Neural connectivity changes related to vocabulary knowledge effects during second language comprehension

Vocabulary knowledge is a key factor of second language listening success. Our previous behavioral and electrophysiological studies showed faster processing for sentences containing vocabulary with productive knowledge, accompanied by changes in neural activity during sentence listening tasks. However, it is still unclear how vocabulary knowledge shapes functional connectivity across regions during online comprehension. To address this issue, we estimated phase-based connectivity using the weighted phase lag index, identified between level subnetworks with the Network-Based Statistic, and examined whether vocabulary-level effects emerge as changes in overall network strength, clustered organization, or bridging roles. Connectivity differences were most consistent for productive knowledge, appearing mainly as localized increases in bridging or intermediary roles rather than as broad increases in within-module hubness. However, none of these differences survived family-wise correction, and the findings are best treated as exploratory targets for confirmatory replication.

For second language learners, processing spoken sentences quickly and smoothly is critical for both communication and evaluation settings. This struggle typically decreases with proficiency as comprehension becomes more automatic (Segalowitz and Hulstijn, 2005). Vocabulary is generally acquired in stages, from recognizing a word's meaning to being able to use it productively (González-Fernández and Schmitt, 2019; Read, 2000; Webb, 2012). While knowing the meanings of words is necessary for comprehension, explaining processing difficulty and processing time requires taking into account not only vocabulary size but also the degree of vocabulary mastery.

In our previous behavioral and electrophysiological (EEG) studies, we classified target words into productive, only comprehensive, and non-comprehensive knowledge and found that productive knowledge facilitates second language sentence comprehension speed via changes in neural activity, rather than mere processing acceleration (Allal-Sumoto et al., 2023, 2024). Related EEG and magnetoencephalographic (MEG) studies indicate that

oscillatory activity reflects lexical processing, as well as attentional and control demands during language comprehension (Grant et al., 2022; Piai et al., 2020) and may support integrative processes in speech comprehension (Zioga et al., 2023). Nevertheless, this work has primarily relied on local measures such as EEG power, focusing on changes in individual electrodes or within a limited set of regions.

In contrast to the above, how multiple brain regions are functionally connected during second language comprehension is unclear. The brain is now widely understood as a complex network of connections among many regions (Bullmore and Sporns, 2009; Sato et al., 2021), and network-level measures make it possible to evaluate aspects of information routing, integration, and segregation that are difficult to capture from local activity alone (Rubinov and Sporns, 2010). At the network level, previous studies have reported that connectivity patterns can differ across proficiency levels, even when participants perform the same task (Fraga-González et al., 2016; Fraga-González et al., 2021; Kliesch et al., 2022), suggesting that, in second language sentence comprehension, differences in vocabulary knowledge may also be reflected in changes to functional connectivity patterns, rather than in local activity alone.

The aim of this study is to investigate whether differences in vocabulary knowledge levels (productive, only comprehensive, and non-comprehensive) are accompanied by changes in brain functional connectivity and how such changes manifest as differences in connectivity patterns and node roles during second language listening comprehension. Using the same EEG dataset as in our previous studies, we quantified phase-based connectivity (wPLI; Vinck et al., 2011), identified subnetworks that differed between vocabulary levels (NBS; Zalesky et al., 2010), and characterized their modular organization and node roles using established graph-theoretic measures (Guimerà and Nunes Amaral, 2005; Rubinov and Sporns, 2010; Hsieh et al., 2024; Muller and Meyer, 2014; Power et al., 2011).

We note that the present analyses were exploratory rather than confirmatory. Rather than testing a small set of pre-registered hypotheses about specific bands, time windows, or contrasts, we surveyed candidate subnetworks and node roles across the full space of bands, windows, and vocabulary-level contrasts to generate hypotheses for future confirmatory work. The statistical thresholds and correction procedures reported below should be interpreted accordingly.

