Unaware processing of words activates experience-derived information in conceptual-semantic brain networks Open Access

Our brain constantly receives sensory input from around and within us and reacts to it, also in anticipatory mode, by activating conceptual-semantic representations derived from previous experience with similar percepts. For example, visual perception of manipulable objects (MO) engages not only bottom-up visual encoding but also top-down conceptual-semantic decoding based on stored MO concepts (Noppeney et al., 2006). These are considered to encompass high-level semantic (e.g., categorical) as well as visual, motor, and functional information derived from our previous experience with MO exemplars (Binder & Desai, 2011; Lambon Ralph et al., 2017). The representation, storage, and retrieval of visual properties rely on bilateral ventral occipito-temporal areas, while motor and functional properties draw on portions of the left-hemispheric parietal and premotor cortices that support the motor programming of actions related to the object function (Chen et al., 2017; Ishibashi et al., 2016). This MO-specific visuomotor network (MOVN) interacts with the general semantic brain network (Binder et al., 2009) to form inferences about the incoming visual input that allow us to ultimately identify the object and possibly to plan a motor response (Binder & Desai, 2011; Lambon Ralph et al., 2017).

Notably, MO conceptual-semantic representations seem to be activated also under unaware visual perception conditions instantiated by Continuous Flash Suppression (CFS; Fahrenfort et al., 2017; Ludwig et al., 2016; Tsuchiya & Koch, 2005). By applying CFS in combination with fMRI, we showed the recruitment and effective connectivity of the MOVN both under aware and unaware processing of MO pictures (Ghio et al., 2022; Tettamanti et al., 2017). A possible concern, however, is that the MOVN was recruited due to the visual processing of irrelevant perceptual properties such as the prototypical elongated shape of MO, rather than to the retrieval of semantic features such as manipulability and affordance. The first aim of the current study was to address this concern by evaluating the subliminally-induced activation of the MOVN in processing MO stimuli that lack a visual depiction of the objects, as is the case for words. While both pictorial and verbal stimuli trigger conceptual-semantic processing, the latter are arbitrarily associated with their referents through learning and do not carry any object-related visuomotor information. In the present study, we combined fMRI and CFS to achieve a high anatomical localization specificity, thus complementing previous evidence obtained by means of electrophysiological measurements, which showed that the unconscious activation of motor features occurs already in the early stages of word processing (Trumpp et al., 2013, 2014). Evidence of a recruitment of the MOVN under subliminal processing of MO words would speak for the hypothesis that conceptual-semantic processing activating experience-derived MO information occurs automatically.

The second aim was to examine the generalizability of the hypothesis that experience-derived information constituting conceptual-semantic representations is automatically recruited in subliminal perception, by extending our investigation to emotions. Conceptual representations of discrete emotions (EM; e.g., happiness, anger) are characterized by valence, arousal, as well as interoceptive and somatic information derived from our previous experience of EM instances (Borghi et al., 2018; Conca et al., 2021; Ghio & Tettamanti, 2016). This information is supposed to be represented and stored in the same limbic network that supports the experience of EM and comprises, bilaterally, the amygdala, insula, anterior cingulate, and orbitofrontal cortices, putamen, and midbrain (Lindquist et al., 2012). This EM-specific limbic network (EMLN) has been shown to be activated when we visually perceive an instance of an EM in, for example, a face picture, even presented below the perceptual threshold (Dahlén et al., 2022; Jiang & He, 2006; Lindquist et al., 2012). As for MO, however, to overcome the potential criticism that subliminal activation is due to visual properties of the pictures, we used subliminally presented EM words. Previous behavioral and event-related potential studies using CFS proved that EM word meaning is also accessed under subliminal conditions (Lei et al., 2017; Sklar et al., 2012; Vinson et al., 2011; Y. H. Yang & Yeh, 2011), without providing any insight, however, into the automatic recruitment of the EMLN.

In the present fMRI study, we thus tested the aforementioned hypothesis and its generalization by assessing whether the subliminal processing of MO and EM words specifically engages, respectively, the MOVN and the EMLN (see Fig. 1A). Because valence is a dominant feature for emotion conceptual-semantic representations (Conca et al., 2021), with respect to which EM words are consistently characterized across languages (Jackson et al., 2019), EM words were parametrically distributed along the negative-to-positive valence dimension, including an equal number of negative (EMneg) and positive (EMpos) words. In addition, neutral words NEU (e.g., thought) served as a control condition to test the effect of valence within the EMLN.

To instantiate subliminal word processing, we applied a CFS paradigm (Fig. 1B). After determining the perceptual threshold for each participant, the word stimuli (Word Type: MO, EMneg, EMpos, NEU) were presented at five incremental visual contrast levels (Word Contrast: two levels below, one at participant’s perceptual threshold, and two above), with void stimuli as perceptual control. Our CFS paradigm also required to indicate subjective ratings along a perceptual awareness scale (PAS; Ramsøy & Overgaard, 2004), which allowed us to determine the presence or absence of subjective perceptual awareness of the lexical or sublexical stimulus form, independently of word contrast.

We first applied univariate analyses (Fig. 1C) to assess subliminal fMRI activation (1) as a function of Word Contrast (reflecting the subliminal vs. supraliminal conditions defined based on the participant’s perceptual threshold), and (2) as a function of PAS ratings (reflecting the subjective perceptual awareness of the lexical or sublexical stimulus form). In both analyses, we tested the hypothesis that the MOVN and the EMLN specifically respond to, respectively, MO and EM words following a “subliminal” activation amplitude profile, which captures an approximately stable activation either (1) across the five incremental contrast levels or (2) with and without subjective awareness. In addition to the subliminal profile, which would speak for unaware processing, we also tested other, possibly concomitant, response profiles, namely a “linear” one, modeling activation increase with increasing word contrast (only in 1) and a “stepwise” profile, that is, drop of activation either (1) for word contrast below perceptual threshold or (2) without subjective awareness. Moreover, for EM we tested the effects of valence by distinguishing EMneg, EMpos and NEU.

Secondly, we applied multivariate pattern analyses (Fig. 1D), again (1) as a function of Word Contrast and (2) as a function of PAS ratings. We hypothesized that the involvement of MOVN and EMLN for the representation and subliminal processing of, respectively, MO and EM words would be reflected by the successful decoding of the word types (1) not only above but also below perceptual threshold and (2) not only with but also without subjective perceptual awareness.

Twenty-five Italian native speakers voluntarily participated in the study. Two participants were excluded due to the finding of brain structural anomalies. Two other participants did not comply with the CFS task and were discarded from the analyses. The sample of the remaining 21 subjects comprised 7 male and 14 female participants (mean age = 22.31 years, SD = 4.01; mean years of education = 15.76, SD = 3.25).

