In foreign language learning (FLL) research, older adults have been an understudied population (Mackey and Sachs, 2012), which may be caused by the increased difficulty in picking up a new language after the “critical period” (Lenneberg, 1967; Hartshorne et al., 2018, see also Wang, 2018). However, there has been a recent change in attitude, due in part to a series of pioneering studies on the effect of lifelong bilingualism on brain structures and cognitive functions (Bialystok et al., 2007; Luk et al., 2011; Costa and Sebastián-Gallés, 2014; Olsen et al., 2015) and on the neuroplasticity induced by intensive FLL in younger adults (Wong and Perrachione, 2007; Mårtensson et al., 2012; Stein et al., 2012; Zatorre, 2013; Qi et al., 2019).

Relative to older monolinguals, older bilinguals were reported to have greater protection against dementia, better inhibition function, and improved executive control (Bialystok et al., 2007; van den Noort et al., 2019). The bilingual advantage has also been reported for other cognitive domains, e.g., episodic and verbal memory (Schroeder and Marian, 2012; Grant et al., 2014, but see Paap et al., 2015). Neurally, the correlates of bilingual experiences have been studied using many anatomical segmentation tools (Li et al., 2014; Hämäläinen et al., 2018; Maschio et al., 2019). Using a technique known as surface-based morphometry (SBM; Fischl and Dale, 2000; Winkler et al., 2010; Luders et al., 2012), early acquisition of two languages is found to be associated with larger surface area in the left pars opercularis of the inferior frontal gyrus (IFG) and right superior temporal gyrus (STG), while late acquisition is associated with increased mean curvature in the left STG (Hämäläinen et al., 2018). Bilingualism-induced neural differences have also been demonstrated using related anatomical measures such as sub-cortical volumes (DeLuca et al., 2019b) or gray matter density (Abutalebi et al., 2015), obtained by volume-based morphometry and voxel-based morphometry. For example, an expansion of the thalamus was observed for individuals with greater immersion in the second language (Pliatsikas et al., 2017; Deluca et al., 2019a).

As for learning-induced neuroplasticity in younger adults, five months of intensive FLL was sufficient to bring about increases in gray matter volume in the left IFG and left anterior temporal lobe (Stein et al., 2012). Intensive vocabulary learning in young interpreters was found to promote increases in cortical thickness across various regions in the left-lateralized language network, including left middle frontal gyrus, IFG, and STG, along with increased volume in the bilateral hippocampus (Mårtensson et al., 2012). Consistent with the classroom-based learning above, a recent experimental work based on either the picture–word association or virtual reality as the FLL paradigm revealed both structural and functional changes in younger adults, including increased cortical thickness over both the core language areas in the left hemisphere and many regions in the right hemisphere (Legault et al., 2019a). In another study, increased brain volume was found in the right hippocampus (Bellander et al., 2016).

Taken together, the findings on the bilingual advantage and learning-induced neuroplasticity have raised the possibility to use FLL as a non-pharmacological means for promoting successful aging in older adults (Antoniou et al., 2013; Antoniou and Wright, 2017; Ware et al., 2017; Klimova, 2018; Pfenninger and Polz, 2018; Klimova and Pikhart, 2020). Consistent with this conjecture, a few studies reported cognitive improvements in older learners after intensive FLL (Bak et al., 2016; Bubbico et al., 2019; Wong et al., 2019). For example, improvement in global cognition was found following six months of individual computer-based cognitively stimulating activities, including FLL and cognitive games (Wong et al., 2019). In another study, older learners also showed significant improvement in global cognition, along with increased resting-state functional connectivities (RSFCs) between posterior cingulate cortex and three right-hemispheric regions, including right IFG, right superior frontal gyrus, and left superior parietal lobule (Bubbico et al., 2019). This interesting finding suggests that the right-hemisphere homologs of the left-hemisphere language areas may provide “scaffolding” in the early phase of second language learning, before a right-to-left shift occurs (Yang et al., 2015; Qi and Legault, 2020). Despite the positive findings above, results have been mixed, with a few studies reporting no improvement for all cognitive skills examined (Ware et al., 2017; Berggren et al., 2020), some degree of enhancement that was however not significantly different from that of a control group (Valis et al., 2019), or neither learning-induced anatomical nor behavioral changes (Nilsson et al., 2021).

To account for the neuroplastic changes or cognitive gains that could be driven by FLL in older adults, an understanding about the neurocognitive bases of FLL is essential, because the learning-induced changes in a brain region likely arise due to its increased recruitment over the learning programme (Wong et al., 2007). Toward this end, some prominent neuroanatomical models of language (e.g., Ullman, 2001a, 2004; Hickok and Poeppel, 2007) have provided the guiding theoretical framework. In the declarative–procedural model (DP model; Ullman, 2001a, 2004), the declarative memory (comprising semantic and episodic memory), as subserved by a cortical–hippocampal system (Mcclelland et al., 1995; Eichenbaum, 2000), is posited to be involved for learning words. In contrast, the procedural memory, as subserved by the insula and basal ganglia (especially the striatum), is recruited in phonological and rule-based grammar learning. Functional activation and connectivities studies have been central to the quest for the neurocognitive bases of FLL (e.g., Mestres-Missé et al., 2008; Abutalebi and Green, 2016; Bakker-Marshall et al., 2021, for a recent review, see Qi and Legault, 2020). However, one source of information that could not be understated is individual differences—the focus of the present article. Traditionally, they are sometimes considered as “noise” that may obscure the effect of interest. Nonetheless, they have proved reliable for linking brain functions and cognitive behavior (e.g., Kanai and Rees, 2011; Chiarello et al., 2013).