The present study reanalyzed the EEG dataset reported in our previous work on oscillatory dynamics during second language listening comprehension (Allal-Sumoto et al., 2024). The participants, stimuli, task procedure, and EEG recording procedures were identical to those described in that study and are summarized briefly below, and a full description of the selection criteria and frequency distributions is provided in that previous work. The novel contribution of the current paper lies in the connectivity and graph-theoretical analyses of the same dataset.

Participants

Twenty-seven undergraduate and graduate students at Kyoto University (14 males, 13 females) took part in the study; two were excluded after the screening experiment for not reaching the required 30 correctly classified phrases in each knowledge category, leaving 25 for the EEG session. Sample size for the main analysis was determined a priori by power analysis (G * Power v3.1.9.7; two-tailed, α = 0.05, 1–β = 0.8, d = 0.8), yielding a target of 15. The 25 participants were split into a preliminary sample (10, used to finalize preprocessing and analysis parameters) and a main sample (15: 5 males, 10 females); one male participant was later excluded from the main sample due to an exceptionally high overall trial-rejection rate across all three knowledge levels, leaving a final sample of N = 14 (4 males, 10 females). Mean age was 22.4 ± 2.06 years (range 20–30). Inclusion criteria were: native Japanese speaker; right-handed (Edinburgh Handedness Inventory; Oldfield 1971); no hearing problems; and an English proficiency score within the past 2 years meeting at least one of TOEIC > 600, TOEFL ITP > 503, or TOEFL iBT > 61. Among the 14 participants in the final sample, proficiency scores at the qualifying test averaged 815.0 ± 79.8 for TOEIC (n = 9, range 695–960), 542.3 ± 10.4 for TOEFL ITP (n = 3, range 530–550), and 69.0 ± 4.2 for TOEFL iBT (n = 2, range 66–72). These cutoffs correspond to a minimum of CEFR B1 for each respective test, based on published ETS score-to-CEFR concordance tables (TOEIC Listening and Reading, TOEFL ITP, and TOEFL iBT); participants' actual scores ranged well above this minimum. All participants received an honorarium. The study protocol was approved by the Research Ethics Committee of the Psychological Science Unit at Kyoto University (approval number: 1-P-2).

Stimuli

Target items were formulaic sequences, defined as word strings stored and retrieved holistically from memory (Wray and Perkins, 2000), specifically three- to four-word adverbial or adjectival sequences that can occur in sentence-final position, excluding phrases with a repeated word (e.g., “one by one”), antonym pairs (e.g., “more or less”), or possessive pronouns (e.g., “above one's head”). Two hundred phrases spanning a wide frequency range were selected from the Corpus of Contemporary American English (7–36,475 occurrences) and the British National Corpus (2–6,875 occurrences). Four example sentences per phrase were drawn from learner dictionaries (Cambridge, Oxford, Collins Online) and lightly edited to eight to 10 words. Representative example sentences included “I will take my umbrella just in case” (productive), “She has been working 10 days in a row” (only comprehensive), and “He is a great teacher by all accounts” (non-comprehensive). Category assignment was participant-specific rather than fixed to particular phrases. Audio was synthesized via Google Text-to-Speech at 150 words per minute, the approximate average native-speaker rate (Wang, 2021). No filler items were included; all trials analyzed here were target trials.

Task procedure

The study consisted of two phases: an initial screening phase and the subsequent EEG experiment. The screening phase was necessary to establish each participant's individualized vocabulary knowledge profile in advance, so that stimulus sets for the EEG session could be constructed per participant. It was administered online and comprised a production task and a comprehension task, each with two sets of 100 trials. In the production task, participants read an English sentence with its Japanese translation, the final word of the target phrase omitted, and typed the missing word (near-correct spellings accepted). In the comprehension task, participants read a complete English sentence and typed a Japanese translation of the underlined target phrase. Although there was no strict time limit, participants were advised to complete each set of 100 items within approximately 30 min. Phrases were classified per participant into three knowledge categories accordingly: productive, only comprehensive, or non-comprehensive. All responses were scored by a single rater. Only participants with at least 30 items in each category proceeded to the EEG session.

