We first determined whether the voxelwise models based on a word vector space appropriately predict movie-evoked brain responses. The performance of brain-response prediction was evaluated using the following two measures: (1) prediction accuracy calculated as the correlation coefficients between predicted and measured brain responses in the test dataset, and (2) the fraction of significant voxels, among all cortical voxels, for which the prediction accuracy reached a significance threshold (p < 0.05 after the correction for multiple comparisons using the false discovery rate [FDR]; for more details, see Methods). The prediction accuracy averaged across all cortical voxels and across participants for the fastText, GloVe, and word2vec vector spaces was 0.0928, 0.0955, and 0.0967, respectively, for movie set 1 (28 participants) and 0.0865, 0.0884, and 0.0891, respectively, for movie set 2 (40 participants); no participant showed mean prediction accuracy less than 0 (Fig 3A, 3C, 3E, 3G, 3I and 3K). This accuracy was sufficiently high because it was averaged over all cortical voxels, and previous studies on voxelwise modeling have reported a similar tendency [20–22]. The fraction of significant voxels averaged across participants for the fastText, GloVe, and word2vec vector spaces was 0.575, 0.589, and 0.594, respectively, for movie set 1 and 0.526, 0.541, and 0.543, respectively, for movie set 2 (Fig 3B, 3D, 3F, 3H, 3J and 3I). These results indicate that the voxelwise models trained showed sufficient performance in the modeling of semantic representations consistently across different stimulus sets.

In the comparison of prediction performance between different types of word vectors, word2vec vectors showed significantly higher accuracy than the other vectors for both movie sets (Wilcoxon test, p < 0.05, FDR corrected) while GloVe vectors showed significantly higher accuracy than fastText vectors for both movie sets (p < 0.0001, FDR corrected). In the comparison of the fraction of significant voxels, word2vec and GloVe vectors showed significantly higher fractions than fastText vectors for both movie sets (p < 0.0001, FDR corrected) while the differences between word2vec and GloVe vectors were not significant for both movie sets (p > 0.06, FDR corrected). Nonetheless, the differences of the prediction performance between these vectors were totally small (<0.004 for mean prediction accuracy and <0.02 for the mean fraction of significant voxels).

Although the vector dimensionality of the word vector spaces had little effect on the model performance, there was no clear tendency toward vector-dimensionality dependency. The change of model performance was not monotonic against the change of vector dimensions (Kendall rank correlation, τ = −0.071–0.708, p > 0.17) and was inconsistent across movie sets and different vector types (S1–S3 Figs).

Previous studies performed voxelwise modeling using simpler, discrete word features, such as binary labels of objects and actions occurring in movie scenes (e.g., [23]). However, because such discrete word features have been examined separately from word vectors in voxelwise modeling, it is still unclear whether word vectors are more effective to model semantic representations in the brain than discrete word features. To address this, we compared these two types of semantic features, first of all, in terms of the performance of brain-response prediction. For this purpose, we constructed a voxelwise model using the binary labeling of semantic contents in each movie scene (binary-labeling model). Although this model was similar to one used in a previous study [23], we obtained binary labels of semantic contents by extracting words from the manual scene descriptions rather than manual word labeling performed in the previous study (for more details, see Methods). In this model, a semantic feature in a given scene was represented by a binary vector with 100–2000 dimension that corresponded to the occurrence of 100–2000 specific words in manual descriptions in that scene.

We found that the binary-labeling model constructed for each participant showed sufficiently high performance in brain-response prediction especially when high-dimensional vectors were used (S4 Fig). For 1000 and 2000 vector dimensions, the performance of the binary-labeling model was rather significantly higher than that of all the word vector-based models (Wilcoxon test, p < 0.05, FDR corrected); nonetheless, the performance differences between models were not large (<0.006 for mean prediction accuracy and <0.05 for the mean fraction of significant voxels). This result suggests that brain response to semantic information in natural scenes can be effectively predicted even using discrete word features.

Does this finding imply that the semantic relational structure of words, captured by word vectors [3,4,6], is ineffective in the modeling of cortical semantic representations? Comparing the word vector-based and binary-labeling model, however, cannot purely assess the effectiveness of semantic relational structure of word vectors because the procedures to transform scene descriptions to semantic features were different between these models. To examine the effectiveness more fairly, we compared the performance of brain-response prediction between the voxelwise model based on original, trained word vectors and that based on untrained word vectors. The same procedure as used for original word vectors was employed to obtain the untrained word vectors, except that the statistical training of word vectors was omitted. This procedure yielded 1000-dimensional random vectors, each assigned to each word in the vocabulary shared with original word vectors. Hence, the untrained vectors still had signatures of individual words but not the semantic relational structure of words at all. Then, the voxelwise model based on the untrained word vectors were constructed using the same procedure as used for the voxelwise model based on original, trained word vectors (see also Methods).