In FLL research, individual differences in brain functions are posited to put strong constraints on the learning success (Zatorre, 2013; Li and Grant, 2016). Previous behavioral and neural findings have suggested that phonological skills, such as phonological short-term memory (PSTM) and phonological awareness, are the determining factors of language learning success (Papagno et al., 1991; Baddeley et al., 1998; Koda, 1998; Linck et al., 2014; Gillon, 2017). Corroborating evidence has been found in both classroom-based and laboratory-based studies with younger learners, for phonemic learning (Golestani and Zatorre, 2009; Silbert et al., 2015; Fuhrmeister and Myers, 2021), lexical tone learning (Wong and Perrachione, 2007; Chandrasekaran et al., 2010), and artificial grammar learning (Yang and Li, 2012; Kepinska et al., 2017). In agreement with the view that PSTM plays a crucial role in FLL, functional connectivity or activation studies have reported an association between FLL success and regions important for phonological processing, e.g., left precentral gyrus (Veroude et al., 2010), left supramarginal gyrus (SMG; Veroude et al., 2010; Sliwinska et al., 2012), and left posterior STG (e.g., Wong et al., 2007). Also, consistent with the hypothesis that phonological awareness plays a critical role in language learning success, individual variation in learning outcomes can be predicted by individual variation in baseline sensitivity to non-lexical pitch patterns (Wong et al., 2007).

Of special interest to the present study are the anatomical associations with learning performance. For example, the success in linguistic pitch learning or direction discrimination were predicted by the volume of left Mescal's gyrus (Wong et al., 2008) and by the cortical thickness of the right hemispheric homolog of the Broca's area (Novén et al., 2019). In the study by Legault et al. (2019a), reviewed in the last section, the FLL performance was associated with the cortical thickness of right IFG in the picture–word group, and with the cortical thickness of right inferior parietal lobe in the VR group. The emphasis of phonological skills has extended to sub-cortical regions. Recent empirical works (Legault et al., 2019a) and theoretical papers (Abutalebi and Green, 2007; Li et al., 2014) have highlighted the roles of the striatum—comprising caudate nucleus and putamen—in phonological learning. For example, the caudate nucleus is involved in phonemic fluency (rather than semantic fluency; Grogan et al., 2009) and procedural memory (Ullman, 2004), while the putamen has been implicated in articulatory planning and detecting phonological errors (e.g., Abutalebi and Green, 2007).

While the emphasis on phonological skills in previous studies is understandable, in light of the DP model, it is surprising that the roles played by semantic and verbal episodic memory have seldom been tested in neural research on FLL. Since the DP model was initially proposed, functional imaging and neurological studies have established that semantic functions are supported by a widespread network that includes the ventrolateral prefrontal cortex, anterior temporal lobe, middle temporal gyrus, etc. (Patterson et al., 2007; Binder et al., 2009; Ralph et al., 2017), along with sub-cortical regions like the thalamus. In addition to being a main relay station of sensory information, the thalamus may also play an important role in language production by selecting lexical and semantic representations (Abutalebi and Green, 2016), by virtue of its connection with the left IFG (Ford et al., 2013). The episodic memory is subserved by the hippocampus and its surrounding medial temporal lobe structures (including the entorhinal cortex and the parahippocampal gyrus), along with the prefrontal cortex (e.g., Breitenstein et al., 2005). The hippocampus is well-known for its role in transforming short-term memory to long-term memory, while the entorhinal cortex and the parahippocampal gyrus are implicated during the processing of object/event information and spatial-temporal context, respectively (Eichenbaum et al., 2012). Another study also reported that the entorhinal cortex reflects elementary memory processes related to novelty detection, while the parahippocampus is more involved in the formation and subsequent reactivation of the memory (Daselaar et al., 2004). A few recent studies have reported evidence that supports a role of hippocampus in younger adults in FLL (Kepinska et al., 2017, 2018). For example, the connectivity of the hippocampus and Broca's area was implicated during the acquisition of a novel grammar (Kepinska et al., 2018).