The EEG session was conducted at least 1 week after the screening phase, to minimize the effect of memories of the earlier tasks. Although the target phrases themselves were shared between the two phases, the surrounding sentence context differed: each phrase was embedded in a different carrier sentence at screening and at test, further reducing the risk of item familiarization. During the EEG recording, participants performed a listening comprehension task using individualized stimulus sets constructed from their screening results, with the number of trials per condition varying slightly across participants (35.4 ± 4.2 non-comprehensive, 43.9 ± 1.6 only comprehensive, 40.7 ± 5.1 productive). The task consisted of two blocks of 60 trials, each lasting approximately 20 min, pre-sented via Psychtoolbox (Brainard, 1997; Kleiner et al., 2007; Pelli, 1997) in MATLAB (MathWorks). At the start of the session, participants adjusted the sound volume and completed five practice trials (Figure 1). Each trial began with a 300 ms beep followed by a silent interval of 1,000, 1,500, or 2,000 ms (randomly selected), after which an English sentence was played whose final three or four words formed the target phrase. After each sentence, participants rated their confidence in understanding the sentence on a 9-point Likert scale (1–9), primarily to promote sustained attention to the listening task. No cue, visual or auditory, signaled the response window; participants were simply instructed to respond within a self-paced window of up to 10,000 ms, with no instruction to respond quickly.

EEG recording

EEG was recorded with a 32-channel BrainVision MR amplifier (Brain Products, Germany) using an electrode cap arranged according to the international 10–20 system, comprising 28 EEG channels and 4 EOG channels to monitor vertical and horizontal eye movements. Signals were referenced to linked earlobes, grounded at the inion, and sampled at 5,000 Hz with online 1 Hz high-pass, 250 Hz low-pass, and 60 Hz notch filtering. Impedances were kept at or below approximately 10 kΩ. Recordings were conducted in a magnetically shielded, sound-attenuated room, and participants were instructed to keep their eyes closed and to minimize movement throughout the session.

EEG preprocessing

EEG preprocessing was conducted in MATLAB using the FieldTrip toolbox (Oostenveld et al., 2010) for the weighted phase lag index analysis (wPLI) and functional EEG network analysis. The pipeline proceeded in the following order: trial definition and segmentation, EOG artifact reduction, resampling, filtering and average re-referencing, and bad trial rejection. In the trial definition, two types of triggers were recorded (target word onsets and audio file terminations); however, all connectivity analyses were time-locked to the target word onset, which corresponded to the final word of each formulaic sequence. Continuous EEG data were epoched from 1000 ms before to 1,500 ms after the target word onset. Within each epoch, five non-overlapping 500 ms time windows of interest (TOIs; −1000 to −500 ms, −500 to 0 ms, 0 to 500 ms, 500 to 1,000 ms, and 1,000 to 1,500 ms, relative to target word onset) were defined, and wPLI was estimated separately for each TOI.

For EOG artifact reduction, four electro-oculogram channels (HEOGL, HEOGR, VEOGU, VEOGL) were used as predictors in a regression-based correction: For each trial, EEG channels were regressed on the EOG signals, and the residuals were retained as EOG corrected EEG (Gratton et al., 1983). Data were downsampled to 200 Hz and linearly detrended, then band-pass filtered between 1 and 30 Hz using zero-phase Butterworth filters (fourth-order high-pass and sixth-order low-pass) and re-referenced to the average of all non-EOG EEG channels. Bad trials were automatically rejected based on amplitude and derivative criteria within −1,000 to 1,500 ms: A trial was marked as bad if at least three non-EOG EEG channels exceeded 75 μV in absolute amplitude, exceeded 300 μV/ms in smoothed temporal derivative, or showed sustained low amplitude below 0.3 μV for ≥150 ms. These criteria excluded, on average, 7.36 ± 9.42% of trials in the final N = 14 sample, and the remaining trials were saved as long-epoch data (−1,000 to 1,500 ms). Retained trial counts (post-artifact-rejection) averaged 33.3 ± 4.8 (non-comprehensive), 39.9 ± 5.4 (only comprehensive), and 37.8 ± 6.2 (productive) trials per participant (overall range 22–47).