The sample size was determined by means of a statistical power analysis, based on our previous fMRI study investigating the processing of MO pictures under unaware perceptual conditions instantiated through CFS (Tettamanti et al., 2017). We used the fmripower toolbox (Mumford & Nichols, 2008) to estimate the required sample size with respect to the two effects of interest of our previous fMRI study, namely stimulus type and its interaction with subjective awareness, which paralleled those in the current study. The power estimation was restricted to the MOVN brain regions of interest. A one-sample t-test (α = .05) indicated that a sample size of 21 participants yields at least 80% power to detect a significant effect of stimulus type (assumed Z score = 3.59). In turn, a paired t-test (α = .05) indicated that a sample size of 16 participants yields at least 80% power to detect a significant interaction between stimulus type and subliminal activation profile (assumed Z score = 2.86). Based on these estimates, we set the target sample size for the current study to 25 participants. We motivated this decision from a conservative perspective (Mumford, 2012), taking into account the similarities, but also the number of differences between the experimental designs of our previous versus current study. While the focus on stimuli referring to MO and on regions of the MOVN (Fig. 1A) and trial structure (Fig. 1B) were essentially identical in the two studies, differences concerned the additional focus on EM-specific responses in the EMLN in the present study, stimulus format (i.e., pictures in the previous vs. words in the current study), the number of stimuli and of fMRI runs (see Section 2.3.4). As noted above, 4 participants had to be excluded from the final sample (n = 21), which remained nevertheless the same size as our largest power estimate in the two power analysis tests just described.

All participants reported to be right-handed, which was in line with the Edinburgh inventory assessment (Oldfield, 1971; mean score = 0.80, SD = 0.16, with 20 participants scoring in the range from 0.5 to 1 and thus classified as right-handed and one participant scoring 0.4 and classified as ambidextrous). The Miles test (Miles, 1929) was administered for evaluating eye dominance for the CFS task: 8 participants were left- and 13 right eye dominant. All participants reported neither sensory deficits nor developmental, neurological, or psychiatric diagnoses. No participants had reading deficits according to their performance with real word (mean score = 100%, SD = 0%) and pseudoword (mean score = 99%, SD = 1%) reading tests included in the Italian battery for assessing aphasic deficits (Miceli et al., 1994). All participants gave written consent and received a reimbursement of 30 euro for their participation in the study. The study was approved by the ethics committee of the University of Trento, Italy.

As experimental word stimuli we used 75 nouns referring to MO and 75 to EM. EM nouns were parametrically distributed along the negative-to-positive valence dimension, including 25 EMneg and 25 EMpos nouns, with additional 25 NEU nouns as control condition for testing the valence effects. EMneg and EMpos were defined as referring to “a mental state that could be felt” (e.g., anger or joy; Jackson et al., 2019), also known as emotion-label nouns (Conca et al., 2021). Emotion-laden nouns (e.g., war or party) were avoided. NEU were defined as referring to “neutral mental states” (Jackson et al., 2019) that “can not be felt” (e.g., thought). In line with these definitions, we checked that in the lexical database Multiwordnet (Pianta et al., 2002) all EMneg, EMpos, and NEU nouns are classified as psychological features or factotum (semantic domain level), but only EMneg and EMpos are defined as emotions or traits/attitudes (affective level). The complete list of stimuli is available as Supplementary Table 1A.

To validate the variability of EM words along the valence dimension, all nouns were rated on a bipolar Likert scale (-3 = negative, 0 = neutral, +3 = positive) by an independent group of 16 volunteers (10 female and 5 male participants, 1 did not specify the sex; mean age = 29.25 years, SD = 5.48). Results showed that EMneg, EMpos, and NEU nouns were clearly distributed along the negative-to-positive valence dimension, and significantly differed from MO (Table 1A for summary statistics and Supplementary Table 1B for single word rating scores). Ratings for arousal (1 = calm or relaxation, 7 = high excitement or agitation) were also collected, as valence and arousal are considered to be independent dimensions of the semantic space (Osgood & Suci, 1955). Results showed that EMpos, NEU, and MO did not differ in arousal, although all were lower than EMneg (Table 1A). Finally, ratings on a unipolar Likert scale for imageability (1 = barely imaginable, 7 = easily imaginable), which is considered as a proxy for concreteness (Hoffman, 2016), showed that EM and NEU were less imageable than MO words. Experimental stimuli were rated for valence, arousal, and imageability and by all (but one) fMRI participants (on average 8.55 days, SD = 9.97, after the MRI session). The results showed a high correlation between the ratings of the fMRI participants and those of the participants who only did the validation word rating (Spearman’s correlation coefficient: valence: rho = 0.971; arousal: rho = 0.937; imageability: rho = 0.994, all p values < .001).

Experimental stimuli were balanced for relevant sub-lexical and lexical variables (Table 1B for detailed descriptive and inferential statistics). Lexical frequency, which has been demonstrated to influence speed in word recognition (Brysbaert et al., 2016; but see Heyman & Moors, 2014), was extracted from the SUBTLEX-IT database (https://osf.io/zg7sc) and matched across the four types of words. Moreover, since sub-lexical processing has also been found to affect word recognition (Westbury & Buchanan, 2002), MO, EMneg, EMpos, and NEU were matched for the frequency of the first, last, and average medial bigrams of each word, plus the overall average bigram frequency. Bigram frequencies were extracted from the PhonItalia database (Goslin et al., 2014). Stimuli were also matched for their number of orthographic neighbors (again extracted from the PhonItalia database, Goslin et al., 2014), for it has been demonstrated that a large orthographic neighborhood size makes words easier to read (Duñabeitia & Vidal-Abarca, 2008). Finally, stimuli were balanced for their number of syllables, number of letters, and visual length calculated in millimeters using the software GIMP 2.10.4 (www.gimp.org).

We used MATLAB R2017b to create 80 mask images with cyan background to make the mask visible only to the dominant eye when wearing the red-cyan anaglyph goggles (see below). Over the cyan background, we displayed black patterns, which were similar in shape (e.g., closed unfilled geometrical shapes such as ellipses and rectangles, and open u-shaped figures with different orientations) and perceptual features (e.g., line width, dimension) to the Latin alphabet letters of the word stimuli (Axelrod et al., 2015; Kido & Makioka, 2014). Then, we overlaid a black frame and a black fixation cross at the center of every mask, using GIMP 2.10.4 (www.gimp.org).