There are reasons that the brain networks and cognitive mechanisms could be different in older learners, which could result in different learning outcomes (Service and Craik, 1993; Glass et al., 2012; Ingvalson et al., 2017). To start with, older adults had poorer speech discrimination for phonemes and pitch (Shen et al., 2016), which may promote the use of alternative strategies in vocabulary learning. Due to lower working memory capacity (Service and Craik, 1993; Mackey and Sachs, 2012), they may also rely less on rote memorization. In addition, the inhibition function is generally in decline (Hasher et al., 1991, but see Veríssimo et al., 2021), making it difficult for them to retrieve the L2 word because it is more difficult for them to suppress the L1 word. In contrast, older adults often have better crystallized knowledge (e.g., world and vocabulary knowledge) than younger adults (Salthouse, 2012; Hartshorne and Germine, 2015). It is plausible that declarative memory may play a crucial role in older learners (Ullman, 2001b). For example, they may take advantage of their better semantic knowledge to remember the vocabularies as a chunk (e.g., by forming a story to connect the words being learnt), or to proactively derive semantic associations for use as retrieval cues in episodic memory recall. However, individual differences studies on FLL in older adults have remained scarce, and even fewer have examined the neural correlates of FLL in older adults. To our knowledge, one such study has focused on vocabulary learning, and it suggested that the hippocampal volume and the associative memory ability prior to language learning are robust predictors of vocabulary proficiency at the end of training (Nilsson et al., 2021). Another study showed that the artificial grammar learning performance was influenced by structural and functional connectivity of the Broca's area and its right hemisphere homolog (Antonenko et al., 2012).

The present study investigates the neurocognitive factors that influence the immediate learning success and long-term retention of the vocabularies during an intensive FLL programme, based on a group of Cantonese-speaking Hong Kong older learners aged 58–69, all of whom were familiar with English as a second language. Italian was chosen as the target language due to its highly regular grapheme–phoneme correspondence, its use of the Latin alphabet, the unfamiliarity with the language by the general population in Hong Kong, and the popularity of Italy as a tourist destination. By design, while both vocabulary and basic grammar rules were taught throughout the learning programme, a stronger emphasis was put on the vocabulary since it plays a fundamental role in both spoken and written language comprehension across all stages of learning (Yum et al., 2014). Also, older learners may be less motivated in learning a new language in its entirety, and they may only wish to learn the essential vocabularies for simple communications and travels (Antoniou et al., 2013; Pfenninger and Polz, 2018). Our main hypothesis is that, for foreign vocabulary learning in older adults, both semantic and episodic memory are more strongly associated with the learning success than phonological skills. To examine this hypothesis, two groups of analyses were conducted, using MRI and cognitive data to predict the immediate learning success and long-term retention.

The first group of analyses (MRI analysis) comprised a cortical analysis and a complementary sub-cortical analysis. For the cortical analysis, using SBM, four cortical morphological measures (cortical thickness, cortical surface area, cortical volume, and mean curvature) were extracted over a broad range of regions of interest (ROIs); these measures were then used as predictors to test the associations of each ROI with the in-class performances and the final test scores. The ROIs were selected because they have consistently been found to support semantic memory (Patterson et al., 2007; Lau et al., 2008; Binder et al., 2009; Ralph et al., 2017), have exhibited learning-induced neuroplasticity (Lee et al., 2007; Stein et al., 2012; Li et al., 2014), have previously been associated with second language learning performance in general (Ullman, 2016; Tagarelli et al., 2019), or are part of the language control network (Abutalebi and Green, 2007). Although language processing is strongly left-lateralized in the brain, the right hemisphere is also heavily engaged for learning a second language (Hosoda et al., 2013; Bubbico et al., 2019; Qi et al., 2019; Chen et al., 2021). Thus, the homologous areas in both hemispheres were included as ROIs. Our hypothesis would be supported if the cortical measures of the brain regions centrally involved in semantic functions (pars orbitalis and temporal pole) or episodic memory (caudal middle frontal cortex and entorhinal cortex) are more associated with learning performance than those that underlie phonological functions (precentral gyrus and SMG).

For the sub-cortical analysis, sub-cortical volume measures were used to predict in-class and final test scores. Due to our relatively small sample size and the already extended set of cortical regions examined, exploration was limited to four ROIs: hippocampus, thalamus, caudate nucleus, and putamen. This analysis complemented the cortical analysis in providing further evidence consistent with the main hypothesis. For example, if the hippocampal volume was associated with learning performance, it would strengthen the view that the episodic memory is associated with FLL in older adults.

In the second group of analyses (cognitive analysis), a broad range of cognitive and phonological measures, derived from a cognitive and phonological test battery, were used as the predictors of language learning performance. Behavioral support for our main hypothesis would be obtained if the semantic function score (semantic fluency and picture naming) and verbal episodic memory score (Hong Kong List Learning test) are more important predictors of the learning outcomes than phonological function scores (phonological discrimination and awareness).