For connectivity estimation, phase information was derived using a Hilbert-based approach. Hilbert transforms were applied to band-pass-filtered signals separately for theta (4–8 Hz), alpha (8–13 Hz), and beta (13–30 Hz), to ensure band-specific phase estimation. For each frequency band, the preprocessed EEG (excluding EOG channels) was band-pass filtered with a zero-phase fourth-order Butterworth filter with symmetric reflection padding (0.5 s at both ends), and the analytic signal was obtained via the Hilbert transform. For each TOI, the imaginary component of cross products was averaged over time within the TOI and then combined across trials to compute the debiased wPLI (wPLI_debiased), yielding an N × N connectivity matrix per subject, level, band, and TOI.

Weighted phase lag index analysis

Because this study surveyed connectivity differences across the full set of 3 frequency bands × 5 time windows × 3 pairwise contrasts without pre-registered, narrower hypotheses, we treat the entire set of 45 NBS runs as a single exploratory analysis family and interpret all reported effects as candidate targets for future confirmatory replication. For each TOI and frequency band, weighted phase lag index analyses (wPLI) were performed to assess the neural connectivity change based on the participants' lexical knowledge. Debiased wPLI matrices were computed separately for each 500 ms TOI and each band as described above, yielding subject-wise connectivity matrices for each level. Between-level network differences were tested with the Network-Based Statistic (NBS; Zalesky, Fornito and Bullmore 2010) in the theta, alpha, and beta bands within the predefined TOI windows. The final sample (N = 14, df = 13) was used for all 45 band × time-window × contrast combinations. The NBS parameters were: a primary (cluster-forming) threshold t = 2.160 (df = 13), 10,000 sign-flip permutations, an extent-based component measure, two-tailed testing, and FWE control at α = 0.05. For each contrast, the SigMask was taken as the union of edges across all FWE-significant components in either direction; contrasts with no significant component were excluded from downstream analysis. Because 45 such tests were run, we additionally applied Benjamini-Hochberg FDR correction across all raw FWE p -values (excluding the additional theta 0–1,000 ms window described below) as a family-wise robustness check.

EEG-based functional network analysis

For each FWE-significant subnetwork, the SigMask was converted into a symmetric weighted adjacency matrix (weighted by | t |), and community structure was estimated using the Louvain algorithm (Blondel et al., 2008; community_louvain, Brain Connectivity Toolbox), repeated 100 times per subnetwork given its stochastic optimization step; the partition with the highest modularity (Q) was retained. Observed modularity was tested against a Maslov-Sneppen degree-preserving rewiring null and a spatially constrained null based on inter-electrode distance (2,000 iterations each), with one-sided p -values defined as the proportion of null Q ≥ Q_observed.

Participation coefficient (P) and within-module degree z-score (Zwithin) instead used predefined anatomical scalp-region groupings based on standard 10–20 electrode prefixes (electrodes with ambiguous prefixes assigned to an “Others” group; full mapping in Supplementary Table 1), independent of the Louvain partition above: Louvain addresses whether the SigMask shows community structure at all, while P and Zwithin characterize each node's bridging vs. locally embedded role relative to a fixed partition, which is necessary for comparing role indices across subjects, levels, and contrasts on a common basis.

Because BC, P, and Zwithin are computed on matrices restricted to edges already selected by the same contrast under evaluation, we treat NBS purely as a localization tool and report these indices descriptively for each FWE-significant contrast, listing the highest-scoring nodes at each level rather than testing between-level or absolute-role differences.

