of decoding was mostly limited to occipital cortex and superior frontal and parietal regions.
Some of these effects may reflect differences in non-semantic processing demands to the two sentence types: passive sentences are necessarily longer than their active equivalents and thus are likely to require greater visual analysis, additional eye movements and greater working memory recruitment. Significant decoding was also observed in a posterior portion of left IFG (pars opericularis; BA44), somewhat posterior to the peak decoding in the eventdiscrimination analyses. This is consistent with the idea of graded specialisation in left IFG, with semantic processing engaging more anterior regions and syntactic processing more posterior areas (Friederici, 2012;Rodd et al., 2015). In general, however, our results support the view that few regions in the language network exclusively code syntactic information, independently of the semantic content of language (Fedorenko et al., 2012;Blank et al., 2016).
Finally, we tested for areas that discriminated between different lexical forms used to describe the same events. The searchlight analysis found that this decoding was restricted to early visual and left occipitotemporal regions, which is likely to be a consequence of orthographic rather than semantic processing. In the ROI analyses, however, pMTG showed above-chance discrimination. This effect might indicate differences in lexical access processes supported by posterior temporal regions (Lau et al., 2008;Lewis and Poeppel, 2014;Davis, 2016). Alternatively, the distinct neural patterns observed here may be a consequence of subtle semantic distinctions between the different lexical forms. Although we designed the materials such that lexical substitutes were as conceptually similar as possible, there were inevitably some differences between them (cows and bulls are not identical, nor are fences and gates, and so on). With a larger sample, we may have been able to decode these subtle semantic distinctions more widely within the semantic network. It is clear from the picture across analyses, however, that activation patterns in the semantic network were primarily driven by the conceptual distinctions between the different events and not by the lexical or syntactic forms used to describe them. Thus, to conclude, the present study indicates that a distributed network of language and semantic regions code event-level semantic information in a manner that generalises across lexical and syntactic variation in the sentences used to describe them. This suggests that the core function of this network is to abstract away from perceptual input and represent the underlying conceptual significance of our experiences.