acceptodds
Under review as a conference paper at ICLR 2027

Comparing cortical organization and semantic tuning across bilingual brains

Abstract

In bilingual individuals, the brain regions involved in processing semantics in native and non-native languages are broadly overlapping and represent largely similar semantic concepts, although fine-grained differences exist. However, how these broad similarities and fine-grained differences vary across individuals remains unclear. Here, we compared the spatial cortical organization and semantic tuning of semantic representations using functional magnetic resonance imaging (fMRI) data from six Chinese-English bilingual participants who read several hours of narratives in both languages. We fit voxelwise encoding models separately for each participant and language and applied hierarchical clustering to group-averaged estimated model weights to derive a common set of semantic clusters. We then compared the cortical organization and semantic tuning of these clusters across participants and languages. For all clusters, cortical overlap was higher across languages within a participant than across participants within a language. Semantic tuning similarity showed the opposite pattern, suggesting largely shared tuning across participants within the same language. Cross-participant semantic tuning similarity was lower for the communication and family clusters than for the other clusters, suggesting greater individual variability for social concepts. Together, these findings reveal a participant-specific spatial cortical organization that is preserved across languages and fine-grained language-specific patterns of semantic tuning largely shared across participants.

Primary area:
applications to neuroscience & cognitive science
Keywords:
fMRI, bilingual, semantic tuning, cortical organization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_VWfYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper compares fMRI data from six Chinese-English bilingual participants who read several hours of narratives in both languages. Specifically, they compare spatial cortical organisation and semantic tuning of semantic representations across subjects. To make these comparisons, they use some common brain alignment techniques (voxelwise encoding models) and build upon the analysis of Chen et al. by doing a clustering of model weights with 5 centroids that they associate with categories and then basing their analysis on the association to these clusters. The findings are that the cortical overlap was higher across languages than among participants, while the semantic tuning similarity showed the opposite pattern. But they admit that these results depend on the mapping of the individual cortical surfaces to the common fsaverage space.

Critical strengths

The paper is well written and explained, and the questions are well motivated. The experiments are clear, and the procedure is correct. They use common brain alignment techniques, banded ridge regression, and temporal correction, and fMRI data that was already used in the literature.

Critical weaknesses
  • (W1) While ICLR welcomes computational neuroscience work, such as encoding models or RSA comparing model and brain representations, the machine learning angle of this paper is limited. The contribution is primarily neuroscientific since it compares the organisation of semantic representations across bilingual participants, and the rest of the pipeline (voxelwise encoding models, projection to a common surface space, and clustering of model weights) combines well-established techniques. The paper does not offer new insights into model representations or into the relationship between model and brain representations and does not frame its contribution to an ML audience.

  • (W2) The contribution appears incremental relative to prior work with an almost identical design and pipeline (Chen et al. 2026). The main difference is that, with the same data and the same methodology, once Chen et al. established cross-language consistency within each brain, this paper adds the cross-participant comparison; it now compares 15 pairs per language (30 total), using Pearson correlation on the re-averaged weight vectors for semantic tuning and the Dice coefficient for cortical overlap.

  • (W3) The paper's central claim is precisely that the cross-participant overlap is low. Comparing across subject brains required projecting all subjects onto a common space (fsaverage). In Chen et al., they also do this projection, but their comparisons were within the same brain, so the answer did not depend on the space. The problem is that imperfect individual-to-template mapping mechanically lowers the cross-participant Dice. The artefact and the expected result point in the same direction, so a low Dice cannot by itself support the conclusion. The authors acknowledge this in their own limitations section: “Comparisons of cortical organisation across participants rely on mapping individual cortical surfaces to the common fsaverage space. Imperfect alignment may reduce the measured cortical overlap across participants.”

Questions for authors
  • (W3) Run the experiments proposed in the future work to examine how robust the pattern found is to different alignment methods.