The performance of brain-response prediction (i.e., prediction accuracy and the fraction of significant voxels) was compared between the voxelwise model based on trained vectors and that based on untrained vectors. We found that the prediction performance of voxelwise models was consistently higher for trained vectors than for untrained vectors regardless of performance measures, word-vector types, and movie sets (S5 Fig). This result suggests that the semantic relational structure of words, captured by trained word vectors, is effective in the prediction of movie-evoked brain response, namely, in the modeling of cortical semantic representations.

We next identified which cortical regions were predictable by the word vector-based models in order to determine that the models could capture the semantic representations in cortical regions which are considered to be involved in audiovisual semantic processing. For this purpose, the prediction accuracy averaged across participants was calculated in each of the 148 cortical regions that were anatomically segmented using FreeSurfer [24,25] on the basis of the Destrieux atlas [26]. Then, the averaged prediction accuracy for each region was mapped onto the cortical surface of a reference brain (Fig 4 and Table 1). The highly predictable regions for each word vector-based model were localized over widespread cortical regions, including the occipital, superior and inferior temporal, and posterior parietal regions. This localization was consistent with previous reports, in which the semantic representations in the brain were modeled with word features [12,23]. There was a strong correlation of mean prediction accuracy over 148 cortical regions between different types of word vectors (Pearson’s r > 0.999; Spearman’s ρ > 0.999) and between movie sets 1 and 2 (Pearson’s r > 0.972; Spearman’s ρ > 0.973); this indicates the consistency of model predictability of these cortical regions across vector types and stimulus sets. In addition, there were similar tendencies across vector dimensionality (S6–S8 Figs and S1–S6 Tables). These results indicate that the word vector-based voxelwise models trained reliably showed the localization of highly predictable regions.

These results showed that the mean prediction accuracy in each brain region ranged up to 0.37 across word-vector types and movie sets. It might be argued that the prediction accuracy observed in this study was deemed much lower than that observed in previous studies (up to 0.6–0.8) [21,23]. However, this discrepancy is likely to be explained by the difference in spatial granularity used for calculating prediction accuracy; namely, we showed brain region-wise prediction accuracy whereas the previous studies reported voxelwise prediction accuracy. At the voxel level, our voxelwise models, including binary-labeling models, for individual brains exhibited prediction accuracy ranging up to ~0.7 (S9 and S10 Figs), which is consistent with the results from the previous studies [21,23].

A similar localization was also observed when the fraction of significant voxels in each cortical region was mapped onto the cortical surface (Fig 5 and Table 2). The fraction was relatively large in the occipital, superior and inferior temporal, and posterior parietal regions compared with the other regions. This localization pattern was observed to be highly consistent across vector types, movie sets, and vector dimensions (S11–S13 Figs and S7–S12 Tables). Taken together, these results suggest that the word vector-based models capture the semantic representations in appropriate cortical regions.

It should be noted that, however, the fraction of significant voxels was not small even in the other cortical regions, including the prefrontal cortex. The minimum fraction within any individual region was 0.237, which is significantly higher than 0.05 (corresponding to the significance level of 0.05; Wilcoxson test, p < 0.00001). In addition, even when the chance level of the fraction for each region was estimated from control voxelwise models in which word vectors were shuffled across vector dimensions for each vector (see also Methods), the fraction of significant voxels was significantly above the chance level in all the cortical regions (Wilcoxson test, p < 0.00001, FDR corrected; S14 Fig). These results suggest that although high prediction accuracy was observed in specific regions (Figs 4 and S6–S8), the word vector-based model can potentially capture semantic information even from other regions across the cortex.

In comparison between word vectors and discreate word features, we observed similar localization patterns of prediction accuracy (S15 Fig and S13 Table) and the fraction of significant voxels (S16 Fig and S14 Table). Therefore, even regarding the localization of predictable cortical regions, we did not find any clear differences between word vectors and discrete word features.

