Pith. sign in

REVIEW 4 major objections 6 minor 89 references

A language model's internal language tree closely matches the Indo-European family tree, and the strength of that match predicts its cross-lingual accuracy.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review

2026-08-01 19:55 UTC pith:TBOO52FM

load-bearing objection A careful tree-recovery measurement that overreaches on its headline correlation, but deserves a serious referee because the method and the honest reporting are solid. the 4 major comments →

arxiv 2607.22699 v1 pith:TBOO52FM submitted 2026-07-18 cs.AI

Similarity All The Way Up: Multilingual Generalization in LLMs Relies on Language-Level Similarity Structures

Supantho Rakshit , Adele Goldberg , Henry Conklin This is my paper
classification cs.AI
keywords multilingual generalizationsimilarity structureIndo-European language treeXNLIrepresentation geometrycophenetic correlationpath kernelcross-lingual transfer
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper argues that a language model's ability to generalize from one language to another depends on how its internal representation space organizes languages: models that place languages from the same subfamily close together transfer better. Analyzing hidden states from twelve LLMs, the authors show that latent representations largely recover the hierarchical structure of the Indo-European family tree without explicit supervision. The strength of this recovery correlates with performance on the multilingual XNLI benchmark: Spearman rho = 0.81 for cophenetic correlation and 0.79 for path kernel across eight models. The authors interpret this as an extension of the classic cognitive-science idea that generalization is governed by similarity structure, now operating at the scale of whole languages. Why it matters: if correct, it gives a representational diagnostic for why some models generalize multilingually and others do not, and points to hierarchical similarity — not just pairwise lexical overlap — as the operative quantity.

Core claim

On the paper's own terms, the central discovery is that large language models implicitly encode the hierarchical similarity relations among languages — grouping Germanic, Romance, Slavic, Indo-Aryan, and Baltic languages into coherent clusters in embedding space — and that the fidelity of this encoding, measured by tree-distance statistics, is significantly correlated with zero-shot cross-lingual performance on XNLI. The authors report that the global measures (cophenetic correlation and path kernel) but not the local subtree kernel predict performance, indicating that preserving broad subfamily organization matters more than exact nearest-neighbor matches. They further show that explicitly

What carries the argument

The load-bearing machinery is a pipeline that turns hidden states into language-level trees. For each of 38 Indo-European languages, the method estimates a conditional distribution over a random codebook on the unit hypersphere (a soft-density estimator), aggregates these distributions over 7,500 sentences per language, computes pairwise Jensen-Shannon divergences per layer, averages them, and applies UPGMA clustering to obtain a dendrogram. This dendrogram is then compared to a reference tree derived from a Bayesian phylogenetic reconstruction of Indo-European, using three tree-similarity measures: cophenetic correlation (preservation of pairwise merging heights), path kernel (shared root-t

Load-bearing premise

The headline correlation rests on eight models whose XNLI accuracy values and evaluation protocol are never reported in the paper, and model size is not controlled for; if the accuracy numbers were produced under inconsistent conditions, or if the correlation disappears once parameter count is partialled out, the claim that similarity structure drives multilingual performance collapses.