  • (W3) The limitation discusses how mapping individual cortical surfaces to fsaveragre may affect the cortical organisation results. Could the authors also discuss how this mapping may affect the semantic tuning results? This mapping, beyond being used to compute the centroids, is also used to determine the cluster assignment that results in the mean weight vector for the cross-participant comparisons (15 per language).

  • What is the pairing scheme in the paired t-tests on Figures 3 and 4?

1 comment
AuthorsOpenReview ↗

Critical weaknesses and questions

We thank the reviewer for their thorough and constructive review. We appreciate that they found the paper well written and explained, and the question well motivated. We are encouraged that they found the experiments clear and the procedure correct.

Critical weaknesses: W1) We agree that our main contributions are neuroscientific. We have added suggestions for future work aimed at a machine learning audience. Lines 510-513 in the Future work section now read: “Our comparison across participants and languages motivates a parallel question for multilingual language models. How do semantic representations vary across independently trained instances of the same model when they process the same language, and how do they vary across languages within a single instance?”

W2) Our study advances beyond Chen et al. (2026) in three ways. First, we explicitly differentiate the cortical location of semantic representations from their semantic tuning. Second, we compare variation across participants with variation across languages, thereby placing bilingual differences in the context of individual differences. Third, this comparison reveals opposite patterns: semantic tuning is more similar across participants within a language, whereas cortical location is more consistent across languages within a participant. This shows that semantic tuning of representations can be shared across participants even when their cortical location is participant-specific. Chen et al.’s within-participant comparisons could not reveal this pattern.

Questions:

(W3) Run the experiments proposed in the future work to examine how robust the pattern found is to different alignment methods. We agree that imperfect alignment can reduce cross-participant cortical overlap. To assess whether the pattern holds under a different anatomical mapping, we repeated the analysis in MNI space instead of fsaverage. We used affine registration to map each participant’s normalized model weights and prediction scores to FSL’s 2 mm MNI152 T1 brain template. We then performed the clustering and the comparisons as in the original analysis. The results are reported in Appendix section A.4. All analyses were conducted in MNI space. The resulting clusters were projected onto fsaverage only for visualization on a single cortical map. For both mBERT and fastText, as in our original analysis, cortical overlap is significantly higher across languages within a participant than across participants within a language. And the opposite pattern is observed for semantic tuning similarity as well. Thus, we observe the same ordering in both the MNI and fsaverage template spaces. This suggests that our findings are not specific to the fsaverage projection. This does not eliminate alignment error as a contributing factor, and we have made this qualification more explicit in the Limitations.

Could the authors also discuss how this mapping may affect the semantic tuning results? Cluster assignment is based entirely on model weight vectors in embedding space, without using cortical coordinates or requiring vertices assigned to the same cluster to occupy corresponding locations across participants. Centroid estimation depends on anatomical alignment, which affects how weights are averaged across participants at corresponding template vertices and which vertices pass the group-average prediction threshold in both languages. We verified that assigning clusters after projecting weights into fsaverage gives the same assignments as assigning clusters in native space and then projecting the labels into fsaverage.

What is the pairing scheme in the paired t-tests on Figures 3 and 4? For both cortical overlap and semantic tuning similarity, we computed one cross-language average and one cross-participant average per participant, then compared these six pairs using a paired t-test.

Reviewer_J3PNOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how semantic representations vary across individuals and languages using fMRI data from six Chinese-English bilingual participants. The authors construct voxelwise encoding models and identify five semantic clusters using hierarchical clustering. The results suggest that cortical organization is more consistent across languages within individuals, while semantic tuning is more similar across individuals within the same language. The study also finds greater individual variability in representations of social concepts.

Critical strengths
  1. The paper provides an interesting comparison of cross-language and cross-participant semantic representations, distinguishing cortical organization from semantic tuning.
  2. The experiments use naturalistic bilingual stimuli and two embedding models, mBERT and fastText, providing complementary evidence for the main findings.
Critical weaknesses
  1. The study includes only six participants, all Chinese-English bilinguals. This substantially limits generalizability to other individuals, language pairs, and bilingual populations.
Questions for authors

Given my limited expertise in this specific neuroscience domain, I do not have additional technical questions beyond the concerns noted above.