To assess the effectiveness of the sematic relational structure of word vectors in response prediction in each cortical region, we compared prediction performance (i.e., prediction performance and the fraction of significant voxels) between the voxelwise model based on trained word vectors and that based on untrained word vectors within each region. The results reveal that the model based on trained vectors exhibited significantly higher response-prediction performance in the majority of the 148 regions (S17 Fig; 130–144 regions for prediction accuracy; 136–148 regions for the fraction of significant voxels; Wilcoxon test, p < 0.05, FDR corrected). Even across regions, the model based on trained vectors outperformed the model based on untrained vectors regardless of performance measures, word-vector types, and datasets (Wilcoxon test, p < 0.05, FDR corrected). Although performance differences between these two models were not large, the improvement of the performance was observed comparably across regions rather than conspicuously in specific regions. These results suggest that the semantic relational structure of words, captured by word vectors, improves the modeling of semantic representations in widespread brain regions.

Next, we tested the correlation between the word dissimilarity matrix derived from the voxelwise models and that derived from behavioral data to clarify the behavioral correlates of modeled semantic representations. Word representations in the modeled brain space were calculated by multiplying the original fastText, GloVe, or word2vec word vectors by model weights, and the brain-derived word dissimilarity matrix was obtained from the correlation distance between all possible pairs of word representations (Fig 1B). Meanwhile, the behavior-derived word dissimilarity matrix was measured from the behavioral data from the word-arrangement task (Fig 2), in which 36 participants completed 18.8 (SD = 10.7) trials on average for each session (see Methods for more details). These brain- and behavior-derived word dissimilarity matrices were constructed by averaging the matrices over all brain models or all behavioral data separately for nouns and adjectives. In addition, to determine whether the behavioral correlates of word dissimilarity change through the transformation from original word vector representations to brain representations, we also calculated word vector-derived word dissimilarity matrices from the correlation distance between all possible pairs of original fastText, GloVe, or word2vec word vectors separately for nouns and adjectives.

Fig 6 shows the behavior-derived word dissimilarity matrices and the brain- and word vector-derived word dissimilarity matrices constructed using 1000-dimensional fastText word vectors. The brain-derived matrices were highly consistent across movie sets (Spearman’s ρ = 0.926 and 0.943 for nouns and adjectives, respectively). For both nouns (Fig 6A) and adjectives (Fig 6B), we found significant correlations between the brain- and behavior-derived matrices (brain–behavior correlations; permutation test, p < 0.0001). The correlation coefficients were larger for nouns than for adjectives (p < 0.0001). Although we also found significant correlations between the word vector- and behavior-derived matrices (word vector–behavior correlations; p < 0.0001), the brain–behavior correlations for nouns (but not for adjectives) were significantly stronger than the word vector–behavior correlations (p < 0.0001).

Fig 7 summarizes the brain–and word vector–behavior correlations for each word-vector type (i.e., fastText, GloVe, or word2vec; vector dimensionality = 1000). For all the vector types and for both nouns and adjectives, there were significant brain–behavior correlations as well as word vector–behavior correlations (permutation test, p < 0.0001, FDR corrected). Such significant correlations were observed regardless of vector dimensionality (S18–S20 Figs); although the correlation coefficients differed across vector dimensions, no clear tendency of changes in correlation coefficients across vector dimensions was determined. These results indicate that the brain-derived dissimilarity structure of semantic representations correlates with the behavior-derived dissimilarity structure regardless of vector types and dimensionality.

We then performed three different comparisons of behavioral correlations in Fig 7. In the comparison between nouns and adjectives, the brain–behavior correlations of noun dissimilarity were significantly stronger than those of adjective dissimilarity for all the vector types (permutation test, p < 0.0001, FDR corrected). In the comparison between brain–behavior and word vector–behavior correlations, there was a significant difference only for fastText noun dissimilarity (p < 0.0001, FDR corrected). In the comparison between vector types, the vector-type difference of brain–behavior correlations was partly significant but small (p < 0.05, FDR corrected) whereas the word vector–behavior correlations for fastText vectors was much weaker than those for the other types of vectors (p < 0.0001, FDR corrected).

In these cases, the populations from which we obtained these dissimilarity matrices were different between brain- and behavior-derived data. However, we observed a significant correlation for both nouns and adjectives and for both movie sets even when these matrices were collected from and averaged over the same population (six participants; permutation test, p < 0.0001; S21 Fig). At the level of individual participants, the brain–behavior correlation of noun dissimilarity was significant for each of all the six participants from this population (p < 0.01) whereas the correlation of adjective dissimilarity was significantly higher than chance level for only three participants of them (p < 0.05; S22 Fig). Together, the word vector-based models could capture perceptual noun and adjective dissimilarities at the population level and perceptual noun dissimilarity at the individual level even when the brain- and behavior-derived word dissimilarities were originated from the same population.