Behavioral relevance analysis

Masked network strength was correlated with participants' trial-by-trial confidence ratings (1–9 Likert scale), averaged per participant and per level to match the trial-averaged nature of wPLI (trials with no response, Key = 0 at the 5,000 ms ceiling, were excluded). For each of the eight FWE-significant contrasts, we computed the paired difference in mean confidence and in masked network strength (logit scale) across participants, and correlated these paired differences (Pearson r, Spearman ρ), consistent with the paired design used throughout. Benjamini-Hochberg FDR correction was applied across the eight resulting p -values.

Weighted phase lag index analysis (wPLI)

Network-Based Statistics (NBS) was applied to Hilbert-based wPLI connectivity, and functional connectivity differences among vocabulary knowledge levels were assessed using masked network strength, defined as the mean wPLI within FWE-significant NBS masks. Prior to target onset, alpha-band connectivity was greater for sentences that included phrases with non-comprehensive knowledge than only-comprehensive knowledge at both −1,000 to −500 ms (L1 > L2; p = 0.036; 27 edges) (Figures 2a, b) and −500 to 0 ms (L1 > L2; p = 0.017; 47 edges) (Figures 2c, d), and greater for non-comprehensive than productive knowledge at −500 to 0 ms (L1 > L3; p = 0.013; 37 edges; the largest effect size observed in this study, Cohen's dz = −2.10, 95% CI [−2.64, −1.55]) (Figures 2e, f).

Immediately after target onset (0–500 ms), connectivity was stronger for productive knowledge in theta (L2 < L3; p = 0.044; 15 edges) (Figures 3a, b) and beta (L1 < L3; p = 0.005; 43 edges; L2 < L3; p = 0.009; 41 edges) (Figures 3c–f). In the following window (500–1,000 ms), productive knowledge again showed stronger connectivity in theta (L2 < L3; p = 0.043; 16 edges) (Figures 3g, h) and alpha (L1 < L3; p = 0.022; 30 edges) (Figures 3i, j). No component survived FWE correction in the 1,000–1,500 ms window. Effect sizes and 95% confidence intervals for all eight FWE-significant contrasts are reported in Supplementary Table 2.

Because 45 NBS tests were conducted (5 time windows × 3 bands × 3 contrasts), we additionally applied Benjamini-Hochberg FDR correction across all 45 raw FWE p -values. None of the eight nominally significant contrasts survived this correction (all q ≥ 0.19). We therefore treat the results above as exploratory rather than confirmatory, and report the full 45-run table (raw p and FDR q) as Supplementary Table 3.

Because 500 ms windows contain only 2–4 cycles at theta frequencies, which may be marginal for stable phase estimation, we additionally tested a single, longer 0–1,000 ms theta window spanning the two significant 500 ms theta windows above (L2 vs. L3). This contrast remained FWE-significant (14 edges), with the same direction of effect (L2 < L3). Because this window overlaps substantially with the two shorter theta windows already included in the 45-run family, it was analyzed separately and excluded from that FDR correction.

Change in confidence rating did not correlate with change in masked network strength for any of the eight contrasts after FDR correction (all q ≥ 0.38). The strongest uncorrected associations, for theta 500–1,000 ms (L2 vs. L3; r = −0.50, p = 0.066) and beta 0–500 ms (L1 vs. L3; ρ = −0.53, p = 0.052), were both in the opposite direction to what a confidence-driven account would predict (full results in Supplementary Table 4).