1 comment
AuthorsOpenReview ↗

Critical weaknesses

We thank the reviewer for their positive review. We are glad that they found the study interesting and that they appreciated the experimental setup and the use of two complementary feature spaces. Our participant sample only consists of 6 participants, but extensive amounts of data (more than 2 hours in each language, requiring approximately 8 hours of scanning session across multiple days) were collected in each participant. We are currently preparing to collect data on other L1 - L2 language pairs on a different set of participants, which will allow us to explicitly test generalizability. We have made these limitations more explicit in the Discussion and Limitation sections.

Reviewer_TJdeOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper reuses an existing fMRI dataset (Chen et al., 2026) in which six Chinese–English bilinguals read about 2 hours of narratives in each language. It fits voxelwise banded-ridge encoding models on mBERT layer-9 or aligned fastText features, after regressing out seven low-level sensory feature spaces. It then projects each participant’s weights to fsaverage and runs Ward clustering on the group-averaged weights, which gives 5 clusters: numbers/names, location, cognition, communication and family. Each vertex in each participant and language is assigned to its nearest cluster centroid.

Two comparisons follow:

Cortical overlap (Dice score): higher across languages within a participant than across participants within a language. Semantic tuning similarity (Pearson correlation of each cluster’s mean weight vector): the reverse, higher across participants within a language than across languages within a participant.

Secondary findings:

With mBERT only, cross-participant similarity is higher in Chinese (the native language) than in English. The communication and family (“social”) clusters are more participant-specific.

Critical strengths
  1. A clear, interesting question: is variability in bilingual semantic representations mostly about where (between people) or how they’re tuned (between languages)? The opposite patterns for the two measures are a neat headline.

  2. Rich within-participant data: several hours per language per participant, a held-out test narrative, banded ridge, and low-level confounds regressed out.

  3. Replication across two feature spaces (mBERT and fastText), and the Dice results are reported to agree with IoU and match rate.

Critical weaknesses
  1. The tuning result may be an artefact of the feature space. This is my main concern. Tuning similarity is computed on weight vectors in feature space. Within a language, every participant shares the same design matrix, and ridge weights inherit structure from the stimulus covariance, so you’d expect some cross-participant correlation even without shared neural tuning. Across languages, the design matrices differ: different words, word order and timing (the Chinese is a translation), and the embeddings themselves aren’t fully aligned. mBERT in particular keeps language-specific directions. So even with identical tuning in both languages, the weight vectors would differ. That alone predicts the reported pattern. The fact that the native-language effect in Fig. 5A depends on which embedding model is used fits this explanation. It would be good to see: (a) a simulation that pushes identical tuning through both languages’ features, to get a ceiling for cross-language correlation; (b) tuning compared in a common space, for example, predicted responses to translation-equivalent probe items; (c) a split-half reliability within participant and language.

  2. Circularity in the cluster analysis. The centroids come from the average over all participants and both languages, and vertices are then assigned to the nearest centroid. Averaging the weights within a cluster pulls everyone’s mean vector toward the shared centroid, which inflates similarity. Leave-one-participant-out centroids would help. k = 5 is taken from the earlier paper without a robustness check here, and I couldn’t work out how the fastText clusters were matched to the mBERT labels; the colour order differs between Figs. 6 and 7.

  3. Statistics. With six participants, the tests are run over pairwise values: 30 cross-language versus 150 cross-participant. Those aren’t independent, because pairs share participants and the clusters are correlated. I also don’t see how a “paired” t-test works with groups of different sizes. I’d expect participant-level permutation tests or a mixed model; the p < 0.001 values are probably too small as they stand.