The brain-derived dissimilarity matrices described above (Figs 6 and 7 and S18–S22) were calculated from the model weights of all cortical voxels (the number of voxels was 54914–73301 [mean ± SD = 62245 ± 5018] for movie set 1 and 52832–73301 [mean ± SD = 62150 ± 5190] for movie set 2). However, the high performance of the voxelwise models was observed in localized cortical regions (Figs 4, 5, S6–S8 and S11–S13). Hence, the correlation of brain- and behavior-derived dissimilarity matrices may be stronger when only the cortical regions with high model performance were used. To test this possibility, we constructed brain-derived dissimilarity matrices using only those voxels with the highest prediction accuracy (top 2000, 5000, 10000, 30000, or 50000 voxels). Then, their correlations with behavior-derived matrices were compared with the original correlations calculated using all cortical voxels. However, we found the strongest correlation when using all cortical voxels regardless of word vector types (Figs 8 and S23–S25). This result suggests that semantic information correlated with behavior are distributed broadly across the cortex and can be captured using word vector-based voxelwise models.

To gain a further understanding of cortical localization of behaviorally correlated sematic representations modeled by word vectors, we examined brain–behavior correlations of word dissimilarity within each cortical region. We found that although moderate or strong correlations were observed across the cortex (Spearman’s ρ = 0.356–0.621 for noun dissimilarity, 0.246–0.400 for adjective dissimilarity), the correlations were relatively weak in low-level audiovisual regions, such as the posterior occipital cortex and the superior temporal area (Figs 9 and 10). This localization pattern of brain–behavior correlations across the cortex was deemed different from that of model prediction performance (Figs 4, 5, S6–S8 and S11–S13). This discrepancy is likely to cause the reduced brain–behavior correlation when voxels used for the analysis were selected according to high prediction accuracy (Figs 8 and S23–S25). Considering that the fraction of significantly predictable voxels was sufficiently large even in the high-level cortical regions showing strong brain–behavior correlations (Fig 5 and S11–S13), these results suggest that word vector-based models effectively capture semantic information in these regions, which is more closely associated with human perception than the low-level audiovisual regions.

Previous studies have reported that cortical representations in the object-selective visual areas show clear distinction between animate and inanimate categories [27–29]. Since the nouns we used for constructing word dissimilarity matrices consisted of six different semantic categories along animate/inanimate dimension (humans, non-human animals, non-animal nature things, constructs, vehicles, and other artifacts; see also Methods), it might be argued that the observed behavioral correlates of brain-derived word dissimilarity simply reflected such clear distinction of animate and inanimate representations. However, even when brain–behavior correlations were examined separately for each of the six categories, we found significant brain–behavior correlations for each of the “humans”, “non-animal natural things”, “vehicles”, and “other artifacts” categories (p < 0.05, FDR corrected; S26 Fig). This result indicates that the observed brain–behavior correlations of word dissimilarity cannot be explained only by simple representational distinction between animate and inanimate categories.

Finally, we compared the brain-derived word dissimilarity modeled by word vectors and that modeled by discrete word features in terms of their behavioral correlates. In contrast to the performance of brain-response prediction (S4, S15 and S16 Figs), the brain–behavior correlations of word dissimilarity for the binary-labeling models were much weaker than those for the word vector-based models regardless of datasets and vector dimensionality (S27 Fig). This result suggests that the voxelwise modeling based on word vectors have a clear advantage in capturing the similarity structure of cortical semantic representations associated with human perception.

The experimental protocol was approved by the ethics and safety committees of the National Institute of Information and Communications Technology. Written informed consent was obtained from all of experimental participants.

In total, 52 healthy Japanese participants (21 females and 31 males; age 20–61, mean ± SD = 26.8 ± 8.9 years) were recruited for the 2 sets of fMRI experiments. Among the recruits, 36 participated in one or the other of these fMRI experiments, and 16 participated in both. In addition, another 36 healthy Japanese participants (20 females and 16 males; age 18–58, mean ± SD = 25.9 ± 10.1 years) were recruited for psychological experiments. Of these, 6 were also participants of the fMRI experiments and 30 were unique participants. All participants had normal or corrected-to-normal vision. The fMRI data, but not the psychological data, used here were also utilized in a previous publication [22].