Participants first performed a training session to familiarize with the CFS task outside the MRI scanner room. Then, they entered into the MRI scanner and performed a pre-fMRI behavioral session of the CFS task to assess their perceptual threshold in the same contextual and perceptual conditions as for the fMRI CFS task. Finally, they performed the experimental CFS task session, while their brain activity was measured with fMRI. In all these three sessions, participants wore MRI-compatible red-cyan anaglyph goggles, with the cyan lens placed above the dominant eye. The CFS trial structure was identical across the three sessions, as well as the use of Psychtoolbox-3 (v3.0.14, running in MATLAB 2017b) for stimulus delivery and behavioral response collection.

Each CFS trial began with the 200-ms presentation of a fixation cross on a red background overlaid by an empty cyan image with 50% transparency (alpha blending). Then, a single word was presented in black on a red background to make it visible only to the non-dominant eye when wearing the red-cyan anaglyph goggles. Each word was written in uppercase (font: Helvetica, font dimension: 80) and displayed in the center of the screen at a trial-specific (see below) visual contrast level, for a duration of 1600 ms, during which a set of 16 different masks presented at 10 Hz was overlaid on the word with 50% transparency (Fig. 1B). For every trial, the 16 masks were randomly extracted without replacement from the whole pool of 80 mask images and then presented in a randomized order. To reduce after image effects (Hong & Blake, 2009), which can occur at the offset of the last mask, we displayed for 1000 ms an image with multiple white capital X letters on a black background. After a variable within-trial delay (pseudo-randomized duration of either 1425, 1925, or 2425 ms in 4:2:1 proportion), participants were shown the display “1 2 3 4” (duration 1000 ms), which represented the four levels of the PAS (Ramsøy & Overgaard, 2004), with definitions which we adapted to match the lexical nature of our task: 1 = no experience; 2 = brief glimpse, without recognition of any letter; 3 = almost clear perception, with recognition of one or more letters but not of the entire word; and 4 = absolutely clear perception, with recognition of the entire word. Participants were asked to press one of the four buttons of a response pad according to their subjective level of visual perception of the word. Importantly, PAS scores 1 and 2 signaled the absence of orthographic or lexical recognition, meaning that the participants lacked subjective awareness of the lexical or sublexical form of the word stimulus. Throughout the manuscript, we therefore refer to PAS score levels 1 and 2 as corresponding, in short, to the absence of subjective awareness, and group these two levels together in the analysis as a function of PAS (see Fig. 1C and Section 2.5.2.2). Because we wanted to measure spontaneous activation devoid of explicit semantic judgments, we did not require our participants to overtly engage in mental imagery or perform any task requiring deep semantic processing. The variable within-trial delay was introduced in order to reduce the temporal correlation between the stimulus visual presentation and the motor response to the PAS scale, thus preventing possible biases in interpreting blood oxygen level dependent (BOLD) signal fluctuations. At the end of the trial, a delay of variable duration (short: 2925 or 3425 ms, medium: 4075 or 4575 ms, long: variable from 5075 to 10575 ms, in 4:2:1 proportion) was included before the start of the next trial. Within-trial and between-trial delays were determined by OPTseq2 (https://surfer.nmr.mgh.harvard.edu/optseq) to maximize the sensitivity of the hemodynamic signal in the fMRI CFS event-related design (Dale, 1999). Trial timings were kept identical also for the training and the pre-fMRI behavioral sessions.

Before entering the MRI scanner, participants were trained to perform the CFS task on a laptop computer. After performing 6 practice trials, participants started the training session, which comprised 105 consecutive trials. For 100 trials, we used nouns referring to jobs and careers (e.g., biologist), which were matched with the experimental stimuli with respect to lexical frequency, number of syllables, and number of letters (all p > .072; Supplementary Table 2A, B). These nouns were presented at a specific contrast level according to the following procedure (similarly to Tettamanti et al., 2017). For each participant, the 100 nouns were randomly ordered in a list and then sequentially drawn without replacement from such list and assigned, in batches of 5, to 1 of 20 transparency levels (alpha blending value ranging from 8 to 255 = fully opaque, by steps of 13 units). In addition, for 5 trials we applied a transparency level with alpha = 0, which resulted in 5 empty trials (i.e., trials devoid of words, only including the red background and the flickering cyan masks). The presentation order of the 105 trials was pseudo-randomized across participants.

The perceptual threshold of each participant was assessed before the fMRI data acquisition session. Participants were positioned in the MRI scanner and performed the CFS task on 105 consecutive trials, including 5 empty trials and 100 trials, during which 100 nouns referring to living/natural items (e.g., elephant, sunflower, storm, forest) were presented at 20 different alpha transparency levels (defined as in the training session described above). This set of nouns was matched with the experimental stimuli with respect to lexical frequency, number of syllables, number of letters, bigram frequency (initial, final, medial, average), and number of orthographic neighbors (all p > .103; Supplementary Table 3A, B). The same presentation procedure as in the training session was applied.

To estimate the subjective perceptual threshold, we adopted the same procedure as in Tettamanti et al. (2017). First, for each participant we plotted the number of responses with a PAS score equal to 3 or 4 (indicating high to maximal subjective perceptual awareness) for each of the 20 transparency levels plus the empty trials. By using nonlinear least-squares estimation in MATLAB 2017b, we fitted these values with a psychometric sigmoid function f = L + (R – L) / {1 + exp [– (x – i) / W]}, where x = transparency level, L = left horizontal asymptote (initial value = 0); R = right horizontal asymptote (initial value = 5), i = point of inflection (initial value = 2.5), and W = width of the rising interval (initial value = 1). We, thus, defined the subjective perceptual threshold (PT) as the first transparency level above the estimated point of inflection. Based on the subjective perceptual threshold, five transparency levels were individually tailored to be used in the fMRI session (and henceforth defined as Word Contrast levels). Specifically, two tailored contrast levels were below PT: one defined as PT -25% (PT-2) and the other as PT -10% (PT-1). One contrast level was at PT. Two tailored contrast levels were above PT: one defined as PT +10% (PT+1), while the other was set for all participants at the maximum available contrast (Full, i.e., corresponding to alpha = 255). In addition, for all participants, we included a null contrast control condition (Empty), consisting of a red square devoid of any words. For participants with a subjective PT equal or lower than the transparency alpha level value of 69, the contrast levels below PT had to be tailored to a narrower range. Specifically, for participants (n = 9) with a subjective PT between alpha = 36 and 69, we set PT-2 to alpha = 5. For participants (n = 4) with a subjective PT lower than alpha = 36, we set PT-2 to alpha = 5 and PT-1 to alpha = 10.