  4. Missing basics: No prediction accuracies or noise ceilings are reported for either language. The threshold is inconsistent: √R² > 0.05 in §3.7, but 0.1 in the Fig. 2 caption and the text around line 256. Relevant literature is missing: hyperalignment (Haxby et al.), the shared response model, and individual-specific functional mapping (Fedorenko’s functional regions of interest; Braga & Buckner).

This uses the same data, features, clustering and k as Chen et al. (2026); the new part is the cross-participant comparison. There’s no methodological contribution and no link back to representation learning. For example, does mBERT’s own cross-lingual alignment predict which clusters differ across languages? As it stands, it reads like a neuroscience follow-up paper.

Questions for authors

What cross-language weight correlation do you get if the underlying tuning is identical?

What are the split-half ceilings for Dice and for tuning similarity?

Which vertices are used for Dice?

Do the results survive leave-one-out centroids, other values of k, and participant-level permutation tests?

Does the social-cluster effect hold after controlling for cluster size and R²?

2 comments
AuthorsOpenReview ↗

Critical weaknesses

We thank the reviewer for their thorough and constructive review. We appreciate that they found the question clear and interesting and that the findings constitute a neat headline. We are encouraged that they highlighted the rich dataset and adequate encoding modelling as well as the replication with two feature spaces.

Critical weaknesses:

  1. We thank the reviewer for raising the issue that using different features across languages could affect the comparison of semantic tuning. To address this potential confound, we replicated our analysis while holding the features constant across languages. That is, we estimated representations from both Chinese and English brain responses using Chinese features, and then repeated this analysis using English features, pairing the features and brain responses run by run. We then compared cross-participant and cross-language tuning similarity separately for each shared feature space. In these analyses, the features are the same, the only difference are the participants’ brain responses in each language. The results are reported in Appendix section A.2. For both mBERT and fastText, and with either Chinese or English shared features, semantic tuning similarity is higher across participants within language than across languages within participant, which replicates our original finding. This difference is statistically significant in three of the four conditions. For mBERT with shared Chinese features, the difference is in the same direction but does not reach statistical significance (one-sided paired t-test, p = 0.086). Several cross-participant weight correlations in this condition fell below 0.20 (Fig. 8-A). Differences in timing and content across translations may have left features and brain responses imperfectly aligned even after trimming them to the same length. This misalignment may have contributed to lower prediction accuracy in the cross-language fits. The lower accuracy may reflect less reliable model weight estimates, which could help explain some of the low cross-participant correlations. Despite this imperfect alignment, the consistent direction across all four conditions and significant differences in three suggest that the original finding is not solely attributable to using different features across languages.

  2. We replicated the analysis using a leave-one-out procedure: For each participant, the centroids were computed using all other participants. The results are reported in Appendix section A.1. The original pattern is consistent in all 6 cross-validation folds, suggesting that including each participant in centroid estimation does not account for the findings. We retained k = 5 for consistency with Chen et al., since we use the same brain responses and features. The clusters were matched by visual inspection of their cortical location, and we updated the order of the colors in the caption of Fig. 13 and Fig. 14 (previously Fig. 6 and 7).

  3. The tests were not run on the 30 versus 150 values. For both cortical overlap and semantic tuning similarity, we computed one cross-language average and one cross-participant average per participant, then compared these six pairs using a paired t-test. We will clarify this in the Methods. However, the cross-participant averages still share pairwise measurements, so averaging by participant does not fully remove the dependence.

  4. We added cortical maps of prediction accuracy for both languages and both models in Appendix Figures 11 and 12. These figures also show the prediction accuracy when features from one language were used to predict responses in the other. We corrected the typo for the R^2 value in §3.7. We also added the suggested literature in the Related work in Lines 115-123.