A group of 25 older learners aged 58–69 were initially enrolled into the Italian learning programme. They were native Cantonese speakers, had no known neurological disorders and normal/correct-to-normal vision. All of them had limited exposure to Italian, had visited Italy for no more than a month, and had at least 6 years of prior experience in learning English. The cohort was recruited via the Institute of Active Aging, Hong Kong Polytechnic University. Hong Kong version of Montreal Cognitive Assessment (Wong et al., 2009) was used to confirm that they were cognitively normal (mean = 27.9, SD = 1.7). The attendance for the intensive Italian learning programme was high, with 21 of 25 participants completing all the 21 possible visits. There were four drop-outs midway through the programme, including three who completed 5 visits and one who completed 10 visits. One female participant was left-handed (handedness = −60) according to the Edinburgh inventory (Oldfield, 1971); only the remaining 20 participants were included in the present analysis. Their mean age, education, and HK-MoCA were 63.7 (SD = 2.9, range = 58.6–69.3), 15.2 (SD = 3.1, range = 11–21), and 27.8 (SD = 1.7, range = 25–30), respectively. All procedures were approved by the Ethical Review Committee, Hong Kong Polytechnic University. Written signed informed consent was obtained from all participants. They were reimbursed HKD 800 for their participation.

The participants individually attended an intensive Italian vocabulary learning programme (Figure 1). The whole programme comprised an introductory lesson (visit 1), ten computerized vocabulary lessons (visits 3–4, 6–7, 9–10, 12–13, and 15–16), four revisions sessions (visit 8, 11, 14, 17), an Italian final test (visit 18), and a range of tests arranged at various points of the programme. These included an MRI session (visit 2), initial-learning phase cognitive tests (visit 5), initial-learning phase phonological tests (at the beginning of visits 3, 4, and 6), post-learning phase phonological tests (visits 18–19), post-learning phase cognitive tests (visit 20), and a debriefing session (visit 21; see Supplementary Section S1). The MRI, phonological, and cognitive testing were not conducted right at the beginning, for two reasons. (1) In reality, many language learners have some degree of knowledge of the target foreign language before proceeding to take a formal course. The introductory session serves to provide a more uniform experience prior to the MRI visit. Also, the introductory session gave them a taste of learning a new foreign language and offered a buffer period for them to commit to participating in the structural MRI session and the subsequent intensive Italian learning programme, which would likely reduce the dropout rate. (2) Upon enrolling to the study, most participants were anxious and most motivated to start learning Italian. As such, arranging too many tests that are unrelated to Italian learning right at the beginning would likely be seen as non-ideal arrangement from the learner's view, which might increase the dropout rate. Hence, the phonological and cognitive testing were not arranged immediately after the MRI session, but they were spread over sessions 3–5 as a compromise, especially because it was not our primary goals to investigate learning-induced cognitive gains. Overall, the whole programme can be divided into three phases—initial learning (visits 1–5), intensive-learning (visits 6–17), and post-learning (visits 18–21). The division of visits 1–17 into two phases was an arbitrary one, but the main difference was that a variety of tests were throughout the initial learning stage, while a revision session was arranged every two lessons only during the intensive-learning phases. Also, there was more separation between visits in the initial-learning phase (4.5 days on average, SD = 1.1 days) than in intensive-learning (3.2 days, SD = 0.7 days), due to scheduling difficulties. The completion time of the first twenty visits of the programme was 68.9 days on average (SD = 11.6 days). The details for each component are laid out below in the following sections: Italian vocabulary learning programme, MRI protocol, and behavioral test battery.

Linear mixed-effects (LME) models were constructed to test our main hypothesis that semantic and episodic memory functions play a major role in language learning in older learners. First, baseline demographic models (baseline models thereafter) were constructed for each of the three scores (in-class Ita-to-Can, in-class Can-to-Ita, and final test). In each baseline model, in addition to either Week (for the in-class Ita-to-Can or Can-to-Ita) or TestPart (for the final test), three demographic variables (age, years of education, and gender) were included as fixed-effects predictors, while participants were included as a random-effects variable.

Next, two groups of analyses were conducted, each involving comparisons between the full model(s) of interest against the corresponding baseline models. The first analysis comprised both a cortical and a complementary sub-cortical analysis. In the cortical analysis, the surface-based morphological measures were used as predictors to test the neural associations with the in-class performances and the final test scores. The regions of interest (ROIs), illustrated in Figure 2, comprised the Broca's area (pars opercularis and pars triangularis), the regions supporting semantic memory (pars orbitalis, temporal pole, and middle temporal gyrus) and episodic memory (caudal and rostral middle frontal cortex, entorhinal cortex, and parahippocampal gyrus), as well as regions important for phonological processing (SMG, precentral gyrus, banks of superior temporal sulcus), procedural memory (insula), language switching (caudal and rostral anterior cingulate), and orthographical processing (fusiform gyrus). For each ROI, a full model was constructed that include all the predictors in the baseline model, four anatomical measures of the ROI, and the interactions between each anatomical measure with either Week or TestPart. Likelihood ratio tests were used to evaluate whether the full model of each ROI represented an improvement over the baseline model. Multiple comparisons were corrected using the false discovery rate (FDR) procedure. If a significant interaction was found, post hoc trend analysis was conducted to reveal how the degree of association of each anatomical measure with the language learning performance was modulated by Week or TestPart. In the sub-cortical analysis, the sub-cortical volumes were used to predict the in-class and final test scores. The eight regions of interest (ROIs) included the hippocampus, thalamus, caudate nucleus, and putamen on both hemispheres. For each ROI, a full model was constructed using the sub-cortical volume of the ROI and its interaction with either Week or TestPart. The subsequent likelihood ratio tests were FDR-corrected.