Functional and anatomical MRI data were collected via a 3T Siemens MAGNETOM Prisma scanner (Siemens, Germany), with a 64-channel Siemens volume coil. Functional data were collected using a multiband gradient echo EPI sequence [52] (TR = 1000 ms; TE = 30 ms; flip angle = 60°; voxel size = 2 × 2 × 2 mm; matrix size = 96 × 96; FOV = 192 × 192 mm; the number of slices = 72; multiband factor = 6). Anatomical data were also gathered using a T1-weighted MPRAGE sequence (TR = 2530 ms; TE = 3.26 ms; flip angle = 9°; voxel size = 1 × 1 × 1 mm; matrix size = 256 × 256; FOV = 256 × 256 mm; the number of slices = 208) on the same 3T scanner.

In the two sets of fMRI experiments, participants were asked to view movie stimuli on a projector screen inside the scanner (27.9 × 15.5 of visual angle at 30Hz) and used MR-compatible headphones for the sounds. The participants were given no explicit task. The fMRI data for each participant upon viewing of the movies were collected in three separate recording sessions over 3 days for each set of fMRI experiments.

The movie stimuli consisted of Japanese television advertisements for one set of experiments (movie set 1) and Japanese web advertisements for the other set of experiments (movie set 2; see also [22]). Movie set 1 has included 420 ads broadcasted on Japanese TV between 2011 and 2017; meanwhile, movie set 2 included 368 ads broadcasted on the Internet between 2015 and 2018. The ad movies were all unique, include a wide variety of product categories (see S17 Table for more details). The length of each movie was 15 or 30 s. To create the movie stimuli for each experiment, the original movies in each of the movie sets 1 and 2 were sequentially concatenated in a pseudo-random order. For each movie set, 14 non-overlapping movie clips of 610 s in length were obtained.

Individual movie clips were then displayed in separate scans. The initial 10 s of each clip served as a dummy in order to discard hemodynamic transient signals caused by clip onset. fMRI responses collected during the 10 s dummy part were not used for modeling. Twelve clips from each movie set were only presented once. The fMRI responses to these clips were used for the training of the voxelwise models (training dataset; 7200 s in total). The other two clips for each movie set were presented four times each in four separate scans. The fMRI responses to these clips were then averaged across four scans to improve the signal-to-noise ratio [11,23,53]. The averaged responses were further used for the test of the voxelwise models (test dataset; 1200 s in total).

Motion correction in each functional scan was carried out via the statistical parameter mapping toolbox (SPM8, http://www.fil.ion.ucl.ac.uk/spm/software/spm8/). For each subject, all volumes were aligned to the first image from the first functional run. Low-frequency fMRI response drift was then eliminated by subtracting median-filtered signals (within a 120-s window) from raw signals. Then, the response for each voxel was normalized by subtracting the mean response and scaling to the unit variance. FreeSurfer [24,25] was used to identify cortical surfaces from anatomical data and register these surfaces to the voxels of functional data. Then, each voxel was assigned to one of the 148 cortical regions derived from the Destrieux atlas for cortical segmentation [26].

This study used fastText skip-gram [4], GloVe [3], or word2vec skip-gram [2] to construct a word vector space. These algorithms have been originally developed to learn a word vector space based on word co-occurrence statistics in natural language texts.

The training objective of the skip-gram algorithm of fastText and word2vec is to obtain word vector representations that enable the surrounding words to be accurately predicted from a given word in a sentence [2,4]. More formally, given a sequence of training words w₁, w₂,…,w_T, the skip-gram algorithm seeks a K-dimensional vector space that maximizes the average log probability, given as: 1T∑t=1T∑−c≤j≤c,j≠0logp(wt+j|wt) where c is the size of the context window, which corresponds to the number of to-be-predicted words before and after the center word w_t. Therefore, the skip-gram vector space is optimized on the basis of the local co-occurrence statistics of nearby words in the text corpus. The basic formulation of p(w_t+j|w_t) is the softmax function: p(wy|wx)=exp(〈vwx,vwy〉)∑i=0W(〈vwx,vwi〉) where v_wi is the vector representation of w_i, W is the number of words in the vocabulary, and 〈v₁, v₂〉 indicates the inner product of vectors v₁ and v₂. However, because of the high computational cost of this formulation, the negative sampling technique is used to produce a computationally efficient approximation of the softmax function [2]. In addition, fastText introduces sub-word modeling, which is robust for inflected and rare words [4].