The fMRI experiment consisted in a within-subject event-related experimental design, where Word Type (MO, EMneg, EMpos, NEU) and Word Contrast (PT-2, PT-1, PT, PT+1, Full) were factorially manipulated. All participants completed 6 runs. Every run comprised 55 trials: 5 empty, 25 MO, and 25 EM trials. All the 25 EM words in every run belonged to the same valence category (either EMneg, or EMpos, or NEU), and the EM category assignment to the runs was counterbalanced for every participant on the basis of a Latin square, such that all words of the three categories occurred once in the first three runs before being extracted again in the second three runs. Each of the 75 MO words were presented once in the first three runs and a second time in the last three runs. Thus, every word was presented twice throughout the experiment. Within every run, 5 MO and 5 EM were presented at one of the 5 word contrast levels. The presentation order of the different Word Type × Word Contrast combinations was unique for every participant and run, and determined by OPTseq2. The assignment of the specific word stimuli to the Word Type × Word Contrast combinations was pseudo-randomized for every participant and run, in order to minimize the number of stimuli presented at the same contrast in their first occurrence (during one of the first three runs) versus their second occurrence (during one of the last three runs) for a given participant.

MRI scans were acquired with a 3 Tesla Siemens Magnetom Prisma whole-body MRI scanner using a 64-channels head coil. During the PT estimation session, we obtained a T1-weighted MPRAGE anatomical scan (288 slices, TR = 3280 ms, TE = 2.01 ms, flip angle = 7°; FOV = 252 × 252 mm, 0.7 mm isotropic voxel size). During the experimental CFS task session, whole-brain functional images were acquired with a T2*-weighted gradient-echo, EPI pulse sequence, using BOLD contrast (TR = 1000 ms, TE = 28 ms, flip angle = 59°, multi-band acceleration factor = 5). Each functional image comprised 65 transversal slices (2 mm isotropic voxel size), acquired in interleaved mode (FOV = 200 x 200 mm). Each participant underwent 6 consecutive functional scanning sessions, each comprising 520 scans and lasting 8 min and 50 s.

The stimuli were projected on a computer monitor located in the back of the scanner and visible to the participants through a mirror placed on the head coil above their eyes. An MRI-compatible fiber-optic response box with four buttons allowed the participants to give PAS score responses using their left hand. The left hand was chosen to avoid contamination of the BOLD signal acquired in left hemisphere where language processing is lateralized, even though the overlapping between motor responses and word presentation was anyway minimized thank to the within-trial delay.

To evaluate the effects of Word Type and Word Contrast on the PAS scores during fMRI, we ran a series of Generalized Linear Mixed Models using R 4.2.2 (R Core Team, 2022). The first simplest model comprised PAS scores as dependent variable, the six runs as fixed-effect confounds, and the participants as random-effect factor. Word stimuli were also modeled as random-effect factor, although we note that, since their variance equaled zero (SD = 0), their inclusion or omission from the model yields identical results (Pasch et al., 2013). We then specified a series of models with increasing complexity by adding Word Type (MO, EMneg, EMpos, NEU), Word Contrast (PT-2, PT-1, PT, PT+1, Full), and the Word Type by Word Contrast interaction as fixed-effect factors. All models were fit to a Poisson distribution, with a log-link function. To evaluate whether each additional predictor significantly increased the model fit, we ran a χ² log-likelihood ratio test (declared significance α level: .05) between each hierarchically more complex model and the simpler one.

For MRI data preprocessing and univariate statistical analyses, we used SPM12 v7771 (www.fil.ion.ucl.ac.uk/spm). Structural MRI images were segmented and registered to the Montreal Neurological Institute (MNI) standard space. Functional images were corrected for geometrical distortions, spatially realigned, corrected for slice timing, and normalized to the MNI space by using the segment procedure with the subject-specific segmented structural images as customized segmentation priors. Functional images were spatially smoothed with a 4 mm FWHM Gaussian kernel.

By applying a two-stage random-effects statistical approach, we performed univariate statistical analyses of fMRI activation (1) as a function of Word Contrast and (2) as a function of PAS rating. For all the reported univariate analyses, the significance threshold was declared at peak level p = .05, using a small volume familywise error (FWE) type correction for multiple comparisons. For small volume correction, we defined a single mask pooling together the regions of interest constituting the MOVN and the EMLN (Fig. 1A), respectively, each region consisting of a spherical small volume of 6 mm radius centered on the target coordinate (x, y, z cartesian values in mm referring to the MNI space). The coordinates for the MOVN were taken from the meta-analysis by Ishibashi et al. (2016), and included the left ventral premotor cortex (LVPM; x = -50.3, y = 6.3, z = 30.7), the left inferior (LIPL; x = -45.7, y = -31.7, z = 44.8) and superior parietal lobules (LSPL; x = -23.4, y = -61.1, z = 60.9), and the left middle temporal gyrus (LMTG; x = -47.8, y = -65.1, z = -1.8). For EMLN, the coordinates were taken from Lindquist et al. (2012), and included the left (LAMY; x = -20, y = -4, z = -16) and right amygdala (RAMY; x = 22, y = -4, z = -16), the left (LOFC; x = -40, y = 26, z = -6) and right orbito-frontal cortex (ROFC; x = 42, y = 24, z = -4), the left anterior insula (LINS; x = -38, y = 8, z = -10), and the sub-/pre-genual anterior cingulate cortex (sACC; x = 0, y = 38, z = 6).

At the first level, for each participant we specified a general linear model, which comprised six separate runs, each including 11 regressors of interest: Empty, MOPT-2, MOPT-1, MOPT, MOPT+1, MOFull, EMPT-2, EMPT-1, EMPT, EMPT+1, EMFull, with EM corresponding to either EMneg, EMpos, or NEU depending on the run. Evoked responses were aligned to the onset of each trial (appearance of the fixation cross), and their duration was fixed to 2800 ms (comprising 200 ms for the presentation of the fixation cross, 1600 ms for the presentation of the masked stimuli, and 1000 ms for the presentation of the anti-after effect image). Experimental confounds, namely PAS responses (for each trial, onset aligned to the display of the PAS scale and duration equal to the interval between this time point and the participant’s response), task instructions, and head movement parameters yielded by image realignment, were modeled as separate regressors. If there were trials in which participants did not answer to the PAS scale (misses) or answered before the PAS question was displayed, these were modeled in the same confound regressor with task instructions. For all experimental conditions, we modeled the evoked responses as canonical hemodynamic response functions. The time series were high-pass filtered at 128 s and pre-whitened by means of an autoregressive model AR(1). Global normalization was performed to account for possible differences in signal due to the presentation of EMneg, EMpos, and NEU stimuli in different runs. Statistical analyses were restricted by using an explicit mask, including only the voxels with gray matter tissue probability > .1, as determined based on the segmented structural images of each participant.

