Twenty-five native Chinese college students (mean age = 21.68 years, SD = 1.82, 19 female) were collected in this study. All participants’ native language was Chinese, and their second language was English. They started to learn English at the mean age of 7 years (SD = 2.42) and by the time of the experiment had received formal education in the English language for 13 years (SD = 2.62). Their proficiency in English was self-reported on a 5-point questionnaire, which has been widely used in previous research on bilingualism (Jamal et al., 2012; Cao et al., 2017; Kachlicka et al., 2019). The mean scores of reading, writing, dictation and oral expression in English were 4.36 (SD = 0.70), 3.96 (SD = 0.73), 3.84 (SD = 0.94) and 3.84 (SD = 0.62), respectively. Therefore, all the participants were intermediate bilinguals, and had similar proficiency levels in their second language. They were unfamiliar with the Korean language participated in this study. To estimate the number of appropriate participants needed for our design, we used the G*Power analysis and found that 17 participants are sufficient for medium effect size (i.e., 0.25) with 0.80 power in one-way repeated-measures analysis of variance (ANOVA; Cohen, 1988, 1992). All participants’ vision or corrected vision was normal and had no related psychiatric history. They were all right-handed as measured by Snyder and Harris’s handedness inventory (Snyder and Harris, 1993). The above-mentioned inclusion and exclusion criteria were established before the data collection. Before the experiment, all participants signed an informed written consent and the procedures were approved by the IRB at the South China Normal University. No part of the study procedures and analyses was pre-registered prior to the research being conducted.

One group Chinese characters and two groups of artificial language characters were used in this study. Each group consisted of 16 characters. The artificial language characters were used for learning. The Chinese characters were used as a control for learning and consequently only used in the fMRI task. All characters are presented with 226 × 151 pixels in size on a gray background screen (see Figure 1B). Chinese characters had 2–3 units (M = 2.1), and 6–11 strokes (M = 8.5). All of them were medium-frequency characters (M = 16.9 per million, ranging from 10/million to 30/million; Cai and Brysbaert, 2010).

The artificial language characters used in this study were constructed using 16 Hangul letters (eight consonants and eight vowels). The consonants and vowels were then divided into two matched groups. Each group consisted of four consonants and four vowels, which were used to construct 16 artificial language characters (CVC) with 3 units and 6–11 strokes. In total, 32 artificial language characters were constructed. The two groups of artificial language characters (each group had 16 characters) were assigned into the two learning conditions (i.e., the addressed-phonology and assembled-phonology learning) as described in Training Procedure blow and they did not differ in the number of strokes [the first group: M = 8.50; the second group: M = 8.25; t(15) = 0.37, p = 0.72].

The sounds of artificial language letters and characters were initially recorded from one female college students. Adobe Audition CS6 software was used to denoise the speech materials, and normalize them to the same volume and length. The sound of each character and letter lasted for 800 and 600 ms, respectively.

To evaluate participants’ preferred routes in reading, we used two fluency tests to assess the reading fluency of the irregular words and pseudowords. The test for pseudowords was adopted from the Phonemic Decoding Efficiency subtest of TOWRE (Test of Word Reading Efficiency; Torgesen et al., 1999), which consisted of 63 phonemically regular pseudowords. The test of irregular words, consisting of 104 phonemically irregular words (e.g., whose, hour), was developed by the authors in this study. The number of irregular words was determined according to the Sight Word Efficiency subtest of TOWRE, which consisted of 104 real words. The irregular words used in this study have at least one grapheme whose phoneme is not the most common phoneme for that grapheme. Following previous studies (Berndt et al., 1987; Gontijo et al., 2003), we also quantified the pronunciation regularity by calculating the probabilities of graphemes being pronounced as their corresponding phonemes. The mean pronunciation regularity was 0.65 (SD = 0.14), suggesting that they were phonemically irregular words. In both tests, participants were asked to read words/pseudowords one by one as quickly and accurately as possible. Following the instruction of TOWRE, they were scored by using the number of correctly pronounced items within 45 s (see Figure 2A).

Given previous findings that reading irregular words and regular pseudowords, respectively, rely on the lexical pathway (depending on whole-word mapping) and nonlexical pathway (depending on grapheme-to-phoneme mapping), the difference in standardized scores between irregular words and pseudowords was used as a reading strategy index (Destoky et al., 2020). The positive reading strategy index indicates greater reliance on whole-word mapping during word reading, the negative reading strategy index indicates the tendency to use grapheme-to-phoneme mapping.