GloVe learns word vector representations using a global log-bilinear regression model that is subject to the cost function: J=∑i,j=1Vf(Xij)(wiTw˜j+bi+b˜j−logXij)2 where V is the vocabulary size, wi∈Rd is a d-dimensional vector of word i, w˜j∈Rd is a separate context vector of word j, and b_i and b˜j are additional biases. X is the co-occurrence matrix, where X_ij represents the number of occurrences of word i in the context of word j. f(X_ij) is a specific weighting function. GloVe combines the advantages of local context window and global matrix factorization methods [3].

A vector space of each algorithm was constructed from a text corpus of the Japanese Wikipedia dump on April 1, 2020. All Japanese texts in the corpus were segmented into words and lemmatized using MeCab (http://taku910.github.io/mecab). Used were only nouns, verbs, and adjectives. In an attempt to improve the reliability of word-vector learning, the vocabulary size was restricted to ~100,000 words by excluding words that appeared infrequently in the corpus. The learning parameters of fastText and word2vec used were as follows: window size = 10; the number of negative samples = 5; downsampling rate = 0.001; and the number of learning epochs = 10. The learning parameters of GloVe used were as follows: window size = 10; initial learning rate = 0.05; α = 0.75; and the number of training iterations = 50. The vector spaces were constructed using five different numbers of vector dimensions (100, 300, 500, 1000, and 2000). Of these, the 1000-dimension vector space of each algorithm was used for the main analysis (Figs 3–10), and the other dimensions were used to test the effect of vector dimensionality on modeling performance and behavioral correlates. To eliminate the effects of trial variations in the learning quality of vector spaces of each algorithm on the performance of voxelwise modeling, five different vector spaces were learned independently using the same text corpus and the same algorithm with the same parameters. All results were obtained from the average over five learned spaces for each parameter set.

Manual scene descriptions using natural Japanese language were provided for every 1-s scene of each movie in movie sets 1 and 2, in a manner similar to that described previously [11,12,22]. The annotators were native Japanese speakers (movie set 1: 68 females and 28 males, age 19–62 years; movie set 2: 11 females and 2 males, age 20–56 years), who were not the fMRI participants. They were instructed to describe each scene (the middle frame of each 1-s clip) using more than 50 Japanese characters. Multiple annotators (movie set 1, 12–14 annotators; movie set 2, 5 annotators) were randomly assigned for each scene to reduce the potential effect of personal bias. The descriptions contain a variety of expressions reflecting not only objective perceptions but also the subjective perceptions of the annotator (e.g., impression, feeling, association with ideas; for more details, see https://osf.io/3hkwd).

Each description for a given scene was also segmented, lemmatized, and decomposed into nouns, verbs, and adjectives via MeCab as described above; then, they were transformed into fastText, GloVe, or word2vec vectors. The word vectors were then averaged within each description. For each scene, all vectors obtained from the different descriptions were averaged. Through this procedure, a single vector (scene vector) was obtained for each 1-s scene, which was later used for modeling.

To compare discrete word features with word vectors in terms of the modeling of brain semantic representations, the binary labeling of semantic contents in each movie scene was used to construct voxelwise models. The binary labels were obtained from the presence/absence of words in the manual movie-scene descriptions. For each scene, individual words (nouns, verbs, and adjectives) in the vocabulary were counted in all descriptions for the scene. If a given word was present, the binary label of the word was 1. If not, the binary label of the word was 0. This labeling was performed over all movie scenes. Then, scene vectors with the dimensionality of 100, 300, 500, 1000, and 2000 were obtained from the binary labels of 100, 300, 500, 1000, and 2000 words that most frequently appeared in the movie-scene descriptions for each movie set. This type of semantic features is similar to ones used in a previous study on the modeling of cortical semantic representations [23]. However, the previous study performed the binary labeling of movie scenes using the manual labeling of single words by one annotator, unlike our binary labeling.

The procedure of voxelwise modeling was similar to that described previously [12]. A series of fMRI responses evoked in individual N voxel by a series of S movie scenes, represented by a S×N matrix R, was modeled as a weighted linear combination of a K-dimensional scene-vector matrix V plus isotropic Gaussian noise ε.

A set of linear temporal filters was used to model the slow hemodynamic response and its coupling with brain response [53]. In an attempt to capture the hemodynamic delay in the responses, the S×4K scene-vector matrix was constructed by concatenating four sets of K-dimensional scene vectors with delays of 3, 4, 5, and 6 s. The 4K×N weighted matrix W was estimated using an L2-regularized linear least-squares regression, which can obtain good estimates even for models containing a large number of regressors [23].