We defined two different random-effects, full factorial designs. The first one was aimed at testing the main effect of Word Type. Accordingly, within the estimated first-level general linear model we defined a series of 20 condition-specific contrasts with the Empty baseline subtracted (e.g., MOPT-2 – Empty). Namely for each run of the general linear model containing the condition of interest (e.g., MOPT-2), we entered a weight of +1 for the regressor representing that condition, and a weight of -1 for the Empty baseline; a weight of 0 was entered for all the other regressors in the run and, in the case of contrasts involving the word types EMneg, EMpos, or NEU, also for all regressors of the four runs not containing the condition of interest. At the second level, we entered the resulting 20 first-level contrasts into a model including Word Type (4 levels: MO, EMneg, EMpos, NEU) and Word Contrast (5 levels: PT-2, PT-1, PT, PT+1, Full) as within-subject factors. We assumed dependence and equal variance between the levels of both factors. Within this full factorial design, we specified a T-contrast to examine the effects of MO (MO – [EMneg+EMpos]), and a T-contrast to examine the effects of EM ([EMneg+EMpos] – MO). In addition, for EM we tested the simple effect of valence (i.e., with linear contrast weights: -1 for EMneg, 0 for NEU, and +1 for EMpos) and the quadratic effect of valence (i.e., with non-linear weights: +1 for EMneg, -2 for NEU, and +1 for EMpos).

The second random-effects, full factorial design aimed at testing whether the effects of MO and EM as a function of Word Contrast followed either a subliminal (i.e., activation both above and below perceptual threshold), a linear (i.e., gradual activation increase with contrast), or a stepwise (i.e., activation only above perceptual threshold) profile (Fig. 1C). In order to model these three different profiles, within the estimated first-level general linear model, we defined 24 condition-specific contrasts with the Empty baseline not subtracted but explicitly modeled. Namely, for each run of the general linear model containing the Word Type level (e.g., MO) to be factorially combined with a Word Contrast level (e.g., PT-2), we entered a weight of +1 representing the factorial combination of the specific condition of interest (e.g., MOPT-2); a weight of 0 was entered for all the other regressors in the run and, in the case of contrasts involving the word types EMneg, EMpos, or NEU, also for all regressors of the four runs not containing the Word Type level of interest. At the second level, we entered the resulting 24 contrasts into a model including the within-subject factors Word Type (4 levels: MO, EMneg, EMpos, NEU) and Word Contrast (6 levels: Empty, PT-2, PT-1, PT, PT+1, Full), with dependence and equal variance assumed between the levels of both factors. Then, we investigated whether the Word Contrast level modulated the [MO – (EMneg+EMpos)] and [(EMneg+EMpos) – MO] activation effects, following either a subliminal (weights for each Word Contrast level are: [Empty: -5, PT-2: +1, PT-1: +1, PT: +1, PT+1: +1, Full: +1]), a linear [Empty: -2.5, PT-2: -1.5, PT-1: -0.5, PT: +0.5, PT+1: +1.5, Full: +2.5], or a stepwise [Empty: -1, PT-2: -1, PT-1: -1; PT: +1, PT+1: +1, Full: +1] BOLD amplitude model.

To test the MO and EM effects without subjective awareness, but independently of Word Contrast, we further analyzed BOLD responses as a function of PAS ratings. Specifically, for each type of words, all trials in which the reported PAS score was either 1 or 2 (PAS{1,2}) were modeled as indicating no subjective awareness of the lexical or sublexical forms, whereas all trials in which the PAS score was either 3 or 4 (PAS{3,4}) were modeled as indicating subjective awareness. We restricted this analysis to only those subjects that produced comparable response frequencies across all PAS levels. To exclude participants who did not satisfy this criterion, we applied the same procedure as in Tettamanti et al. (2017): first, we estimated the maximum likelihood of the individual Contrast-to-PAS linear function to a Contrast-to-PAS linear model function with R. Then, we ranked participants based on the resulting Akaike information criterion. None of the participants markedly departed from a linear increase of PAS score with increasing contrast level. However, one participant was excluded from the following analyses because of an exceedingly low number of PAS{3,4} responses, even for the PT+1 and Full contrast levels (see Supplementary Fig. 1). Thus, these analyses were performed on a sub-sample of 20 participants.

At the first level, we specified a general linear model with the same parameters as defined above for the analysis as a function of Word Contrast. However, here we modeled the six runs by assigning the CFS trial events on the basis of PAS score and including the Empty baseline condition. Specifically, each run included 5 regressors of interest: Empty, MOPAS{1,2}, MOPAS{3,4}, EMPAS{1,2}, EMPAS{3,4}, with EM being either EMneg, EMpos, or NEU depending on the run. All the confound regressors were defined as in the analysis as a function of contrast (see above). Within the estimated first-level general linear model, we defined a set of 12 condition-specific contrasts. Namely, for each run of the general linear model containing the Word Type level (e.g., MO) to be factorially combined with a PAS level (e.g., PAS{1,2}), we entered a weight of +1 for the regressor representing the factorial combination of interest (e.g., MOPAS{1,2}); a weight of 0 was entered for all the other regressors in the run and, in the case of contrasts involving the word types EMneg, EMpos, or NEU, also for all regressors of the four runs not containing the Word Type level of interest.

For the second-level, random-effects full factorial design, we entered the 12 condition-specific contrasts of each participant into a model including the within-subject factors Word Type (4 levels: MO, EMneg, EMpos, NEU) and PAS (3 levels: Empty, PAS{1,2}, PAS{3,4}), with dependence and equal variance assumed between the levels of both factors. In this full factorial design, we investigated whether the absence or presence of subjective perceptual awareness (as reflected by the PAS score) modulated the activation effects of [MO – (EMneg+EMpos)] and [(EMneg+EMpos) – MO], following a subliminal (weights for each PAS level are: [Empty: -2, PAS{1,2}: +1, PAS{3,4}: +1]) or a stepwise [Empty: -1, PAS{1,2}: -1, PAS{3,4}; +2] BOLD amplitude model.

MVPA was carried out using the PyMVPA 2.6.5 toolbox (www.pymvpa.org) running under Python 2.7.18. We first re-estimated the SPM12 univariate first-level analyses on preprocessed but spatially unsmoothed fMRI data. We then entered the resulting spmT map images into the MVPA (Misaki et al., 2010). All the included spmT maps incorporated the Empty baseline subtraction (e.g., [MOPT-2 – Empty]). We applied no Z-scoring or averaging within the PyMVPA toolbox. We restricted the analysis to voxels with gray matter tissue probability > .1, which were comprised within 12-mm radius spherical regions of interest in the MOVN and EMLN (Fig. 1A), yielding a single voxel mask of 7550 voxels in total.