In the second group of analyses—cognitive analysis—the cognitive predictors, derived from the neuropsychological data, were included as the additional predictors for modeling the three scores (Bonferroni correction was applied). For each model, backward elimination was conducted to yield the final model.

For completeness, we tested the changes in cognitive functions and phonological skills incurred between the initial- and post-learning, despite the lack of a control group to assess whether the changes were due to repetition effects or induced by language learning. For each cognitive/phonological measure, a pairwise Wilcoxon signed rank test (two-tailed) was conducted with FDR-correction.

Table 2 shows the summary statistics for the mean performance for the two in-class quizzes (Ita-to-Can and Can-to-Ita) and the final test, which were used as indices for immediate learning success and language retention, respectively. Figure 3 illustrates the overall in-class learning performance across the five weeks; the verb and noun components were separately plotted to reflect the structure of the in-class quizzes, although only the overall learning performance was selected for statistical analysis. To test whether the improvement was significant across the five weeks, a linear mixed-effects model was fitted for each in-class quiz, using the demographic variables for predicting the performance (Table 3). In both cases, the main effect of Week was significant, ps < 0.001. For Ita-to-Can, post hoc pairwise comparisons with Tukey's correction showed that the performance for week 1 was significantly worse than all other weeks, while that for week 5 was significantly better than all other weeks, with no differences among Week 2–4 being significant. For Can-to-Ita, the post hoc pairwise comparisons indicated that the improvements over two weeks or more were all significant, although the improvement between successive weeks was only significant going from Week 1 to 2. These results suggested that the participants were generally able to learn an increasing number of Italian words across the five weeks.

For the final test, the participants were able to dictate 59.5% of the Italian words, backward translate 80.2% of Italian words into Cantonese verbally, forward translate 56.4% of Cantonese words into Italian in written form, and achieve a score of 92.9% for grammar. It is worth noting that the first three parts were closed-note tests while participants were allowed to refer to the notes in the grammar part, which likely accounts for the higher grammar score. In the baseline model, the main effect of TestPart was significant, p < 0.001. Post hoc pairwise comparisons showed that the test score significantly differed across the four parts, except between dictation and Can-to-Ita.

Pairwise Wilcoxon signed rank tests revealed marginally shorter picture naming latency, p = 0.057, with an effect size (King et al., 2011) of r_c = −0.67, marginally better Raven score (p = 0.057, r_c = 0.67) and HKLLT (p = 0.094, r_c = 0.67). There were also marginally higher Italian discrimination score, p = 0.057, r_c = 0.67, as well as higher scores in the two Cantonese tests: CanDiscrim, p = 0.094, r_c = 0.57, and CanSpoon, p = 0.094, r_c = 0.62 (Table 7). However, due to the lack of a control group, it is unclear whether the change arose due to repetition effect or it was truly induced by intensive vocabulary learning.

Taking an individual-differences approach, the present study used both anatomical and cognitive measurements to investigate the basis of FLL in older adults, focusing on whether semantic and episodic memory play an especially important role in vocabulary learning within this special group of language learners.

In support of the hypothesis that semantic functions are important for older language learners, the left pars orbitalis, which is central to semantic control (Sabb et al., 2007; Binder et al., 2009; Ralph et al., 2017) by virtue of its white matter pathway (uncinate fasciculus) to the semantic hub in the temporal pole (Harvey et al., 2013), was found to be associated with the performance in the in-class quiz and the final test. In particular, its adjusted curvature was positively associated with the in-class Can-to-Ita performance while its adjusted volume was significantly associated with the first three parts of the final test (dictation, Ita-to-Can, and Can-to-Ita). This suggests that high-performing learners tend to have higher adjusted curvature and volume in left pars orbitalis. To our knowledge, the importance of the left pars orbitalis has seldom been reported in previous works in which the learner group comprised mostly younger adults, even though a whole-brain analysis was conducted in these studies (e.g., Bellander et al., 2016). Instead, it is generally understood that the anatomical measures of the left IFG, as the seat of the Broca's area, are sensitive to language processes and bilingualism. For example, its gray matter volume was positively correlated with L2 performance (Mårtensson et al., 2012; Stein et al., 2012). Its cortical thickness was also larger for bilinguals that acquired their L2 in late childhood (8–13 years) than in early childhood (4–7 years) (Klein et al., 2014). Our finding is consistent with this general understanding, but further suggests that the pars orbitalis is especially associated with both immediate and long-term vocabulary retention. It is worth noting that the left temporal pole has barely failed to reach statistical significance in predicting in-class Can-to-Ita (FDR-corrected p = 0.053), which would have provided another piece of convergent evidence.