Our study advances beyond Chen et al. (2026) in three ways. First, we explicitly differentiate the cortical location of semantic representations from their semantic tuning. Second, we compare variation across participants with variation across languages, thereby placing bilingual differences in the context of individual differences. Third, this comparison reveals opposite patterns: semantic tuning is more similar across participants within a language, whereas cortical location is more consistent across languages within a participant. This shows that semantic tuning of representations can be shared across participants even when their cortical location is participant-specific. Chen et al.’s within-participant comparisons could not reveal this pattern.

Questions:

“What cross-language weight correlation do you get if the underlying tuning is identical?” We interpret this question as asking what cross-language correlation we would recover between the estimated weights if the underlying tuning were identical. The reviewer suggested a simulation that pushes identical tuning through both languages’ features. However, simulating such data requires assumptions about the magnitude and the spatial and temporal structure of fMRI noise, both of which can affect the recovered cross-language weight correlation. To address the concern about differences between the feature matrices, we repeated the analysis while holding the features constant across languages. We fit models to the Chinese and English brain responses using the same Chinese features and then repeated the analysis using the same English features. Within each shared feature space, we compared cross-participant tuning similarity within language with cross-language tuning similarity within participant. This control tests whether the observed difference persists when both languages are modeled using the same features. For both mBERT and fastText and with either set of shared features, tuning similarity was higher across participants within language than across languages within participant. This difference was statistically significant in three of the four conditions. This suggests that differences between the feature matrices alone are unlikely to explain the original finding.

“What are the split-half ceilings for Dice and for tuning similarity?” We thank the reviewer for suggesting this analysis. The feature space and clustering controls address potential confounds that could account for the main pattern. Split-half ceilings would complement these controls by quantifying how reproducible Dice and tuning similarity are within the same participant and language. We agree that these estimates would provide useful benchmarks for interpreting our results and plan to examine them in future work.

“Which vertices are used for Dice?” After selecting vertices whose group-average prediction score exceeds the threshold in both languages, we compute Dice for each cluster using all selected vertices assigned to that cluster in either of the two maps being compared.

“Do the results survive leave-one-out centroids, other values of k, and participant-level permutation tests?” The results survive leave-one-out centroids as shown in Appendix section A.1. In future work, we would like to assess how sensitive the results are to other values of k, and evaluate their robustness to participant-level permutation tests.

“Does the social-cluster effect hold after controlling for cluster size and R²?” To assess whether the social-cluster difference persisted after accounting for cluster size and prediction performance, we fit regressions between these factors and semantic tuning similarity. For mBERT and fastText separately, we predicted mean cross-participant within language tuning similarity from mean vertex count, mean sqrt(R^2), and an indicator for Communication and Family. The estimated social-cluster difference was negative in both models (mBERT: beta=-0.085, p=0.506; fastText: beta=-0.216, p=0.235). These estimates suggest that the social-cluster difference may persist after accounting for cluster size and prediction performance. However, neither estimate reached statistical significance, and each regression was based on five cluster means.

Reviewer_n6f2OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper examines how semantic representations vary across people and languages using narrative-reading fMRI data from six Chinese–English bilinguals. Encoding models based on mBERT and fastText define five semantic clusters. Their cortical locations are more similar across languages within a person, while their average tuning profiles are more similar across people reading the same language. The authors also find greater between-person variation in the communication and family clusters and discuss differences between contextual and lexical representations.

Critical strengths

The contrast between cortical location and semantic tuning is the part I find most interesting. Two people need not recruit precisely corresponding cortical locations to represent similar semantic information, and the bilingual setting offers a useful way to examine this distinction. Figures 3 and 4 bring these two aspects together clearly. The opposing patterns are worth reporting and could help refine what we mean when we describe semantic representations as shared.

Using corresponding narratives in both languages is a strength of the design, since the comparison concerns broadly matched content rather than unrelated language samples. The recurrence of the main pattern with both mBERT and fastText also makes the observation more interesting than a result tied to one particular representation model. I do not think this work needs a new encoding architecture to make a contribution; a convincing finding about bilingual semantic organization would be sufficient.