We trained a Gaussian Naive Bayes classifier on spmT maps to distinguish the levels of the Word Type factor (MO, EMneg, EMpos) either between all three levels (3 x 3 classification problems) or between pairwise combinations of these levels (2 x 2 classification problems), based on a leave-one-subject-out cross-validation procedure, analogous to a second-level, random effects analysis (Ghio et al., 2016). More specifically, we performed a searchlight analysis (Kriegeskorte et al., 2006) with 4-mm radius spheres, and for each sphere the mean classification accuracy, as well as the confusion matrix of predicted against actual classes, were calculated across leave-one-subject-out folds.

A Monte Carlo procedure was applied to assess the significance of the classification accuracies. In particular, the spmT map labels were randomly permuted over the factor levels in each sphere 160000 times, in order to compare the actual classification accuracy against the random permutation distribution with a declared Bonferroni-corrected p < .05 threshold (equivalent to a p < 6.62 x 10^-6, that is, p < .05 / 7550 spheres/comparisons). In case of null results, comparisons at an uncorrected p < .0001, or subsequently at an uncorrected p < .001, were carried out.

The mean classification accuracies, together with the corresponding confusion matrices, are reported for the above threshold searchlight spheres, resulting from the analysis of 2 different spmT map sets. A first set entered the 3 x 3 and 2 x 2 multivariate pattern analyses as a function of Word Contrast, with each classification problem assessed: i) across the 5 Word Contrast levels (PT-2, PT-1, PT, PT+1, Full); ii) across levels below perceptual threshold (PT-2, PT-1); and iii) across levels above perceptual threshold (PT+1, Full). The second set entered the 3 x 3 and 2 x 2 multivariate pattern analyses as a function of PAS ratings, with each classification problem assessed: i) across all PAS scores (PAS{1,2}, PAS{3,4}); ii) across PAS{1,2}; and iii) across PAS{3,4}.

For the 2 x 2 classification problems, we assessed the average spmT value across the voxels of any significant searchlight spheres, and counted in each of the regions of interest within the MOVN and the EMLN the presence of spheres whose average estimate reflected a greater activation for condition A (e.g., MO) minus B (e.g., EMneg), or vice versa for condition B minus A (note that each region of interest may contain spheres with average activation in just one of these two directions, in both, or no significant spheres at all).

The Generalized Linear Mixed Model analysis on the PAS behavioral responses during the fMRI session showed that the Word Contrast level significantly modulated the level of subjective awareness, χ²(4) = 2286.3, p = 2.2 x 10^–16 (Fig. 2A). The mean PAS response to words presented with null visual contrast (Empty baseline) was 1.169 (95% confidence interval = 0.036; n = 21; Fig. 2A), thus close to the lowest PAS rating score (i.e., 1) and indicating no perceptual experience. Post-hoc multiple comparison tests (Tukey’s corrected) revealed that every pairwise difference between levels of Word Contrast was significant (all p < .001). The Word Type main effect (χ²(3) = 4.107, p = .250) and the Word Type by Word Contrast interaction (χ²(12) = 0.847, p = 1) were not significant predictors of the PAS scores.

We found that processing MO elicited greater activations than EM (i.e., EMneg+EMpos) in all the regions of the MOVN (Table 2; Fig. 2B). In turn, when considering EM against MO, we found activations in the sACC, the left OFC, and the LAMY (Table 2; Fig. 2B). We found no specific activation for neither the simple nor the quadratic effect of valence, which therefore were not further investigated.

We examined whether the levels of Word Contrast modulated activations within the MOVN and the EMLN following a subliminal profile (i.e., activation both above and below perceptual threshold), according to our experimental hypothesis. For comparison, we also tested a linear (i.e., gradual activation increase with contrast) and a stepwise (i.e., activation only above perceptual threshold) BOLD amplitude model.

Subliminal BOLD modulations for MO compared to EM were found in three regions of the MOVN, that is, LVPM, LIPL, and LSPL (Table 3; Fig. 2C). Linear and stepwise modulations for MO compared to EM were found in all regions of the MOVN (Table 3). In turn, we did not find any subliminal, linear, nor stepwise BOLD modulations for EM compared to MO.

To test the presence of Word Type effects in the absence of subjective awareness, independently of Word Contrast, we further analyzed BOLD responses as a function of PAS ratings. We tested whether activations within the regions of interest conformed to a subliminal and/or a stepwise BOLD amplitude model.

We found stronger subliminal and stepwise BOLD modulations for MO compared to EM in all regions of the MOVN (Table 4; Fig. 2D). We did not found any stronger subliminal nor stepwise BOLD modulations for EM compared to MO.

To evaluate whether activation differences between MO and EM, as resulting from the univariate fMRI analyses, reflected distinct category-specific neural codes that could be distinguished by a machine learning classifier, we also ran a searchlight MVPA on the same fMRI data.

In the 3 x 3 analyses as a function of Word Contrast (Fig. 3A), reflecting the main effect of Word Type, significant spheres distinguishing among the 3 types of words with accuracy above chance (chance level = 33.3%) were found, either when we considered all Word Contrast levels (42 spheres distributed over LMTG, RAMY, LOFC, ROFC, LINS, sACC; mean classification accuracy = 46.1%), or just those above PT (7 spheres distributed over LIPL, LMTG, RAMY, sACC; mean accuracy = 53.1%). The Word Contrast levels below PT were only distinguished at an uncorrected significance threshold (p < .0001; 4 spheres over ROFC and LINS; mean accuracy = 50.6%).

In the 2 x 2 analyses as a function of Word Contrast (Table 5), reflecting the pairwise post-hoc comparisons between the different word types, the MVPA classifier was able to distinguish significantly above chance (chance level = 50%) among all word types pairs, either when we considered all Word Contrast levels, just those above or just those below perceptual threshold (6 out of 9 pairwise comparisons passed Bonferroni correction, 2 passed an uncorrected p < .0001, and 1 an uncorrected p < .001). Above threshold classification spheres were variably distributed across the 10 regions of interest, but with a relative greater proportion in regions of the MOVN of spheres with greater activation for MO, and a relatively greater proportion of spheres whose average BOLD estimate reflected a greater activation for EMneg/EMpos in the EMLN regions of interest.

Also in the 3 x 3 analyses as a function of PAS ratings (Fig. 3B), the three word types were significantly distinguished above chance (chance level = 33.3%), either when we considered all PAS levels (3 spheres distributed over LMTG and ROFC; mean classification accuracy = 51.9%), or just those indicating presence of subjective awareness (1 sphere in LVPM; mean accuracy = 58.3%). When we considered just the PAS levels indicating absence of subjective awareness, significant spheres were only found at a lower significance threshold (uncorrected p < .001; 8 spheres over LMTG, LOFC, ROFC, LINS, and sACC; mean accuracy = 54.6%).