In support of the hypothesis that episodic memory is also important for older language learners, associations with the performance of both the in-class quiz and the final test were found at the left caudal middle frontal cortex. In particular, the mean curvature of the left caudal frontal cortex was consistently positively associated with the learning performance, including the in-class Can-to-Ita score regardless of week and the first three parts of the final test. By comparison, the adjusted thickness only significantly predicted the in-class Can-to-Ita over Weeks 1–3 while the adjusted cortical surface area was negatively associated with the two translation scores in the final test. In the literature, the left caudal middle frontal cortex is known to be associated with semantic strategies (Kirchhoff et al., 2014; Yu et al., 2018), referential encoding strategies (Yang et al., 2013), self-initiated elaborative encoding strategies (Husa et al., 2017), and working memory (Petrides et al., 1993). The present finding suggests that better-performing older learners might have employed a variety of encoding strategies in vocabulary learning. The present result corroborated with two previous works. Yang et al. (2015) reported that the middle frontal cortex was more active in successful learners in a tone discrimination task. The authors suggested that the higher level of activation was associated with the path between middle frontal gyrus and inferior parietal lobe, where lexical knowledge was automatically activated and sent to middle frontal gyrus. Another study found that the cortical thickness of MFG was increased after FLL (Legault et al., 2019b). Apart from the left caudal middle frontal cortex, converging evidence was also found in the left entorhinal cortex, which is implicated in the episodic processing of object/event information in relation to the context (Eichenbaum et al., 2012). Specifically, the in-class Can-to-Ita performance was positively associated with its cortical thickness.

In contrast, for regions that are especially important for phonological processing, they were associated with only long-term retention. Specifically, the left precentral gyrus is important for phonological working memory, serving as a locus for phonological rehearsal (Veroude et al., 2010; Novén et al., 2019), while the left SMG is known to play important role in phonological processing skills (Veroude et al., 2010; Sliwinska et al., 2012). In partial support to the hypothesis that phonological skills play a role in learning success, for left precentral gyrus, its adjusted thickness and curvature were significantly or marginally associated with the two translation tasks in the final test. The adjusted volume of left SMG was positively associated with the final test score in general, while the adjusted area was negatively associated with the scores in the two translation scores of the final test. The lack of associations of in-class scores with the precentral gyrus and left SMG associations, coupled with the presence of such associations with left pars orbitalis and left caudal middle frontal gyrus, is consistent with our hypothesis that the semantic and episodic memory functions play a more important role for FLL in older adults.

We speculate that semantic and episodic memory processes would be especially needed for intensive vocabulary learning; they are involved, for example, in strategically and proactively generating idiosyncratic semantics-based mnemonics (Thomas and Wang, 1996; Khoii and Sharififar, 2013) or referential encoding strategies for the new words (Turk et al., 2015). Suppose the learner is not confident that they can directly retrieve the meaning of la mela (“the apple,”) a possible cue for retrieval could be the semantic category “fruit,” which could in turn be encoded in terms of the learner's native language. Referential encoding strategies may also be relevant, as the learner could encode the contextual information, e.g., the noun “la mela” was learnt within the word list associated with the verb mangiare (“to eat”) or it was the first word in the first list. Having such cues facilitates the word recall during the in-class translation quizzes. Learners who used these cues may also tend to have better performance in long-term retention, even though the learners no longer need such strategies after many weeks of consolidation.

The cognitive associations reported are consistent with the putative reasons for the strong association of the FLL performance with the left pars orbitalis and left caudal middle frontal cortex discussed above. Across the three cognitive models, only semantic fluency and the delayed recall in the Hong Kong List Learning Test (HKLLT), as measures of semantic retrieval and verbal episodic memory functions, were significant predictors in more than one final models. Specifically, HKLLT significantly contributed to the two in-class scores but not the final test; this pattern suggested that verbal episodic memory is more associated with the immediate learning than long-term language retention performance. In contrast, semantic fluency significantly predicted both in-class Can-to-Ita and the final test. Thus, the evidence suggested that semantic fluency was associated with both immediate learning and long-term retention. One reason that semantic fluency was not significant for predicting in-class Ita-to-Can is that controlled semantic retrieval ability, as captured by the semantic fluency task, has a strong association with the use of semantic associations and strategies. During vocabulary learning, most learners tend to be able to make a direct association of the L2 (Italian) word with the L1 (Cantonese) word, as reflected by a relatively high performance of the Ita-to-Can task in the present study. In contrast, they were much less accurate in the Can-to-Ita task, indicating that they had difficulty in recalling the same association in the reverse direction, i.e., from the L1 (Cantonese) word to the L2 (Italian) word. Such asymmetry has been a point of emphasis in modeling the acquisition of foreign vocabulary, e.g., the parasitic model of vocabulary acquisition (Ecke, 2015). Due to the relative difficulty in Can-to-Ita, the task is subject to a greater use of mnemonic strategies for retrieval. The stronger association of semantic fluency with Can-to-Ita than with Ita-to-Can suggests that semantic retrieval strategies could have been utilized more in the Can-to-Ita task.

Besides HKLLT and semantic fluency, the remaining cognitive measures were kept in at most only one model, namely, in-class Ita-to-Can. In addition to processing speed, which is known to be a general deciding factor of the performance of older adults across domains (Salthouse, 1996), the in-class Ita-to-Can was associated with faster picture naming latency and better Raven's SPM. In contrast, in terms of phonological function measures, while EngSpoon contributed positively to in-class Can-to-Ita, the opposite was true for digit span forward (an index of phonological STM), suggesting that phonological function measures do not necessarily contribute positively. The inhibition function, measured using Stroop Color and Word Test, was also negatively to the learning performance.