Critical weaknesses

The main result tells us where similarity is greatest, but I am less clear about what it tells us about meaning. A higher tuning correlation between two people reading Chinese than between one person reading Chinese and English does not identify which semantic distinctions are shared or which change with language. For example, the same conceptual representation could be expressed with different emphases during comprehension, without the underlying conceptual distinctions changing. The discussion largely returns to the ordering of the similarities. I would find the contribution stronger if the authors could connect that ordering to identifiable semantic content, rather than leave the reader to infer what “language-specific tuning” means for bilingual understanding.

I am also not persuaded that the category result is specifically about sociality. The analysis shows that the communication and family clusters vary more across people than the other three clusters. That is a useful observation, but these are not independently defined, matched social and nonsocial categories. The paper itself discusses prior findings about abstractness, which offers another possible way to understand such variation. Without distinguishing these accounts, I would read the result as a finding about these two categories, rather than a broader conclusion about social concepts. Establishing what property of their semantic content accounts for the individual variation would give this part of the paper more scientific weight.

Questions for authors
  1. Section 2 states that the study uses the data and semantic clusters of Chen et al. (2026), then what is the main scientific conclusion here that goes beyond that work? I understand the added cross-participant comparison, but would appreciate a precise account of how it changes our understanding of bilingual representation, rather than only extending the set of comparisons.

  2. How far can the Chinese–English difference be interpreted as a difference between native and non-native language processing? All participants have the same direction of language acquisition. Is the proposed explanation expected to reverse in native English speakers who learned Chinese, or could it depend on properties of these particular languages? I am asking about the intended scientific claim, not requesting a new cohort for the response.

  3. Are individual differences in cortical location related to individual differences in tuning? For instance, do participants with more unusual spatial maps also have more unusual semantic profiles, or can the two vary independently? The analyses treat them separately; their relationship could tell us whether different cortical arrangements support essentially the same semantic organization.

  4. What specifically motivates the contextual-processing interpretation of the mBERT–fastText difference? The discussion reasonably presents it as tentative, but is there evidence linking the effect to contextual meaning beyond the fact that one model is contextual and the other is lexical? A concrete prediction of that interpretation would help establish whether the model comparison contributes an additional cognitive finding.

2 comments
AuthorsOpenReview ↗

Critical weaknesses

We thank the reviewer for their thorough and constructive review. We appreciate that they found the contrast between cortical location and semantic tuning interesting and recognized the bilingual setting as a useful way to study this distinction. We are encouraged that they highlighted the corresponding narratives and replication with two feature spaces as strengths.

Critical weaknesses:

We understand the reviewer’s concern about the interpretation of the semantic tuning result. While limited, our results provide a connection to identifiable semantic content: semantic tuning for number/names, location and cognition concepts is more similar across participants than for communication and family concepts. The lower individual variation for these three categories suggests that they may contribute particularly to the language-specific semantic tuning shared across participants. We have added this interpretation to the Discussion in Lines 485-488. We would like to test which semantic dimensions within each cluster vary most and least across participants and languages. Because this requires carefully defining those dimensions and verifying that they capture comparable distinctions in Chinese and English, we plan to pursue this analysis in future work. We added this direction to the Future Work in Lines 509-510.