In the 2 x 2 analyses as a function of PAS ratings (Table 6), the MVPA classifier distinguished above chance all word types pairs and for all three PAS rating level combinations, be it when considering all levels, just those indicating presence, or just those indicating absence of subjective awareness (4 out of 9 pairwise comparisons passed Bonferroni correction, 3 passed an uncorrected p < .0001, and 2 an uncorrected p < .001). Spheres contributing to above-threshold classification were sparsely distributed across the regions of interest. In addition, LAMY and LSPL did not contribute to any of the pairwise classifications. Overall, there was a relatively greater proportion of spheres whose average BOLD estimate reflected a greater activation for EMneg/EMpos than for MO in EMLN regions of interest, but a more balanced distribution in MOVN.

The present fMRI study aimed at testing whether brain regions coding for experience-derived information that constitutes conceptual representations are automatically recruited during the subliminal perception of MO- and EM-related words. Univariate analyses of BOLD signal showed an MO word-specific activation of the MOVN, a left-lateralized visuomotor network involved in object-directed action and object identification (Caspers et al., 2010; Ishibashi et al., 2016), and supposed to represent and process visual and motor-related features such as manipulability and affordances (Buxbaum & Kalénine, 2010; Sakreida et al., 2016). Notably, this activation was found even under subliminal unaware processing conditions, both when the words were presented below perceptual threshold and when participants subjectively reported a lack of stimulus awareness. EM word processing specifically engaged the EMLN, a bilateral limbic network involved in emotion perception and recognition (Dahlén et al., 2022; Dalgleish, 2004), and supposed to represent affective properties (Kober et al., 2008; Lindquist et al, 2012). Univariate analyses, however, did not provide evidence for EM word-specific effects under subliminal unaware processing conditions. MVPA, instead, showed that distinct MO and EM word-specific neural codes could be distinguished above chance in MOVN and EMLN, when words were presented both above and below the perceptual threshold, as well as with and without subjective perceptual awareness.

Previous CFS studies have provided conflicting evidence for the recruitment of the MOVN as a function of awareness when processing tool pictures subliminally: whereas some works found activations in some but not all regions of this left-hemispheric premotor-parietal-temporal network (Almeida et al., 2010; Fang & He, 2005; Suzuki et al., 2014), other studies did not report any significant effects (Fogelson et al., 2014; Hesselmann & Malach, 2011; Hesselmann et al., 2011; Ludwig et al., 2015, 2016). In contrast, we previously showed the recruitment and effective connectivity of the MOVN both under aware and unaware processing of tool pictures (Ghio et al., 2022; Tettamanti et al., 2017). Here, we confirmed and expanded these findings, in line with the hypothesis that semantic features are automatically retrieved also during MO words processing, thus despite the lack of a visual or other sensory depiction of tools, and so excluding the potentially confounding effect of perceptual properties, such as the prototypical elongated shape of manipulable objects. Indeed, we found that MO words elicited a subliminal BOLD response profile as a function of Word Contrast in LVPM, LIPL, and SPL of the MOVN, that is a stable activation regardless of stimulus perception being under aware or unaware conditions. These three regions also exhibited a linear and a stepwise response profile, in concomitance with the subliminal one. Taken together, the significance of all three modulations indicates activation both above and below perceptual threshold, with a relatively stronger activation with than without awareness, and with relatively higher response amplitudes as the word contrast level increased. A fourth MOVN area, namely LMTG, mainly implied in object identification (Bracci et al., 2012), showed only linear and stepwise hemodynamic modulations. However, this effect does not contradict our hypothesis, as it still suggests a significant activation of this region below the perceptual threshold, although weaker than above. Furthermore, assessing fMRI activation as a function of PAS ratings revealed significant subliminal and stepwise modulations for all MOVN regions, indicating that these regions are activated both in the presence and, although to a lesser extent, in the absence of subjective awareness. The fact that brain activity for unaware semantic processing is detectable albeit weaker compared to that for supraliminal processing is consistent with previous fMRI data (Tettamanti et al., 2017), and also with event-related potential evidence that semantic activation may be more transient for unaware than aware stimuli (Kiefer & Spitzer, 2000). Consistently with the univariate analyses, MVPA showed that a classifier trained on the activity in MOVN regions was able to successfully tell apart MO and EM activation maps, not only above but also below perceptual threshold, and both with and without subjective perceptual awareness. Taken together, our results suggest that subliminal perception of visually presented MO words automatically activates conceptual-semantic decoding of visuomotor and functional MO-related information that is likely encoded during previous experiences with tool exemplars and stored in the MOVN (Binder & Desai, 2011; Lambon Ralph et al., 2017). This evidence complements previous findings, which demonstrated subliminal conceptual feature-specific processing against general semantic processing (Trumpp et al., 2013).

Regarding EM word processing, we found that it specifically involved the EMLN, a network engaged in what Lindquist et al. (2012) defined as “core affect” domain. This term refers to the mental representations sustaining the semantic processing of discrete emotions, and comprising bodily sensations—somatovisceral, kinesthetic, proprioceptive, and neurochemical—experienced along two main dimensions: valence and arousal. Previous CFS studies have suggested that valence plays an important role in an emotion stimulus emergence to consciousness (Jiang & He, 2006; Vizueta et al., 2012), even though evidence is inconsistent. While some studies reported a negativity bias (i.e., negatively-valenced words emerging from suppression faster than positively-valenced and neutral stimuli; Y. H. Yang & Yeh, 2011; Sklar et al., 2012), other works reported the opposite pattern (Vinson et al., 2011) or no effect of valence (Cheng et al., 2019; Rabagliati et al., 2018). Our univariate analysis revealed no subliminal response profile, either as a function of Word Contrast or as a function of PAS scores, and no parametric effects of valence, either simple or quadratic. However, the entire EMLN was consistently implicated in the multivariate searchlight decoding of EMneg versus EMpos activation patterns, both when considering contrast levels below perceptual threshold and the absence of subjective awareness. Considering the theoretical and methodological differences between univariate and multivariate analysis (Hebart & Baker, 2018), which make the latter more sensitive to detecting mental representations (Norman et al., 2006), our findings suggest that voxels within EMLN regions are likely not more activated for positively than negatively valenced EM words (or vice versa), but rather exhibit patterns of distributed activations or de-activations that give rise to a clearly identifiable EM-specific neural coding in the EMLN, even under unaware processing.