All participants received artificial language training for 5 days, 1 h per day. Participants were asked to learn the associations between visual forms and sounds of 32 artificial language characters. As mentioned before, the characters were divided into two groups (each had 16 characters) for the two training conditions: addressed-phonology and assembled-phonology training (see Figure 1A). In the addressed-phonology training condition, participants were asked to associate the whole character with its sound by rote memorization. The correspondence between the characters and their original sounds was randomly shuffled to ensure that participants could not implicitly acquire the pronunciations of characters through the grapheme-to-phoneme correspondence (GPC) rules. In the assembled-phonology training condition, participants were instructed to assemble the pronunciations of the whole characters via the GPC rules. They learned the pronunciations of both Hangul letters and the whole characters in the assembled-phonology training condition. The assignment of the two groups of characters into the two learning conditions was counterbalanced across participants to eliminate the effects of materials on learning. In each training session, participants spent the same time in learning the two conditions, with about half an hour for each condition.

As in previous studies (Mei et al., 2015; Li et al., 2019), several learning tasks were designed to facilitate the acquisition of the artificial language characters. The tasks included character learning (i.e., learning the character-sound associations one by one), free learning (i.e., relearning the characters that participants had difficulties), phonological choice (i.e., choosing the correct sound for the target character from four sounds), naming with feedback (i.e., reading each character aloud followed by its correct pronunciation), and fast naming (i.e., reading 10 words as fast and accurately as possible).

After the 5-day artificial language training, participants were scanned once while performing an overt naming task. One group of Chinese characters and two groups artificial language characters (one group in the addressed-phonology learning and the other group in the assembled-phonology learning) were used for the fMRI task. Each group had 16 characters. Those characters were pseudo-randomly presented with the sequence was optimized by using OPTSEQ2.¹ The inter-trial intervals were not jittered because this study used a slow event-related design which was able to precisely estimate single-trial activation patterns. There were three runs in total. Each run contained 48 trials with 16 characters in each condition presented once. Each trial lasted for 12 s. Specifically, it started with a 1 s fixation, followed by a character presented on the screen for 3 s. Participants were asked to read each word as quickly and accurately as possible. Once the character disappeared, participants were asked to perform a visual orientation judgment task for 8 s to prevent further rehearsing for the character and to keep participants undergoing the same cognitive processes after the character (Xue et al., 2013; Zhao et al., 2017; Qu et al., 2021). To make this task engaging, a self-paced procedure was used. Specifically, a Gabor image presented on the screen was tilted 45° either to the left or to the right of vertical. Participants were asked to judge the orientation of the Gabor. The Gabor image disappeared once participants responded and the next image was presented after a blank of 100 ms (see Figure 1B).

Due to the large amount of noise during scanning, participants’ oral responses were recorded in another session of the word naming task outside the scanner. The stimuli sequence and parameters used were identical with the fMRI task.

The fMRI data were collected using a 3.0 T Siemens MRI scanner in the MRI Center at South China Normal University. A single-shot T2*-weighted gradient-echo EPI sequence was used to collect the functional imaging data. The scanning parameters were used as follows: TR = 2,000 ms, TE = 30 ms, flip angle = 90°, FOV = 224 × 224 mm, matrix size = 64 × 64, slice thickness = 3.5 mm, and number of slices = 32. A T1-weighted, gradient-echo pulse sequence was used to collect anatomical data. The parameters were as follows: TR = 1,900 ms, TE = 2.52 ms, flip angle = 9°, FOV = 256 × 256 mm, matrix size = 256 × 256, slice thickness = 1 mm, and number of slices = 176.

The imaging data were processed using FEAT (FMRI Expert Analysis Tool) Version 6.00. The first three images in each run were discarded to allow for T1 equilibrium effects. The remaining functional images were realigned and normalized to the MNI (Montreal Neurological Institute) template (Jenkinson and Smith, 2001). They were smoothed using a 5 mm full-width at half-maximum Gaussian kernel and temporally filtered using a nonlinear high-pass filter with a 60 s cut-off. The head movement was no more than one voxel for any run or participant.

We then modeled the data using the general linear model. In each run, there were 48 separate regressors corresponding to 48 characters. The fixation and the orientation task were not explicitly modeled and served as implicit baseline. A canonical hemodynamic response function (double-gamma) was used to convolve with the onsets and durations of events. Following previous studies (Xue et al., 2010; Mumford et al., 2012; Dong et al., 2021), the single-trial neural response was estimated using least squares estimation and ridge regression. Six motion parameters were used as covariates to improve statistical sensitivity. Contrast images of the three conditions (i.e., Chinese characters, addressed characters, and assembled characters) and their comparisons were calculated for each participant and each run.