Taken together, the anatomical and cognitive models of the learning performance were consistent with our main hypothesis that both semantic and episodic memory functions are prominently implicated in older adults. Phonological skills were also implicated, but their associations with learning performance were not as consistently observed.

Beyond our main hypothesis, the present results could also be viewed from another perspective, in that there were interesting differences in the global pattern of anatomical associations between immediate learning and long-term retention (see Figures 4, 5). As discussed above, the left pars orbitalis and left caudal middle frontal cortex were the only two regions that consistently showed associations with the performance in both immediate learning (Can-to-Ita) and long-term retention. All other regions were only associated with the performance of either the in-class Can-to-Ita or the final test score.

For the in-class Can-to-Ita performance, there were five such regions: left entorhinal cortex, right insula, right pars orbitalis, right caudal anterior cingulate, right middle temporal gyrus. Among them, significant anatomical associations with Can-to-Ita performance were found at left entorhinal cortex (see section 4.1) and right insula. The insula is known to be involved in procedural memory and rule-based learning (Ullman, 2001a, 2004; Yang and Li, 2012). In the present study, the in-class Can-to-Ita score was positively associated with the adjusted volume of the right insula but negatively associated with its adjusted surface area and curvature. In contrast to the right insula, the remaining three right hemisphere regions only showed some subtle changes in the association strength with the in-class Can-to-Ita score, in the sense that while the association was non-significant in any of the five weeks, the magnitude of such association varied significantly across weeks. Taken together, the significant association of right insula and the significant modulation of association over the other three right hemisphere regions suggest that the right hemisphere is implicated in the initial-learning phase. Such right hemisphere associations above added to the literature that reported some associations of the right hemisphere with FLL performance (Hosoda et al., 2013) or bilingualism-induced neuroplasticity (Bubbico et al., 2019; Maschio et al., 2019; Legault et al., 2019b), although the precise roles should be tested using online functional MRI tasks.

For long-term language retention, beyond left pars orbitalis and left caudal middle frontal cortex, seven regions were additionally associated with the final test results: left precentral gyrus (see section 4.1), left SMG (see section 4.1), left pars triangularis, left rostral middle frontal cortex, left banks of superior temporal sulcus, right fusiform gyrus, and right entorhinal cortex. In other words, a more left-dominant network was implicated in long-term vocabulary retention, with seven left hemisphere regions showing significant associations with the final test score when compared to only two right hemisphere regions. Also, unlike the subtle right-hemispheric associations with in-class Can-to-Ita, all regions here showed significant associations with individual test scores. For the left hemisphere, the inclusion of left pars triangularis was not surprising, given that it is a part of the Broca's area. The left rostral middle frontal cortex has been implicated in language learning (Sheppard et al., 2012; Novén et al., 2019), while the left banks of superior temporal sulcus is known to reflect verbal intelligence and receptive language functions in children (Li et al., 2020). For the right hemisphere, the right fusiform gyrus is well-known for its role in object recognition (as opposed to word recognition for the left fusiform gyrus), while the right entorhinal cortex was especially important for long-term memory consolidation (Haist et al., 2001; Piefke et al., 2003).

Overall, while the best anatomical predictor of immediate success was found on semantics-associated (the left pars orbitalis) and episodic memory-related regions (left caudal middle frontal gyrus and the left entorhinal cortex), there were significant association with the right insula and some subtle changes in association with the right pars orbitalis, right rostral anterior cingulate, and right middle temporal gyrus over the course of learning. The increased left-lateralization in the association pattern of the final test scores was consistent with the view that the right-hemisphere homologs of the left-hemisphere language areas may provide “scaffolding” in the early phase of FLL, before a right-to-left shift occurs (Yang et al., 2015; Qi and Legault, 2020).

In previous works, increased hippocampal volume was observed in Swedish individuals studying Italian (Bellander et al., 2016). The hippocampal volume was also a good predictor of the achieved vocabulary proficiency (Nilsson et al., 2021). Apart from the hippocampus, the striatum has been highlighted in previous works (Ullman, 2004; Abutalebi and Green, 2007; Li et al., 2014; Legault et al., 2019a). However, instead of the hippocampus and striatum, our results showed that the thalamus has stronger association with long-term vocabulary retention, with the cortical volumes of the bilateral thalami showing positive associations with the final test scores. The thalamus is implicated in language production by selecting lexical and semantic representations (Abutalebi and Green, 2016), and an expansion of thalamus is associated with greater second language immersion (Pliatsikas et al., 2017; Deluca et al., 2019a). The present finding added to the literature showing that the structural parameters of thalamus are sensitive to aspects of FLL, although the underlying reasons for the associations should be investigated further.