Our results about the subliminal activation of the MOVN and the EMLN suggest that unaware processing of visually presented MO and EM words automatically elicits the retrieval of, respectively, tool-related visuomotor properties (e.g., features useful to identify objects as well as to functionally interact with them, such as manipulability and affordance) and core affect (e.g., interoceptive and somatic) information. Since no overt task implying, for instance, imagery or deep semantic processing was required to participants, and since neither operationalized nor subjective awareness was required for word processing to activate the targeted brain networks, we suggest that the category-specific features encoded by MOVN and EMLN regions are constitutively integrated into the conceptual representation of MO and EM. These findings are consistent with grounded cognition theories of conceptual knowledge, according to which concepts rely, at least partially, on the reuse of specific information derived by previous experience with similar percepts (Kiefer & Harpaintner, 2020; Lambon Ralph et al., 2017; Meteyard et al., 2012). Notably, neither our results nor their interpretation as automatic retrieval of experience-derived properties contradict the grounded cognition hypothesis of context- and task-dependent conceptual flexibility (Binder & Desai, 2011; Kiefer & Harpaintner, 2020; Yee & Thompson-Schill, 2016), for which there is increasing evidence (e.g., Kuhnke et al., 2020). Based on previous studies reporting the recruitment of experiential brain systems (e.g., motor regions) even during shallow tasks (e.g., Hauk et al., 2004; Tettamanti et al., 2005) or with stimulus processing under unaware conditions (e.g., Trumpp et al., 2014; Trumpp et al., 2013), in our study we specifically tested the activation of experiential brain networks (MOVN and EMLN), which have been shown to represent experience-derived properties that are central and highly relevant to the conceptual representation of MO and EM words, such as sensory-motor experiences in the case of MO (see Willems & Casasanto, 2011), and internal (i.e., proprioceptive, somatovisceral) states in the case of EM (Winkielman et al., 2018). Since visuomotor and emotion information is crucial to the meaning of, respectively, MO and EM words, its automatic activation might be necessary also for unaware semantic processing, which still leaves room to concepts being flexibly activated in context- and task-specific ways.

A potential concern with the present study is that, as EM concepts are lower than MO in terms of imageability (which is a proxy for concreteness), we cannot rule out the possibility that the differential engagement of the MOVN and EMLN networks may reflect a difference between concrete and abstract word processing more in general. Indeed, concrete words have been shown to more strongly activate a general semantic network that includes (but is at least partially dissociated from) areas overlapping with the MOVN (Binder et al., 2009). Similarly, abstract concepts have been shown to activate a broader network that includes areas overlapping with the EMLN (Hoffman & Bair, 2024; Hoffman et al., 2015). However, we have here targeted brain activity in regions of interest based on previous semantic category-specific meta-analysis (Caspers et al., 2010; Dahlén et al., 2022; Ishibashi et al., 2016; Lindquist et al., 2012), as well as on previous studies that explicitly manipulated specific semantic features and related the function of such areas to the representation of MO- (Moody & Gennari, 2010; Tettamanti et al., 2017; J. Yang et al., 2011) and EM-related (Tettamanti et al., 2012; Vigliocco et al., 2014) semantic features. Therefore, by exploiting previous evidence, we were able to define and test specific predictions related to the selective recruitment of tool-related visuomotor knowledge and of core affect processes, in the case of MO and EM words respectively.

The hypothesis that these category-specific features are encoded over multiple experiences of EM and MO instances, represented in conceptual representations, and recruited when the corresponding concepts are semantically processed even implicitly and without stimulus awareness, is consistent with the view of the brain as a predictive system, evolved to efficiently meet the ongoing needs of the body (Adams et al., 2013; Clark, 2013). Whether the circumstances require to promptly interact with physical objects outside the body, or to cope with physiological states inside the body, knowing how to plan an appropriate response is essential for survival (Frye & Lindquist, 2025). In this scenario, language plays an important role: words rapidly and automatically activate properties of the entities they refer to, which overlap with those elicited during the perception of and the experience with such entities. By considering a broader perspective on language and semantic processing, this mechanism can be considered in line with the idea suggested by Lupyan and Clark (2015) and Lupyan and Ward (2013) that language acts as a flexible “artificial” context, according to which the brain unconsciously generates predictive beliefs about incoming inputs, possibly through a fast re-tuning of the organism’s entire semantic network, thereby helping constraining which type of top-down information is useful to recruit and what impact it has on lower-level processes as well as on inference and reasoning.

As just noted, we did not require our participants to overtly engage in mental imagery or deep semantic processing, because we wanted to measure spontaneous activation devoid of explicit semantic judgments, owing to the fact that semantic activation is automatically triggered by visual word processing, as best illustrated by the Stroop task (Flaudias & Llorca, 2014; Lorentz et al., 2016). In order to, nevertheless, keep control on the participants’ levels of awareness, we instead requested a subjective perceptual judgment at each trial, by means of the PAS scale (Ramsøy & Overgaard, 2004). Recently, it has been reported that the PAS scale may reduce or even suppress semantic processing in a semantic priming task (Kiefer et al., 2023). However, since there was no semantic priming manipulation in our task, the Kiefer et al. (2023) study does not provide any direct indications that semantic suppression may have occurred in our study. On the contrary, the detectable activation in the MOVN and EMLN even in the absence of awareness suggests that this was not a case of complete response suppression. It is, nevertheless, possible that the inclusion of the PAS rating somewhat altered the stimulus perception, for instance by increasing the participants’ effort in trying to identify the CFS masked word, and as a consequence by shifting the individual perceptual threshold to lower levels. However, even if present, this influence was at most redundant in the context of our task, in which the participants were inherently fostered to minimize the frustration of missing words, by trying to spot as many of them as possible.

In conclusion, previous evidence has demonstrated that visually perceiving objects in the absence of awareness elicits a comparable neurophysiological response signature to that evoked by their conscious perception, thus reflecting not only visual but also semantic processing (Fahrenfort et al., 2017; Tettamanti et al., 2017). We extend these findings by providing evidence that also tool words, which are arbitrarily associated with their referents through learning and do not carry any object-related visuomotor information, recruit a manipulable object-specific visuomotor network. Additionally, we generalize these results to the domain of discrete emotions, by demonstrating that processing emotion label words activates portions of the core affect limbic network. Crucially, manipulable object- and emotion-specific neural activations are decoded even in the absence of perceptual awareness, suggesting that experience-derived information constituting conceptual representations is automatically recruited.

The data and analysis code are available at https://doi.org/10.17632/gfb3xb5yty.1.

M.G., B.C., and M.T. designed the experiment; M.G. and B.C. collected the data; M.G, B.C., and M.T. analyzed the data; and M.G., B.C., and M.T. organized and wrote the manuscript.

M.G. was supported by a postdoctoral researcher grant by the University of Trento, Italy (Decree No. 48, 23.05.2019) awarded to M.T.

The authors declare no competing interests.

Supplementary material for this article is available with the online version here: https://doi.org/10.1162/imag_a_00484.