Post hoc characterization of NBS-significant subnetworks

These analyses are conditioned on the SigMask from Section 3.1 and are descriptive characterizations rather than independent confirmatory tests. For each of the eight FWE-significant subnetworks, we tested whether the SigMask showed community structure beyond chance using the Louvain algorithm (Blondel et al., 2008; 100 repetitions per subnetwork given its stochastic optimization step, SD(Q) ≤ 0.01), against a degree-preserving rewiring null and a spatially constrained null (2,000 iterations each). Only one subnetwork (alpha, −500 to 0 ms, L1 vs. L2) exceeded both nulls (p _rewire = 0.009, p _spatial = 0.015); one (beta, 0–500 ms, L2 vs. L3) exceeded the rewiring null only (p _rewire = 0.021, p _spatial = 0.219), and one (theta, 500–1,000 ms, L2 vs. L3) exceeded the spatial null only (p _rewire = 0.151, p _spatial < 0.001). The remaining five did not differ from either null (all p ≥ 0.09). Theta-band subnetworks were consistently sparse and fragmented (median 15 communities across 28 nodes, largely singleton), limiting interpretability in this band; the sensitivity window (theta, 0–1,000 ms) showed the same pattern (p _rewire = 0.628, p _spatial = 1.000).

Node roles within each SigMask were characterized descriptively using betweenness centrality (BC) and participation coefficient (P), reported separately at each level without between-level testing, to avoid conditional inference on the contrast used to select the mask. Elevated BC and P for productive knowledge were most consistent at right-occipital (O2), midline-frontal (Fz), left fronto-central (FC1, FC2), and left temporo-frontal (T7, F7) sites, though specific nodes and magnitude varied by band and window (full listings in Supplementary Table 5). Within-module degree z-score (Zwithin) was uninformative for most contrasts, as SigMasks were often fragmented into singleton or near-singleton module memberships, for which Zwithin is degenerate; we do not interpret it further. Connector-hub prevalence (P ≥ 0.30, Zwithin ≥ 2.50; Guimerà and Nunes Amaral, 2005), computed on the whole-brain network as an independent check, was low throughout (at most 2–4 of 14 participants per node), consistent with few nodes meeting connector-hub criteria in this dataset.

Neural functional connectivity changes based on vocabulary knowledge

In this study, we examined whether vocabulary knowledge levels are associated with connectivity changes and patterns and node roles during second language listening comprehension.

These findings should be interpreted with several caveats. Although eight subnetworks showed nominally FWE-significant connectivity differences across the 45 tested combinations, none survived Benjamini-Hochberg correction applied across the full analysis family, and we treat them as exploratory and hypothesis-generating. Two patterns were nonetheless consistent across contrasts: stronger pre-target alpha connectivity for non-comprehensive knowledge (three contrasts, both pre-target windows), and stronger post-target connectivity for productive knowledge in theta, beta, and alpha (five contrasts, both post-target windows). Future confirmatory work with narrower, pre-registered hypotheses restricted to these windows would help establish their replicability.

Community structure was likewise null-model-dependent, holding up in only one of the eight subnetworks (alpha, −500 to 0 ms, L1 vs. L2); modular organization elsewhere should be read as descriptive rather than as evidence of a robust community.

Node-role patterns are descriptive rather than confirmatory, as BC, P, and Zwithin were computed on edges already selected by the same contrast. BC and P capture related but distinct aspects of network organization: BC reflects a node's intermediary position along shortest paths, and P reflects how evenly its connections spread across anatomical regions. Elevated values on either or both were most consistent for productive knowledge at right-occipital, midline-frontal, and left fronto-central/temporal sites. Within-module hubness (Zwithin) and connector-hub classification showed no consistent pattern, likely reflecting the sparse, fragmented structure of several subnetworks, particularly in theta, rather than a genuine absence of within-module specialization. As this was a 32-channel sensor-level analysis, anatomical interpretations such as fronto-central or posterior-frontal patterns should be read as approximate rather than source-localized.

This study reanalyzed the same EEG dataset as our earlier report on oscillatory power changes associated with productive vocabulary knowledge (Allal-Sumoto et al., 2024), but addressed a different question: how vocabulary knowledge shapes large-scale functional connectivity and network organization during comprehension, using wPLI, NBS, and graph-theoretic node-role analysis not included in that earlier work. The two studies are complementary, one characterizing local neural activity and the other network-level information routing.