What would settle it

Obtain the XNLI accuracy for each of the eight models and recompute the Spearman correlation between tree-alignment and accuracy while controlling for parameter count; alternatively, re-run the analysis using a different reference tree for Indo-European (e.g., one based on typological features rather than lexical cognates). A non-significant partial correlation or a failure to replicate with a different reference tree would settle the claim.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the correlation is causal, then increasing a model's alignment with language-family structure should improve its cross-lingual transfer; this can be tested by regularizing training toward such structure.
  • Global hierarchical organization (subfamily groupings) matters more for multilingual performance than exact nearest-neighbor relationships, so evaluations and diagnostics should focus on broad structure rather than pairwise accuracy.
  • Multilingual pretraining data induces the relevant similarity structure, while monolingual training does not; adding more languages during pretraining may be the most direct route to better cross-lingual generalization.
  • Language-level structure emerges very early in pretraining, suggesting that the representational foundation for cross-lingual transfer is set early; interventions after that point may have limited effect.
  • The approach offers a new way to measure multilingual capability without running a downstream task: compare a model's language tree to a reference phylogeny.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported correlation uses only eight models spanning a wide range of sizes; a controlled study that varies parameter count at fixed architecture and data would tell whether the effect is really about similarity structure or just about model scale.
  • The recurring misplacement of independent languages and the unexplained Greek–Welsh pairing suggest that tokenization and orthography modulate the signal; one could test this by re-running on transliterated corpora.
  • The method should transfer to non-Indo-European families (e.g., Niger-Congo, Austronesian); if the correlation with cross-lingual performance holds there, the principle would be much more general than the Indo-European case.
  • One could use the tree-alignment scores as a zero-shot predictor of a model's downstream multilingual performance, potentially reducing the need for expensive benchmarks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper investigates whether LLMs' latent representations of languages reflect the hierarchical similarity structure of the Indo-European language family, and whether this structural alignment predicts cross-lingual generalization. Using a soft-density estimator over token embeddings, the authors construct per-model dendrograms for 38 IE languages and compare them to a reference tree derived from Heggarty et al. (2023) via cophenetic correlation, subtree kernel, and path kernel. They report that LLMs largely recover the reference tree and that, across eight LLMs, cophenetic correlation and path kernel alignment correlate with XNLI accuracy (rho=0.81, p=0.02 and rho=0.79, p=0.02), while subtree kernel does not. They also report that multilingual variants (mBERT, XLM-R) align better than monolingual counterparts, and that language-level structure emerges early in Pythia pretraining. The central claim is that hierarchical similarity structure, not merely pairwise similarity, is associated with multilingual generalization.

Significance. If the central correlation is robust, the paper offers a meaningful bridge from classic similarity-based theories of generalization to modern LLM behavior, and it goes beyond prior pairwise analyses by quantifying hierarchical structure. The tree-recovery measurement is thoughtfully designed, with multiple complementary metrics, and the authors honestly report systematic failures (e.g., the Greek–Welsh pairing, the Slavic/Baltic split, the null subtree-kernel result). The monolingual/multilingual matched comparisons (BERT vs. mBERT; RoBERTa vs. XLM-R) are a nice controlled element. However, the headline claim that similarity structure 'relies on' or drives generalization rests on a correlation across only eight models, with no disclosed XNLI evaluation protocol and no control for model scale. These are load-bearing gaps that currently prevent the claim from being accepted as stated.