The model training was conducted using the training dataset (7200 datapoints). In order to estimate the regularization parameters, the training dataset was randomly divided into two subsets containing 80% and 20% of the samples, for model fitting and validation, respectively. This random resampling procedure was repeated 10 times. Regularization parameters were optimized, according to the mean Pearson’s correlation coefficient between the predicted and measured fMRI signals for the 20% validation samples. An optimal parameter was obtained separately for each model.

The brain-response prediction performance of the model was evaluated using the test dataset (1200 datapoints), which was not used for model fitting or parameter estimation. The prediction accuracy of the model was quantified as the Pearson’s correlation coefficient between the predicted and average measured fMRI signals in the test dataset. In addition, the significance of prediction accuracy (Pearson’s correlation) for each of all cortical voxels was evaluated after the correction for multiple comparisons using FDR. The rate of voxels, to all cortical voxels, for which the prediction accuracy reached a significance threshold (p < 0.05, FDR corrected) was calculated as the fraction of significant voxels, which was another measure of brain-response prediction performance for each model.

To test whether the fraction of significant voxels for each cortical region was significantly above chance level, two types of statistical tests were performed. First, using the fraction for each model of each participant as a data sample, whether or not the fraction was significantly higher than a standard significance threshold (p = 0.05) was tested using Wilcoxon test with FDR correction. Second, the chance level of the fraction for each cortical region was more conservatively estimated using control voxelwise models based on word vectors that were shuffled across vector dimensions for each vector. Five different sets of the shuffled word vectors were generated from each word-vector space and used to construct five different control voxelwise models for each participant. Then, the fraction of significant voxels was obtained for each cortical region using these control models and averaged across the five different models, which produces the chance level of the fraction for each participant. Whether the fraction of significant voxels for each original model was significantly above the chance level was tested for each region using Wilcoxon test with FDR correction while the fraction for each participant was used as a data sample.

Word vectors effectively capture the semantic relational structure of words as demonstrated by previous studies that applied word vectors to word similarity judgements [3,4,6]. To determine whether such semantic relational structure of word vectors is effective in the modeling of cortical semantic representations, the response-prediction performance of voxelwise models using original word vectors was compared with that using untrained word vectors. Untrained word vectors were obtained with the same procedure as used for the generation of original, trained word vectors, except that the statistical training of word vectors was omitted. This procedure yielded 1000-dimensional random vectors that were each assigned to each word in the vocabulary shared with the corresponding trained vectors. Hence, untrained vectors had signatures of individual words but not the semantic relational structure of words at all. Then, scene vectors were generated from untrained vector representations of scene descriptions.

Voxelwise models based on untrained vectors were constructed using the same procedure as used for voxelwise models based on trained vectors. Five different voxelwise models for each participant were constructed from five different sets of untrained vectors. The performance of brain-response prediction (i.e., prediction accuracy and the fraction of significant voxels) for each participant was obtained by averaging the performance of these five models. Finally, the performance for the models based on untrained vectors was statistically compared with that for the models based on trained vectors. Higher performance for the models based on trained vectors indicates that the semantic relational structure of word vectors is effective in the modeling of cortical semantic representations.

Participants were asked to perform a word-arrangement task on a PC. This task is considered to be a modified version of the psychological task introduced previously [19]. Kriegeskorte and colleagues originally used this task to examine the behavioral correlates of cortical object representations, which were quantified using the representational similarity of fMRI signals evoked by object stimuli [54,55]. The present study used this task to study the methodological validity of word vector-based brain modeling by testing the behavioral correlates of modeled cortical semantic representations. The psychological data are available online (https://osf.io/um3qg/).

In each trial of this task, the participants were required to arrange ≤60 words (nouns or adjectives) in a two-dimensional space according to their semantic relationship on a computer screen by mouse drag-and-drop operations (Fig 2). This paradigm has allowed for the efficient collection of the perceptual semantic dissimilarity between words from participants [19].

The words used in this task included 60 nouns and 60 adjectives in Japanese (Table 3). These words were selected from the vocabulary of the fastText space (i.e., top 100,000 most frequently used words in the Japanese Wikipedia corpus). Nouns were selected in terms of the following six semantic categories: humans, non-human animals, non-animal natural things, constructs, vehicles, and other artifacts. Ten words were selected for each category. For adjectives, such category-based selection was deemed difficult due to the small number of adjectives in the vocabulary (only 473 words). Nouns and adjectives were separately used in two distinct sessions of the task, which was performed over 2 days.