We then used a fixed-effects model to concatenate data from the three runs for each participant. After that, a random-effects model with FLAME Stage 1 was used to obtain group activations. All reported results were thresholded with a height threshold of Z > 2.6 (i.e., p < 0.005) and a cluster probability of p < 0.05, and the Gaussian random field theory was used to perform multiple comparison corrections for the whole brain (Worsley, 2010).

To investigate differences in activation pattern between assembled and addressed characters, we conducted the multivariate pattern analysis (i.e., MVPA) using the CoSMoMVPA.² Beta-estimates of unsmoothed data in each run were extracted and then were used to train and test the classifiers (Mur et al., 2009; Pereira et al., 2009). To implement all classifications, a linear support vector machine (SVM) classifier with a regularization parameter (c = 1) was used. A leave-one-run-out cross-validation protocol was used to calculate the classification accuracy. Specifically, we trained a classifier to distinguish characters in addressed-phonology learning versus those in assembled-phonology learning in two runs, and then tested the classifier’s ability in the left-out run. The classification accuracies were calculated as the means of the three classification results.

In the above analysis, a searchlight-based method was used. Specifically, a spherical searchlight comprised 100 voxels was moved within the whole brain. The classification accuracy (proportion of correctly classified trials) was mapped back upon the sphere’s central voxel to produce accuracy maps. The analysis was conducted in each participant’s native space and then transformed into standard space to form a 4-D map. Finally, a t-map was generated at the group level to test whether the classification accuracy of each voxel was significantly higher than the chance level (0.5). Based on the performance of those predictions, we could find out which brain regions showed differences in neural pattern between assembled and addressed characters. The group images were thresholded at p < 0.005 (one-tailed), and a cluster probability of p < 0.05.

To examine individual differences in the recruitment of the two pathways for phonological access (i.e., assembled and addressed characters), we correlated activation differences between assembled and addressed characters in the regions identified in MVPA with reading strategy index. Specifically, the abovementioned MVPA revealed nine brain regions showing different neural codes between assembled and addressed characters. The brain regions included the paracingulate gyrus (PCG) and posterior cingulate gyrus (PCC) in the middle line of the brain, the bilateral pars trianguris (PT), orbitofrontal cortex (OFC), middle temporal gyrus (MTG), and the left precuneus cortex (PCun; see Figure 3C). These nine brain regions were functionally defined as regions of interest (ROI) in the correlation analysis. Each ROI was defined as a 6 mm diameter sphere around the local maxima in each cluster. In each ROI, percent signal changes were extracted separately for assembled and addressed characters (Mumford, 2007). We then correlated activation difference between addressed and assembled conditions (i.e., percent signal changes of the addressed condition minus those of the assembled condition) in each ROI with reading strategy index. This analysis would help to reveal the associations between individuals’ preferred reading strategies and the involvement of a certain region in phonological access.

To explore the effects of individuals’ preferred reading strategies on learning, we additionally calculated the correlations between reading strategy index and behavioral differences between addressed and assembled characters (i.e., difference in Z scores of reaction time or accuracy between assembled and addressed characters) during the word naming task after the fMRI scan. The Z scores were separately calculated for the addressed and assembled characters. An individual with greater difference in Z scores of accuracy or with smaller difference in Z scores of reaction time indicates that he is better at addressed-phonology learning relative to assembled-phonology learning than other individuals. Therefore, if individuals’ reading pathway preference affects phonological learning in a new language, reading strategy index would positively correlate with difference in Z scores of accuracy, but negatively correlate with difference in Z scores of reaction time.

One-way repeated measures ANOVA was used to examine behavioral differences in reaction time and accuracy (see Figure 1C). The main effects of word type were significant for both reaction time (F(2, 48) = 328.62, p < 0.001) and accuracy (F(2, 48) = 40.10, p < 0.001). Post hoc comparisons showed that Chinese characters (797.51 ms) were named faster than assembled characters (1402.77 ms; p < 0.001), which were named faster than addressed characters (1548.13 ms; p < 0.001). Consistently, Chinese characters (100.00%) showed higher accuracy than assembled characters (97.92%; p < 0.001), which showed higher accuracy than addressed characters (90.92%; p < 0.001). These results indicate that participants were more familiar with characters in their native language (i.e., Chinese characters) than newly-acquired characters (i.e., assembled and addressed characters). Furthermore, participants performed better on assembled characters relative to addressed characters as the former followed GPC rules. This result is consistent with previous findings that assembled phonology is easier to learn than addressed phonology even for Chinese speakers with much experience on addressed phonology (Hooper, 2001; Aro and Wimmer, 2003; Ellis et al., 2004; Mei et al., 2014, 2015).