major comments (4)
  1. [Methods, 'Measuring LLMs' ability to generalize across languages: XNLI'] The XNLI evaluation protocol is not described. The text says the benchmark is used to quantify zero-shot transfer, but it never states whether models were fine-tuned on English XNLI and evaluated in other languages, whether a linear probe or full fine-tuning was used, which of the 15 XNLI languages were evaluated (the tree covers 38 languages), or any hyperparameters. Figure 2 plots 'XNLI accuracy percentage' but the per-model accuracy values are never enumerated. Without this information the headline correlations in Figure 2 cannot be interpreted or reproduced. Please provide a complete protocol and a table of per-model accuracies.
  2. [Results & Discussion, Figure 2] The eight LLMs span orders of magnitude in scale, from Pythia-160m/410m to XLM-R and mT5. Larger models generally achieve higher XNLI accuracy and may also produce representations that align better with a reference phylogeny for reasons unrelated to the specific hypothesis. The paper does not report a partial correlation controlling for log parameter count (or pretraining compute), and the size-matched BERT/mBERT and RoBERTa/XLM-R pairs, which are the strongest evidence for a specific structural effect, are excluded from the headline correlation. Please report partial correlations and/or a matched analysis. Without this, rho=0.81 does not establish that similarity structure per se drives generalization.
  3. [Results, 'Similarity Structure Correlates with Performance'] With n=8 models that are not independent (two BERT variants, two RoBERTa variants, two Pythia sizes, mT5, XLM), the Spearman p=0.02 is fragile and sensitive to a single observation. The paper also reports a null result for the subtree kernel (rho=0.383, p=0.349), which is discussed qualitatively but not quantitatively. Provide per-model data points, bootstrap confidence intervals for the correlations, and ideally a permutation test based on shuffling model labels. This will establish whether the correlation is robust to the specific set of models chosen.
  4. [Methods, 'LLM Density Estimation', Eq. (1)] The density estimator depends on several free hyperparameters: temperature tau=0.1, codebook size K=3d, number of sentences N=7500, and averaging over layers. No sensitivity analysis is provided. Since all tree metrics derive from this estimator, the headline correlations could change with different hyperparameter choices. Please report the robustness of the Figure 2 correlations to a reasonable range of tau, K, and N, and to the random codebook sampling, at least for a subset of models.
minor comments (6)
  1. [General formatting] The text suffers from missing spaces between words throughout (e.g., the abstract and introduction). This is a typesetting/presentation issue that should be corrected.
  2. [Figure 2] Each data point should be labeled with the model name and its parameter count or scale. Currently the reader cannot see which point corresponds to which model, making it harder to assess the scale-confounding concern.
  3. [Methods, Eq. (5)] The subtree-isomorphism symbol is garbled ('subtree(n1) subtree(n2)'). Use a standard congruence or isomorphism symbol and define it clearly.
  4. [Methods, 'Language sample'] 'independent IE languages([3]: Greek, Armenian, and Albanian)' appears to contain a citation artifact. It should read 'three independent IE languages: Greek, Armenian, and Albanian'.
  5. [Methods, 'LLMs, Data, and Distances'] The manuscript says '12 LLMs' in the Methods opening, but Figure 2 and the performance analysis use eight. Clarify that tree recovery was computed for 12 models but the XNLI correlation used a subset of eight, and explain the selection criterion.
  6. [References and notation] The terms 'mBERT' and 'multilingual BERT' are used interchangeably; define the abbreviation once. Also, the citation to 'Conklin, H. C. (2025)' is a thesis; specify if it is publicly available and, if not, provide the method details in an appendix or supplement.

Circularity Check

0 steps flagged

No significant circularity: the structural measures, reference tree, and performance benchmark are external inputs, and no fitted constant reappears in the headline claim.

full rationale

The paper's derivation chain is non-circular. The reference tree is taken from Heggarty et al. (2023), an external Bayesian phylogeny; the authors explicitly treat it as an independent reference ('we adopt the tree structure induced from a statistical model, namely (Heggarty et al. (2023)), as an independent reference tree'). The representational trees are built from LLM hidden states using the soft-density estimator from Conklin (2025) with hand-chosen hyperparameters (tau=0.1, K=3d, 7500 sentences); these are not fitted to or derived from XNLI scores. The XNLI benchmark provides external performance data. The headline correlations (Cophenetic rho=0.81, Path rho=0.79, Subtree rho=0.383) are observed correlations between model-derived tree-alignment and benchmark accuracy, not quantities that are equal by construction. The single self-citation (Conklin 2025) supplies a measurement tool, not the target result; its parameters are not optimized on XNLI, and no uniqueness or optimality theorem is invoked. Concerns about n=8, model-scale confounding, missing XNLI protocol details, and the contested reference tree are correctness/robustness risks, not circularity.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper introduces no new natural-kind entities. Its free parameters are estimator hyperparameters chosen by hand; they are not fitted to the target (XNLI), which keeps the circularity burden low. The load-bearing assumptions are the validity of the contested reference tree, the transfer of the cognitive-science similarity principle to language-level generalization, and the fidelity of the self-cited density estimator.

