Article Figures & Data
Figures
- Figure 1.
Experimental design. A, Experimental design and trial structure. Our design crossed strength of association (low vs high) and compositionality (list vs comp). In each trial, participants indicated whether the target picture matched the preceding words. Half of the target pictures matched and half did not. Activities recorded from 100 ms prestimulus onset to 1200 ms poststimulus onset were analyzed. B, Schematic diagram showing how cosine value between high-dimensional vectors represents semantic similarity. The smaller the angle is between two vectors, the higher the cosine value and semantic similarity. The angle between “French” and “cheese” is smaller than the one between “Korean” and “cheese” because they share more contextual features. The high-dimensional word embeddings were visualized on the 2D scatter using t-SNE. Extended Data Figures 1-1, 1-2, 1-3 support Figure 1.
- Figure 2.
Schematic illustration of the analysis procedure. A, Two-stage multiple regression analyses. At the first stage, an ordinary least squares regression was applied to each participant's single-trial source estimates for each source within a selected region at each time point of the analysis window. At the second stage, a one-sample t test was performed on the distribution of β values across subjects for each variable at each source and each time point, to test whether their values were significantly different from zero. Significance was determined by TFCE correction with 10,000 permutations. B, Searchlight multivariate pattern classification analyses. A linear SVM was trained on the combination of pseudo-trials from two conditions and tested on a left-out pair with 100 permutations. The same SVM analysis was applied independently to each source and time point within a language mask. Classification accuracy averaged over subjects at source and time point minus the chance level of 50% was submitted to a one-sample t test and significance was determined by TFCE correction with 10,000 permutations. C, Directed connectivity analyses. The RDMs of the source estimates of the two fROIs derived from the regression and the classification analyses at time point t (D(A, t) and D(B, t)) were calculated as 1 minus the correlation between the conditions. Directed activity from region A to region B was quantified as the partial correlation coefficient between D(A, t−dt) and D(B, t), partialling out D(B, tdt), where dt is the time interval between the current time point and the previous time point. Significance of coefficients >0 was determined by 10,000 permutations with an α level of 0.05.
- Figure 3.
Behavioral results. A, Mean predicted accuracy and reaction time for the four conditions, regressing out the effect of word frequency. Association is significant for accuracy (p = 0.049) and marginally significant for reaction time (p = 0.08). Error bars indicate 1 SE. B, Fitted coefficients for all predictors for accuracy and log reaction time. Error bars show 95% confidence intervals. Black point denotes significant coefficients at the level of p < 0.05 and gray point denotes marginally significance at the level of p < 0.1. * p < 0.05; ‵p < 0.1.
- Figure 4.
Multiple regression results. A, Location of the significant cluster sensitive to the interaction effect between association and composition. Light red regions mark the language network within which the analysis was conducted. Color bar indicates t statistics. B, Time courses of β coefficients averaged over the significant cluster. Shaded region denotes the significant time window from 637 to 759 ms (p = 0.03). C, Fitted responses for each condition averaged over the significant cluster and time window. D, Time courses of fitted response for each condition averaged over the significant cluster. Word frequency effects were regressed out of the responses. E, Location of the significant cluster sensitive to the composition effect. F, Time courses of β coefficients averaged over the significant cluster. Shaded region indicates the significant time window from 689 to 882 ms (p = 0.001). G, Fitted responses for each condition averaged over the significant cluster and time window. H, Time courses of fitted response for each condition averaged over the cluster. Word frequency effects were regressed out of the responses. STG, superior temporal gyrus; MTG, middle temporal gyrus; ITG, inferior temporal gyrus; ATL, anterior temporal lobe. * p < 0.05.
- Figure 5.
Searchlight multivariate pattern classification results. A, Color-coded curves report classification time courses for the four linear SVM classifiers averaged over the significant cluster in the LATL. Horizontal lines above the curves mark a statistically significant time window from 740 to 775 ms for the high-association composition versus low-association composition classifier (p = 0.022). B, Classification time courses for the four linear SVM classifiers within the functional ROI for the composition effect identified by the regression analysis. Horizontal line above the curves marks a significant time window from 835 to 895 ms for the low-association composition versus low-association list classifier (p = 0.016). Significance over the chance level of 50% was assessed by a permutation t test with TFCE correction within the mask (light red region) and the analysis time window of 600–1200 ms. * p < 0.05.
- Figure 6.
Directed connectivity between the LATL and the LMTL. A, Contrast between the two directed connectivity measures show significant correlations from LATL to LMTL from ∼0 to 250 ms after the onset of the second word, with a delay from ∼150 to 450 ms (p = 0.03). B, No significant partial correlation coefficients matrix from LMTL to LATL.
- Figure 7.
Pair-wise t test results between each of the two-word conditions and the single word condition within the LATL and the LMTL fROIs. A, The LATL fROI. B, Time course of responses for each condition averaged over the cluster. Shaded region indicates the marginally significant time windows of 634–657 ms (p = 0.06) and significant time window of 1087–1171 ms for high-association composition > single (p = 0.03). C, Fitted responses for each condition averaged over the cluster and the time windows. D, The LMTL fROI. E, Time course of responses for each condition averaged over the cluster. Shaded region indicates the marginally significant time windows of 739–784 ms for low-association composition < single (p = 0.08). F, Fitted responses for each condition averaged over the cluster and the time window. Significance was determined by TFCE correction with 10,000 permutation. The testing time window is 600–1200 ms; *p < 0.05, ′p < 0.1.
- Figure 8.
Effect of lexical frequency. A, The significant cluster for the first word frequency. B, Time course of β coefficient of first word frequency averaged over the cluster. Shaded region indicates the significant time windows from 202 to 294 ms (p = 0.034, TFCE corrected with 10,000 permutation). The analysis time window is 0–600 ms. C, The significant cluster for the second word frequency. D, Time course of β coefficient of second word frequency averaged over the cluster. Shaded region indicates the significant time windows from 870 to 950 ms (p = 0.045, TFCE corrected with 10,000 permutation). The analysis time window is 600–1200 ms. *p < 0.05
- Figure 9.
Directed connectivity between the left SMA and the left M1. A, Partial correlation coefficients matrix from SMA to M1. B, Partial correlation coefficients matrix from M1 to SMA.
Extended Data
Extended Data Figure 1-1
Stimuli lists for experiment 1. Download Figure 1-1, TIF file.
Extended Data Figure 1-2
Stimuli lists for experiment 2. Download Figure 1-2, TIF file.
Extended Data Figure 1-3
Log frequency for all the words in experiments 1 and 2. Download Figure 1-3, TIF file.