To examine the effects of individuals’ preferred reading strategies on character learning, we further calculated the correlations between reading strategy index and behavioral differences across the two types of characters (i.e., addressed and assembled characters) after scan. Results showed that reading strategy index was significantly correlated with differences in accuracy between the two types of characters (r = 0.419, p = 0.037; see Figure 2B). The correlation for reaction time was not significant (r = 0.031, p = 0.884; see Figure 2C). These results indicate that the more lexical pathways participants preferred, the better they performed on newly-acquired addressed characters relative to assembled characters.

Whole-brain activation analysis showed that Chinese characters, addressed characters, and assembled characters had similar activations in an extensive network, including the anterior cingulate cortex, supplementary motor cortex, bilateral prefrontal cortex, occipitoparietal cortex, and occipitotemporal cortex (see Figures 4A–C; Supplementary Table S1). Moreover, direct comparisons between Chinese characters and artificial language characters revealed that, compared with Chinese characters, assembled characters elicited greater activations in the bilateral precentral gyrus, temporoparietal cortex, fusiform gyrus and left inferior frontal gyrus (see Figure 4D; Supplementary Table S2), and addressed characters elicited stronger activations in the paracingulate gyrus, bilateral middle frontal gyrus, fusiform gyrus and left supramarginal gyrus (see Figure 4E; Supplementary Table S2).

Further comparisons between the two types of artificial language characters showed that assembled characters induced greater activations in the right supramarginal gyrus (see Figure 3A; Supplementary Table S3), whereas addressed characters evoked stronger activations in the paracingulate gyrus, bilateral middle frontal gyrus, inferior frontal gyrus, orbitofrontal cortex, precuneus cortex, lingual gyrus and fusiform gyrus (see Figure 3B; Supplementary Table S3).

A searchlight-based MVPA was used to investigate the differences in activation pattern between assembled and addressed characters. In this analysis, we trained and tested classifiers to distinguish assembled versus addressed characters. Results showed that nine brain regions showed significantly higher classification accuracy than the chance level (i.e., 50%), including the paracingulate gyrus (PCG), posterior cingulate gyrus (PCC), bilateral pars trianguris (PT), orbitofrontal cortex (OFC), middle temporal gyrus (MTG), and the left precuneus cortex (PCun; see Figure 3C; Supplementary Table S4). These results are generally consistent with brain regions reported in the activation analysis above.

To examine the associations between individuals’ preferred reading strategies and the involvement of a certain region in phonological access, we conducted correlation analysis on reading strategy index and the activation differences between addressed (relying on whole-word mapping) and assembled (relying on grapheme-to-phoneme mapping) characters in the nine regions identified in MVPA. The activation differences between addressed and assembled characters were computed by subtracting percent signal changes of assembled characters from those of addressed characters. Significant correlations were found in 4 regions, including the PCG (r = −0.52, p < 0.01), bilateral OFC (left: r = −0.51, p < 0.01; right: r = −0.54, p < 0.01), and right PT (r = −0.42, p < 0.05; see Figure 5). The correlation coefficients remained significant in the PCG (p = 0.027) and bilateral OFC (left: p = 0.027; right: p = 0.036) after false discovery rate (FDR) correction (Benjamini and Hochberg, 1995; Qu et al., 2019). It is worth noting that there is an outlier value in the left PT. After excluding that outlier, the correlation coefficients became significant (r = −0.42, p < 0.05). These results indicate that participants who preferred lexical pathway in phonological access show less involvement of brain regions for addressed phonology in the processing of newly-acquired addressed characters.

The involvement of the above five brain regions (i.e., PCG, bilateral OFC, and PT) in addressed phonology were further confirmed by comparison between activations of addressed and assembled characters. Results showed that all the regions showed greater activations for addressed characters than assembled character [the left OFC: t(24) = 4.44, p < 0.001; the right OFC: t(24) = 6.11, p < 0.001; the left PT: t(24) = 3.93, p < 0.01; the right PT: t(24) = 4.78, p < 0.001; and the PCG: t(24) = 7.93, p < 0.001] (see Figure 6).