free parameters (4)
  • softmax temperature tau = 0.1
    Hand-chosen temperature in Eq. 1 controlling the soft assignment of token embeddings to codebook entries; affects all language distribution estimates and hence every tree-alignment score.
  • codebook size K = 3 * hidden_dim
    Chosen to equalize expressive capacity across architectures (K=3d); no sensitivity analysis is reported.
  • sentences per language N = 7500
    Sample size for each language's density estimate from mC4; stability of the resulting trees in N is not analyzed.
  • layer averaging over L layers = uniform mean (1/L)
    Eq. 2 averages pairwise Jensen-Shannon divergences equally across all layers; this weighting is a modeling choice that is not varied or justified.
axioms (4)
  • domain assumption Heggarty et al. (2023) MCC tree, via patristic distances re-clustered by UPGMA, is a valid independent reference for language similarity hierarchy.
    The paper adopts a contested phylogenetic reconstruction (Kassian & Starostin 2025 criticized it) as the similarity reference, and UPGMA re-clustering distorts the original branch lengths.
  • domain assumption Shepard/Tenenbaum similarity-driven generalization transfers from perceptual categories to language-level generalization in LLMs.
    The Introduction's theoretical frame assumes the cognitive-science principle scales up to whole-language generalization; the paper only tests the correlational version of this.
  • domain assumption The soft-density estimator (Conklin 2025) faithfully represents language-level similarity in embedding space.
    This is a self-cited estimator; its fidelity relative to simpler baselines (e.g., token-overlap or script-based distances) is not validated in this paper.
  • domain assumption XNLI zero-shot accuracy is a valid measure of cross-lingual generalization.
    The paper acknowledges XNLI is somewhat unnatural (Weissweiler et al. 2025), and the evaluation protocol is unspecified, so the accuracy values are taken as given.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Similarity All The Way Up: Multilingual Generalization in LLMs Relies on Language-Level Similarity Structures." pith.science (2026). https://pith.science/paper/TBOO52FM

@misc{pith2026260722699,
  author       = {Pith},
  title        = {Pith review of: Similarity All The Way Up: Multilingual Generalization in LLMs Relies on Language-Level Similarity Structures},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TBOO52FM}},
  note         = {Machine review of arXiv:2607.22699}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

As Large Language Models (LLMs) grow more capable across diverse tasks, their (in)ability to generalize remains difficult to quantify and poorly understood beyond limited domains. In particular, LLMs are known to struggle generalizing multilingually, to languages outside of English, and that are poorly attested in their training data. To understand why this may be, and what enables some models to perform better than others, we turn to a long history of work across the cognitive sciences, arguing that successful generalization derives from appropriate representations in similarity space. We look at how well LLMs' representations capture the hierarchical similarity structure between distinct languages. Strikingly, we show LLMs' latent representations largely recover the hierarchical structure of the Indo-European language family tree -- grouping languages that are members of the same subfamily closely together in representation space. Furthermore, we show that the degree to which models reflect the similarity structure of languages correlates with their performance on XNLI, a multilingual natural language inference benchmark. This extends classic work on similarity-driven generalization at scale, showing how models that represent similar languages similarly generalize better from one language to another.

Figures

Figures reproduced from arXiv: 2607.22699 by Adele Goldberg, Henry Conklin, Supantho Rakshit.