In the original manuscript, we interpreted the lower cross-participant similarity of the Family and Communication clusters as related to sociality, based on a qualitative inspection of these concepts and the clusters’ top words. We agree that these categories were not independently defined or matched for sociality. We also agree that abstractness may play a role, given cited literature. To quantify the role of both properties, we measured the sociality and abstractness of each cluster. To do this, we constructed a sociality and an abstractness dimension in fastText space using word rating databases (Diveica et al., 2022; Brysbaert et al., 2014). For each property, we constructed candidate dimensions by subtracting the mean fastText vector of the k lowest-rated words from that of the k highest-rated words, for several values of k. To validate each candidate dimension, we projected the fastText vectors of the stimulus words onto the corresponding dimension and correlated their projection scores with their database ratings. The correlations were highest at k = 500 for both abstractness (r = 0.749) and sociality (r = 0.684), suggesting that these dimensions reliably capture these properties. We used the resulting dimensions to quantify abstractness and sociality in each cluster. We then measured the sociality and abstractness of each cluster in 2 ways. First, we projected the fastText vectors of its top 20 words onto each dimension and summed the projections. Second, we projected its mean model weight vector onto each dimension. The methodology and results of this analysis are reported in the Appendix section A.3. We used only fastText and not mBERT for this analysis because the databases provide ratings for isolated words with no context. Family ranked first in sociality by both measures. Communication ranked second based on the top 20 words and third based on the mean model weight vector. These rankings are consistent with a possible contribution of sociality to the lower cross-participant tuning similarity of Family and Communication. In abstractness, Cognition ranked first by both measures, followed by Communication and Family. Cognition also showed the highest cross-participant tuning similarity of all clusters while Communication and Family showed the lowest (Figs. 5B and 18A). The abstractness rankings thus provide mixed evidence for an association between greater abstractness and lower cross-participant tuning similarity.

Questions:

  1. Our study advances beyond Chen et al. (2026) in three ways. First, we explicitly differentiate the cortical location of semantic representations from their semantic tuning. Second, we compare variation across participants with variation across languages, thereby placing bilingual differences in the context of individual differences. Third, this comparison reveals opposite patterns: semantic tuning is more similar across participants within a language, whereas cortical location is more consistent across languages within a participant. This shows that semantic tuning of representations can be shared across participants even when their cortical location is participant-specific. Chen et al.’s within-participant comparisons could not reveal this pattern.

  2. Using the present data, we cannot distinguish the Chinese-English explanation from the nativeness explanation since all participants are L1 Chinese - L2 English speakers. While it is possible that a specific language property of Chinese or English can explain the difference in cross-participant semantic tuning similarity, we could not find a specific hypothesis to test this possibility. Nevertheless, we agree with the reviewer that our nativeness interpretation of the results is not explicitly tested. We are currently preparing to collect data on other L1 - L2 language pairs on a different set of participants, which will allow us to explicitly test this. However, due to the time-consuming nature of fMRI experiments (~8hrs per participant, across days), data collection will take months. For now, we have clarified this limitation in two places in the revised manuscript. We added “Moreover, all our participants are L1 Chinese - L2 English speakers, so the present study cannot distinguish effects of nativeness from differences specific to Chinese or English.” to Lines 478-480 in the Discussion and “Because all participants were L1 Chinese - L2 English speakers, we cannot determine whether the higher cross-participant semantic tuning similarity observed for Chinese with mBERT reflects nativeness or differences specific to Chinese or English.” to Lines 503-506 in the Limitations.

  3. We thank the reviewer for raising this question about the relationship between cortical organization and semantic tuning. For each participant, we computed mean cortical overlap and mean semantic tuning similarity with the other participants within each language, averaging across languages and clusters. The Pearson correlations across the six participants were r=0.698, p=0.123 (mBERT) and r=-0.286, p = 0.583 (fastText). This exploratory analysis was inconclusive. Given the small sample and the comparisons shared across participant averages, this leaves open whether participants with more unusual cortical maps also have more unusual semantic tuning.

  4. There is no additional evidence linking the effect to contextual meaning beyond the fact that one model is contextual and the other is lexical. We agree that this model comparison alone does not establish contextual processing as the explanation. To test this hypothesis, we plan to repeat the analysis with mBERT embeddings extracted from the same layer, but with no preceding words fed as context. We predict that in this setting, the L1 Chinese - L2 English difference in cross-participant semantic tuning similarity would be smaller. For now, we have made this limitation more explicit in the manuscript. We added “Because mBERT and fastText differ in several respects, this comparison alone cannot isolate the contribution of contextual processing.” in Lines 476-478 of the Discussion.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.