Possible mechanisms of neural functional connectivity during comprehension

The pre-target alpha connectivity pattern may reflect proactive control preparing the system for upcoming input: when lexical representations are less consolidated, listeners may rely more on top-down attention and inhibitory suppression of irrelevant input (Foxe and Snyder, 2011; Jensen and Mazaheri, 2010; Obleser and Weisz, 2012), consistent with an adaptive control strategy rather than richer linguistic representations (Farah et al., 2019).

Immediately after target onset, stronger theta and beta connectivity for productive knowledge may reflect more efficient lexical access and integration: theta connectivity is linked to lexico-semantic binding (Bastiaansen and Hagoort, 2006; Meyer, 2017), while beta connectivity supports syntactic unification into phrase- or sentence-level structure (Bastiaansen et al., 2010; Weiss and Mueller, 2012), together suggesting more efficient semantic binding and structural integration when productive knowledge is available.

In the later window, stronger theta and alpha connectivity for productive knowledge may reflect continued integration and stabilization of the developing interpretation: theta connectivity supports ongoing semantic binding and coherence checking (Bastiaansen and Hagoort, 2006; Meyer, 2017), while alpha connectivity reflects inhibitory stabilization as competition is resolved (Obleser et al., 2012), together suggesting sustained network coordination beyond the initial integration window.

Because confidence ratings rose with knowledge level for 13 of 14 participants, connectivity differences here could in principle reflect confidence rather than knowledge itself; however, participant-level changes in confidence did not correlate with changes in masked network strength for any contrast after FDR correction, arguing against this alternative account, though as an indirect, individual-differences check this does not fully rule out some contribution from confidence.

This study showed the importance of productive vocabulary knowledge during listening comprehension. Knowing the meaning of vocabulary seems to be a minimum requirement for understanding a sentence, but it is not enough in terms of processing speed and efficiency. In real life, listening and speaking in a foreign language is demanding. The load becomes especially heavy when we must comprehend what we hear while simultaneously preparing what to say next. Therefore, to reduce the load of conversation, the acquisition of productive knowledge is highly important.

Limitations

Although the current study suggests that vocabulary knowledge relates to neural connectivity during sentence listening, several limitations should be noted. First, as noted above, downstream node-role indices could not be tested for between-level differences without risking conditional inference, and community structure held up in only one of eight subnetworks, particularly limited in the sparse, fragmented theta band. Second, node role metrics depended on coarse, anatomically defined scalp-region groupings that may not align with task-specific functional communities; future work should refine module specification and complement mask-based analyses with whole-network approaches. Third, although knowledge level was assigned per participant per phrase, phrase-level lexical properties (e.g., corpus frequency, length) were not explicitly modeled here; a separate behavioral study using an overlapping stimulus set found that the effect of knowledge level on comprehension speed remained significant after controlling for these properties (Allal-Sumoto et al., 2023), though residual item-level confounds cannot be fully ruled out. Finally, we did not equalize the number of retained trials across conditions, relying instead on the debiased wPLI estimator (Vinck et al., 2011) to mitigate sample-size-related bias, and although participants had up to 10,000 ms to respond with no instruction to hurry, some contribution of decision-making, response preparation, or motor activity to the analyzed time windows cannot be fully ruled out.

This study examined whether vocabulary knowledge levels are accompanied by time-locked changes in functional EEG connectivity during second language listening comprehension. Across the comprehension time course, connectivity increases were most consistent after target onset for sentences containing productive knowledge, alongside localized elevations in bridging-intermediary network roles rather than widespread shifts in within-module hubness; however, none of these differences survived correction across the full set of tested combinations, and the findings should be treated as exploratory rather than confirmatory. With that caveat, the results are consistent with the idea that vocabulary knowledge shapes comprehension not only through how much the brain activates, but also through how information is routed across functional networks, with differences in knowledge level appearing as differences in route selection and role allocation. These patterns offer candidate hypotheses for future confirmatory work linking lexical knowledge to temporally specific neural coordination.