Figure 1
Figure 1. Figure 1: LLMs Implicitly Recover Relations Between Languages: Similarity structure among 38 Indo-European languages, as derived from XLM-Roberta’s representations (Left) and reconstructed from Heggarty et al. (2023) (Right). Colors represent phylogenetic language subfamilies, with a legend in the top right. across languages, with models converging to representations that are not entirely specific to a single langua… view at source ↗
Figure 2
Figure 2. Figure 2: Correlating Structure with Performance: Each facet shows one of three tree-distance measures (x-axis) plotted against XNLI accuracy percentage (y-axis) for eight LLMs. Above each facet is the Spearman rank correlation 𝜌 between axes. Cophenetic Correlation and Path Kernel show a significant positive relationship with performance; both reflect how well LLM representations capture higher-order structures of … view at source ↗
Figure 3
Figure 3. Figure 3: Multilingual vs. English LLMs | Structure Over Pretraining: (Left) y-axis: degree of alignment with the reference tree measured by the cophenetic correlation; x-axis: one of the model families BERT or RoBERTa. Colors correspond to model IDs. All models share the same architecture, number of layers, and parameters, but differ in training data. Models trained on multilingual data have a white x pattern. Mult… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

89 extracted references · 5 linked inside Pith

  1. [1]

    2025 , publisher=

    Information structure in mappings: an approach to learning, representation and generalisation , author=. 2025 , publisher=

    2025

  2. [2]

    2025 , address =

    Hammarström, Harald and Forkel, Robert and Haspelmath, Martin and Bank, Sebastian , title =. 2025 , address =

    2025

  3. [3]

    and Atkinson, Quentin D

    Gray, Russell D. and Atkinson, Quentin D. , title =. Nature , volume =. 2003 , doi =

    2003

  4. [4]

    and Starostin, George , title =

    Kassian, Alexei S. and Starostin, George , title =. Humanities and Social Sciences Communications , volume =. 2025 , doi =

    2025

  5. [5]

    arXiv preprint arXiv:1907.11692 , year=

    Roberta: A robustly optimized bert pretraining approach , author=. arXiv preprint arXiv:1907.11692 , year=

    Pith/arXiv arXiv 1907

  6. [6]

    arXiv preprint arXiv:1905.05950 , year=

    BERT rediscovers the classical NLP pipeline , author=. arXiv preprint arXiv:1905.05950 , year=

    Pith/arXiv arXiv 1905

  7. [7]

    and Alekseyenko, Alexander V

    Bouckaert, Remco and Lemey, Philippe and Dunn, Michael and Greenhill, Simon J. and Alekseyenko, Alexander V. and Drummond, Alexei J. and Gray, Russell D. and Suchard, Marc A. and Atkinson, Quentin D. , title =. Science , volume =. 2012 , doi =

    2012

  8. [8]

    Science , volume=

    Language trees with sampled ancestors support a hybrid model for the origin of Indo-European languages , author=. Science , volume=. 2023 , publisher=

    2023

  9. [9]

    Cognition and Categorization , editor =

    Tversky, Amos and Gati, Itamar , title =. Cognition and Categorization , editor =. 1978 , pages =

    1978

  10. [10]

    and Goldstone, Robert L

    Medin, Douglas L. and Goldstone, Robert L. and Gentner, Dedre , title =. Psychological Review , year =

  11. [11]

    Psychological Review , year =

    Tversky, Amos , title =. Psychological Review , year =

  12. [12]

    , author=

    Respects for similarity. , author=. Psychological review , volume=. 1993 , publisher=

    1993

  13. [13]

    Language and linguistics compass , volume=

    Tutorial on computational linguistic phylogeny , author=. Language and linguistics compass , volume=. 2008 , publisher=

    2008

  14. [14]

    From semantics to dialectometry: Festschrift in honor of John Nerbonne , pages=

    Exploring the role of extra-linguistic factors in defining dialectal variation patterns through cluster comparison , author=. From semantics to dialectometry: Festschrift in honor of John Nerbonne , pages=. 2017 , publisher=

    2017

  15. [15]

    Languages , volume=

    Language Contact and Population Contact as Sources of Dialect Similarity , author=. Languages , volume=. 2025 , publisher=

    2025

  16. [16]

    International conference on machine learning , pages=

    Xtreme: A massively multilingual multi-task benchmark for evaluating cross-lingual generalisation , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  17. [17]