For hippocampus, we observed some degree of association of the right hippocampal volume with the Can-to-Ita test score, but the association did not survive the FDR correction. Indeed, while the role of hippocampus in associative memory, especially in various learning and recall paradigms, is well-established in patient populations in the neurological literature (Suzuki, 2008), direct associations of hippocampal volume with associative memory have not been consistently found in cognitively normal older adults (Becker et al., 2015; Zheng et al., 2017). Finally, for striatum, the sub-cortical volumes were significantly associated with none of the three learning measures. One possibility is that these structures are non-unitary in functions. A finer-grained atlas may be necessary to reveal the associations.

Previous studies have suggested that structural associations could provide valuable information for inferring functional associations (Kanai and Rees, 2011; Chiarello et al., 2013). In the present study, significant associations with vocabulary learning performances were consistently found for the left pars orbitalis and left caudal middle frontal cortex. These results suggest that the individual variations in structural morphometry of the prefrontal lobe are strongly associated with language learning success in older adults, considering that the prefrontal lobe is known to be involved in many “higher” cognitive skills such as language, reasoning, and planning (Wood and Grafman, 2003). In particular, the structural brain and cognitive models were consistent with our main hypothesis that semantic and episodic memory functions likely play an important role in language learning in older adults, with the regions most implicated with these functions being prominently represented in the overall pattern of associations. It should be noted that the present study had not explicitly tested what cognitive processes were involved during learning, given that an online design was not employed. Indeed, an alternative account for the association pattern observed is that the surface-based morphological features and the cognitive predictors are good indices of language aptitude, as previous studies suggested (e.g., Novén et al., 2019). Which could be a parsimonious and alternative explanation why these measures could significantly predict the FLL performance. Nonetheless, language aptitude also has multiple components that are differentially associated with cognitive functions, which could provide a more domain-general account of the anatomical associations with learning performance.

Our research complemented previous works that put a stronger focus on the acquisition of other linguistic elements such as phonology (Wong and Perrachione, 2007) and grammar (Yang and Li, 2012). However, due to the absence of a young learner group in the present study, no conclusion could be drawn regarding the relative contributions of semantic functions in FLL across older and younger adults. On the other hand, the majority of the studies thus far were on younger learners, and the role of the left pars orbitalis and left caudal middle frontal gyrus have seldom been reported. In this respect, the present findings suggest that semantic functions could play a more important role in older adults, although this hypothesis should be re-examined with online experimental designs.

Also, a strength of the present study was that objective measures of learning performance were collected throughout the Italian vocabulary learning protocol. Such longitudinal data play a crucial role in the sensitivity of the present analysis in revealing the anatomical markers of learning performance, despite our relatively modest sample size. In addition, by incorporating different surface-based measures into our univariate LME models, we successfully revealed a relatively consistent pattern of significant brain–learning associations. Where an association was found, the adjusted cortical volume was consistently a positive predictor of learning performance, with only one exception found at the fusiform gyrus. Similarly, the adjusted curvature of the regions involved was a strong positive predictor of learning performance, with only two exceptions found for the left entorhinal cortex and the right insula. In contrast, the adjusted area was consistently a negative predictor of the learning performance, without any exception. Adjusted thickness showed the least consistent pattern, being a positive predictor at four ROIs but a negative predictor at two ROIs.

The present work was limited in several ways. First, the MRI scans were not acquired at the beginning of the Italian learning programme but after an introductory Italian lesson of 90 min, to acclimate the learners with a more uniform experience with the target language prior to the scanning. Previous studies have shown that measurable changes in diffusion MRI indices could be induced by short-term intervention of 45 min (Tavor et al., 2020) or 120 min (Sagi et al., 2012). Structural changes were found after 20 days of language learning (Legault et al., 2019b), with some even estimating that only 2 h of training could bring about structural changes in the brain (Park and Bischof, 2013). However, considering that such structural changes may not be long-lasting, and the neuroplasticity of older adults' brains are lower than younger adults' (e.g., Freitas et al., 2011), it is unlikely that a 1.5-h lesson could significantly influence the anatomical measures extracted, to a point that their associations with learning performance would change qualitatively. Secondly, we acknowledge that from the experimenter's perspective, it would have been best to arrange the cognitive and phonological tests prior to the introductory Italian lesson. However, at the beginning of the programme, most participants were anxious and most motivated to start learning Italian. As such, arranging the prolonged testing of about 3.5 h that are not directly related to learning Italian could be seen as a non-ideal arrangement from the learner's perspective, which could lower their engagement in the programme, leading to an increased dropout rate. Because it was not our primary goal to investigate the learning-induced cognitive gains, a compromise was made to spread the cognitive and phonological tests over visits 3 to 6. Considering that a cognitive gain or an improvement of phonological skills could have already incurred following the 90-min introductory lesson and the 2-h vocabulary lessons in the initial-learning phase, we acknowledge that our decision to delay these assessments could have led to an underestimation of the cognition gains or phonological skill improvements reported. Finally, no post-learning MRI scan was conducted, such that any learning-related neuroplasticity could not be tested. The lack of a younger group also precluded any formal hypothesis testing on the differential brain–learning associations between young and older learners. These issues will be addressed in our next round of data collection.