The words were arranged in a designated circular area on the computer screen (“arena”). The words were initially displayed outside the arena. The participants used mouse drag-and-drop operations in moving each word item into the arena and arranging them. The arrangement of words was according to the participants’ own judgment of semantic similarity and dissimilarity between words. More similar or dissimilar word pairs should be closer or further apart in the arena, respectively. The participants were allowed to move any words within the arena as many times as they wished. Once all words were moved into the arena, the participants could click a button marked “Next” anytime to move on to the next trial.

On the initial trial of each session, the participants have arranged the entire set of 60 words (nouns or adjectives). On subsequent trials, they arranged subsets of those 60 words. After the end of each trial, the rough estimate of the word dissimilarity matrix and the evidence (0–1) for each word pairwise dissimilarity were computed as described previously [19]. Words in the subsets chosen for each trial were determined so as to increase evidence for pairwise dissimilarities of the words whose evidence estimated on the previous trial was the weakest. The session continued until the evidence of every word pair dissimilarity was above a threshold (0.75) or until the total duration of the session approached 1 h.

The brain-derived word dissimilarity matrix for the word vector-based models was estimated using the following procedure: First, the 4K×N voxelwise model weights for each participant were transformed into K×N by averaging the weights across the four sets of hemodynamic delay terms. Second, all N or top M voxels with the highest prediction accuracy were then selected from the model weights. All voxels were used in the main analysis, and selected voxels (M = 2000, 5000, 10000, 30000, and 50000) were used in an additional analysis that tested the effect of voxel selection on brain–behavior correlation. Third, the K-dimensional fastText, GloVe, or word2vec vectors for the words used in the psychological experiments (Table 3) were multiplied by the K×N weight matrix, which yielded N- or M-dimensional word representations in the modeled brain space. Finally, a word dissimilarity matrix was calculated by the correlation distance (1 –Pearson’s correlation coefficient) between these word representations, which was averaged across participants separately for nouns and adjectives.

The brain-derived word dissimilarity matrix for the voxelwise model based on binary labels (binary-labeling model) was also estimated. For the binary-labeling model, the K×N matrix into which the 4K×N model weights were transformed by averaging across hemodynamic delay terms corresponded to the brain representations of K words. Accordingly, a word dissimilarity matrix of this model was calculated by the correlation distance between these N-dimensional word representations. Note that unlike the word vector-based models, the representations for the binary-labeling model contained only a limited part of the words used in the psychological experiments (Table 3). The number of words contained depended on vector dimensionality and movie sets (S18 Table).

The behavior-derived word dissimilarity matrix was evaluated from multiple word-subset arrangements on the word-arrangement task. The dissimilarity for a given word pair was estimated as a weighted average of the scale-adjusted dissimilarity estimates from individual arrangements as has been described previously [19]. These estimates produced a word dissimilarity matrix for the entire set of words. In this way, the word dissimilarity matrix was estimated separately for nouns and adjectives and was averaged across participants.

In addition, a word vector-derived word dissimilarity matrix was also examined in order to test whether the behavioral correlates of semantic dissimilarity structures changed through the transformation from original word vector representations to brain word representations. The word vector-derived matrix was computed by the correlation distance between fastText, GloVe, or word2vec vectors of the same word sets (Table 3) separately for nouns and adjectives.

The behavioral correlates of brain- or word vector-derived data were evaluated using Spearman’s correlation between the brain- and behavior-derived word dissimilarity matrices (brain–behavior correlation) and between the word vector- and behavior-derived word dissimilarity matrices (word vector–behavior correlation). The upper triangular portion of each matrix was used for calculating correlation coefficients. For the binary-labeling model, correlation coefficients were calculated using only the words contained in the vocabulary of the model (S18 Table).

A permutation test was performed to evaluate the statistical significance of brain–behavior and word vector–behavior correlations. In each repetition of this test, the rows of a brain- or word vector-derived word dissimilarity matrix were randomly shuffled. Then, this shuffled brain- or word vector-derived matrix was used to calculate a Spearman’s correlation with the original behavior-derived matrix. This procedure was repeated 10,000 times to obtain a null distribution of correlation coefficients and estimate the p value of an actual correlation coefficient. The difference of correlation coefficients between different matrix pairs was also tested in a similar manner. In this case, the difference of brain–or word vector–behavior correlation coefficients between different matrix pairs was calculated in each repetition while a brain- or word vector-derived matrix was shuffled separately for each of the different matrix pairs. This procedure was repeated 10,000 times to estimate a null distribution of the correlation coefficient differences.