    2005 , publisher=

    The world atlas of language structures , author=. 2005 , publisher=

    2005

  18. [18]

    2019 , note =

    arXiv:1810.04805 [cs] , author =. 2019 , note =

    Pith/arXiv arXiv 2019

  19. [19]

    1925 , publisher =

    Meillet, Antoine , title =. 1925 , publisher =

    1925

  20. [20]

    2006 , publisher =

    Ringe, Don , title =. 2006 , publisher =

    2006

  21. [21]

    2013 , publisher =

    Campbell, Lyle , title =. 2013 , publisher =

    2013

  22. [22]

    2013 , publisher=

    Historical linguistics: An introduction , author=. 2013 , publisher=

    2013

  23. [23]

    Journal of machine learning research , volume=

    Exploring the limits of transfer learning with a unified text-to-text transformer , author=. Journal of machine learning research , volume=

  24. [24]

    Biderman, Stella and Schoelkopf, Hailey and Anthony, Quentin Gregory and Bradley, Herbie and O’Brien, Kyle and Hallahan, Eric and Khan, Mohammad Aflah and Purohit, Shivanshu and Prashanth, Usvsn Sai and Raff, Edward and Skowron, Aviya and Sutawika, Lintang and Wal, Oskar Van Der , month = jul, year =. Pythia:. Proceedings of the 40th

  25. [25]

    arXiv preprint arXiv:2502.13195 , year=

    Linguistic generalizations are not rules: Impacts on evaluation of LMs , author=. arXiv preprint arXiv:2502.13195 , year=

    arXiv

  26. [26]

    Science , volume=

    Toward a universal law of generalization for psychological science , author=. Science , volume=. 1987 , publisher=

    1987

  27. [27]

    Psychometrika , volume=

    Stimulus and response generalization: A stochastic model relating generalization to distance in psychological space , author=. Psychometrika , volume=. 1957 , publisher=

    1957

  28. [28]

    Psychological Review , volume=

    Features of similarity , author=. Psychological Review , volume=. 1977 , publisher=

    1977

  29. [29]

    Generalization, similarity, and

    Tenenbaum, Joshua B and Griffiths, Thomas L , journal=. Generalization, similarity, and. 2001 , publisher=

    2001

  30. [30]

    Cognitive Science , volume=

    Finding structure in time , author=. Cognitive Science , volume=. 1990 , publisher=

    1990

  31. [31]

    2004 , publisher=

    Semantic cognition: A parallel distributed processing approach , author=. 2004 , publisher=

    2004

  32. [32]

    Nature , volume=

    Learning representations by back-propagating errors , author=. Nature , volume=. 1986 , publisher=

    1986

  33. [33]

    On learning the past tenses of

    Rumelhart, David E and McClelland, James L , booktitle=. On learning the past tenses of. 1986 , publisher=

    1986

  34. [34]

    Proceedings of the Eighth Annual Conference of the Cognitive Science Society , pages=

    Learning distributed representations of concepts , author=. Proceedings of the Eighth Annual Conference of the Cognitive Science Society , pages=

  35. [35]

    Journal of Experimental Psychology: General , volume=

    Attention, similarity, and the identification-categorization relationship , author=. Journal of Experimental Psychology: General , volume=. 1986 , publisher=

    1986

  36. [36]

    Psychological Review , volume=

    Context theory of classification learning , author=. Psychological Review , volume=. 1978 , publisher=

    1978

  37. [37]

    1992 , publisher=

    Kruschke, John K , journal=. 1992 , publisher=

    1992

  38. [38]

    Journal of Experimental Psychology , volume=

    On the genesis of abstract ideas , author=. Journal of Experimental Psychology , volume=. 1968 , publisher=

    1968

  39. [39]

    Frontiers in Systems Neuroscience , volume=

    Representational similarity analysis--connecting the branches of systems neuroscience , author=. Frontiers in Systems Neuroscience , volume=. 2008 , publisher=

    2008

  40. [40]

    Trends in Cognitive Sciences , volume=

    Representational geometry: Integrating cognition, computation, and the brain , author=. Trends in Cognitive Sciences , volume=. 2013 , publisher=

    2013

  41. [41]

    Behavioral and Brain Sciences , volume=

    Representation is representation of similarities , author=. Behavioral and Brain Sciences , volume=. 1998 , publisher=

    1998

  42. [42]

    2000 , publisher=

    Conceptual spaces: The geometry of thought , author=. 2000 , publisher=

    2000

  43. [43]

    Cognitive Science , volume=

    Structure-mapping: A theoretical framework for analogy , author=. Cognitive Science , volume=. 1983 , publisher=

    1983

  44. [44]

    Cognitive Science , volume=

    Analogical mapping by constraint satisfaction , author=. Cognitive Science , volume=. 1989 , publisher=

    1989

  45. [45]

    Cognition , volume=

    Connectionism and cognitive architecture: A critical analysis , author=. Cognition , volume=. 1988 , publisher=

    1988

  46. [46]

    Cognitive Psychology , volume=

    Rethinking eliminative connectionism , author=. Cognitive Psychology , volume=. 1998 , publisher=

    1998

  47. [47]

    Proceedings of the 35th International Conference on Machine Learning , pages=

    Generalization without systematicity: On the compositional skills of sequence-to-sequence recurrent networks , author=. Proceedings of the 35th International Conference on Machine Learning , pages=. 2018 , organization=

    2018

  48. [48]

    Nature Communications , volume=

    Capturing human categorization of natural images by combining deep networks and cognitive models , author=. Nature Communications , volume=. 2020 , publisher=

    2020

  49. [49]

    Cognitive Science , volume=

    Evaluating (and improving) the correspondence between deep neural networks and human representations , author=. Cognitive Science , volume=. 2018 , publisher=

    2018

  50. [50]

    International Conference on Learning Representations , year=

    Human alignment of neural network representations , author=. International Conference on Learning Representations , year=

  51. [51]

    Attention is

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N and Kaiser, Łukasz and Polosukhin, Illia , year =. Attention is

  52. [52]

    arXiv preprint arXiv:1309.4168 , year=

    Exploiting similarities among languages for machine translation , author=. arXiv preprint arXiv:1309.4168 , year=

    Pith/arXiv arXiv

  53. [53]

    International Conference on Learning Representations , year=

    Word translation without parallel data , author=. International Conference on Learning Representations , year=

  54. [54]

    How multilingual is multilingual

    Pires, Telmo and Schlinger, Eva and Garrette, Dan , booktitle=. How multilingual is multilingual. 2019 , organization=

    2019

  55. [55]

    Beto, Bentz, Becas: The surprising cross-lingual effectiveness of

    Wu, Shijie and Dredze, Mark , booktitle=. Beto, Bentz, Becas: The surprising cross-lingual effectiveness of. 2019 , organization=

    2019

  56. [56]

    Language trees with sampled ancestors support a hybrid model for the origin of

    Heggarty, Paul and Shimelman, Aviva and Stark, Megan and others , journal=. Language trees with sampled ancestors support a hybrid model for the origin of. 2023 , publisher=

    2023

  57. [57]

    Taxon , volume=

    The comparison of dendrograms by objective methods , author=. Taxon , volume=. 1962 , publisher=

    1962

  58. [58]

    Systematic Zoology , volume=

    On the cophenetic correlation coefficient , author=. Systematic Zoology , volume=. 1969 , publisher=

    1969

  59. [59]

    Science , volume=

    Multidimensional scaling, tree-fitting, and clustering , author=. Science , volume=. 1980 , publisher=

    1980

  60. [60]

    Advances in Neural Information Processing Systems , volume=

    Convolution kernels for natural language , author=. Advances in Neural Information Processing Systems , volume=

  61. [61]

    Journal of Machine Learning Research , volume=

    Graph kernels , author=. Journal of Machine Learning Research , volume=

  62. [62]

    Proceedings of the 2021 conference of the North American chapter of the association for computational linguistics: Human language technologies , pages=

    mT5: A massively multilingual pre-trained text-to-text transformer , author=. Proceedings of the 2021 conference of the North American chapter of the association for computational linguistics: Human language technologies , pages=

    2021

  63. [63]

    arXiv preprint arXiv:2010.12821 , year=

    Rethinking embedding coupling in pre-trained language models , author=. arXiv preprint arXiv:2010.12821 , year=

    Pith/arXiv arXiv 2010

  64. [64]

    Advances in neural information processing systems , volume=

    Cross-lingual language model pretraining , author=. Advances in neural information processing systems , volume=

  65. [65]

    The annals of mathematical statistics , volume=

    On estimation of a probability density function and mode , author=. The annals of mathematical statistics , volume=. 1962 , publisher=

    1962

  66. [66]

    Proceedings of the 17th European Conference on Machine Learning , pages=

    Efficient convolution kernels for dependency and constituent syntactic trees , author=. Proceedings of the 17th European Conference on Machine Learning , pages=. 2006 , organization=

    2006

  67. [67]

    Proceedings of the National Academy of Sciences , volume=

    A mathematical theory of semantic development in deep neural networks , author=. Proceedings of the National Academy of Sciences , volume=. 2019 , publisher=

    2019

  68. [68]

    Proceedings of the 20th International Conference on Machine Learning , pages=

    Marginalized kernels between labeled graphs , author=. Proceedings of the 20th International Conference on Machine Learning , pages=

  69. [69]

    ACM SIGKDD Explorations Newsletter , volume=

    A survey of kernels for structured data , author=. ACM SIGKDD Explorations Newsletter , volume=. 2003 , publisher=

    2003

  70. [70]

    Proceedings of the 25th International Conference on Machine Learning , pages=

    Graph kernels between point clouds , author=. Proceedings of the 25th International Conference on Machine Learning , pages=

  71. [71]

    Advances in Neural Information Processing Systems , volume=

    Generalized shape metrics on neural representations , author=. Advances in Neural Information Processing Systems , volume=

  72. [72]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=

    Unsupervised cross-lingual representation learning at scale , author=. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , pages=. 2020 , organization=

    2020

  73. [73]

    From zero to hero: On the limitations of zero-shot cross-lingual transfer from

    Lauscher, Anne and Ravishankar, Vinit and Vuli. From zero to hero: On the limitations of zero-shot cross-lingual transfer from. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages=. 2020 , organization=

    2020

  74. [74]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=

    Bilingual lexicon induction with semi-supervision in non-isometric embedding spaces , author=. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics , pages=. 2019 , organization=

    2019

  75. [75]

    Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , pages=

    On the limitations of unsupervised bilingual dictionary induction , author=. Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics , pages=. 2018 , organization=

    2018

  76. [76]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages=

    Are all good word vector spaces isomorphic? , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages=. 2020 , organization=

    2020

  77. [77]

    2019 , organization=

    Singh, Jasdeep and McCann, Bryan and Socher, Richard and Xiong, Caiming , booktitle=. 2019 , organization=

    2019

  78. [78]

    Finding universal grammatical relations in multilingual

    Chi, Ethan A and Hewitt, John and Manning, Christopher D , booktitle=. Finding universal grammatical relations in multilingual. 2020 , organization=

    2020

  79. [79]

    Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    From phonology to syntax: Unsupervised linguistic typology at different levels with language embeddings , author=. Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=. 2018 , organization=

    2018

  80. [80]

    Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages=

    Zero-shot cross-lingual transfer with meta-learning , author=. Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing , pages=. 2020 , organization=

    2020

Showing first 80 references.

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.