Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Disentangled Contrastive Learning for Zero-Shot Multilingual Dense Retrieval

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read By separating semantic and linguistic features into two subspaces, this paper claims that zero-shot multilingual dense retrieval improves consistently, with 1.5 MRR@10 and 2.7 nDCG@10 gains over strong English-supervised baselines, using on

desk verdict Promising disentanglement recipe for multilingual retrieval, but the non-parallel consistency loss cannot do what the paper claims, and the evaluation hides several key details. read the letter →

arxiv 2608.02189 v1 pith:4UN5XNQ7 submitted 2026-08-03 cs.IR cs.CL

Chao Huang , Yufeng Chen , Changhao Guan , Guang Yang , Dongze Chen , Kaiyu Huang This is my paper
classification cs.IRcs.CL
keywords multilingualdenseretrievalzero-shotcross-lingualtransferdisentangledrepresentationlearningcontrastivehierarchicalsemanticalignmentlanguagedebiasingsemantic-linguisticdisentanglement
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that zero-shot cross-lingual dense retrieval — searching documents in a language for which no retrieval training data exists — fails partly because the shared multilingual representation entangles semantic content with language-specific surface patterns. The proposed method, disentangled contrastive learning, splits each sentence representation into a semantic subspace and a language subspace, then trains the semantic subspace to be language-invariant via sentence- and token-level cross-lingual contrastive alignment while the language subspace absorbs linguistic variation via same-language pulling and different-language pushing. Jointly optimizing these objectives with the standard retrieval loss yields consistent gains over strong English-supervised baselines — 1.5 MRR@10 points on mMARCO and 2.7 nDCG@10 points on MIRACL — with no multilingual retrieval data. A sympathetic reader would care because the method is a training-time addition to an ordinary multilingual retriever, and its gains appear across language families and resource levels, suggesting the entanglement problem is real and addressable.

What carries the argument

The load-bearing mechanism is a language-semantic disentangler: two trainable projection matrices W_s and W_l that map a shared encoder representation z into a semantic subspace z_s and a linguistic subspace z_l, coupled with a Frobenius-norm orthogonality loss on the batch-level matrices Z_s and Z_l. On top of this, hierarchical semantic alignment supplies contrastive supervision at sentence and token levels using parallel data (with mined hard negatives), and indirect consistency losses for non-parallel monolingual sentences; language debiasing contrastive learning pulls z_l of same-language sentences together and pushes different-language sentences apart. The whole disentanglement loss is

What would settle it

A concrete test: take a set of monolingual sentences that have known translations, measure the distance of each sentence to its true translation before and after training with the consistency loss; if the constraint does not reliably shrink the distance to true translations relative to random sentences, the indirect alignment signal is not doing what the paper claims. Alternatively, ablate only the non-parallel consistency term on a language with no parallel data and check whether genuine-translation distances degrade; if they do not, the claimed indirect transfer is not occurring.

Watch

Extended reading notes

Core claim

The central claim is that the entanglement of semantic and linguistic features in multilingual encoders is a first-order bottleneck for zero-shot retrieval transfer, and that explicitly forcing a split is better than merely encouraging language-agnostic alignment within a single space. Concretely, the paper proposes two projection heads applied to the encoder output: one produces a semantic vector used for ranking, the other captures language-specific information; a batch-level orthogonality loss discourages correlation between them. Semantic alignment is performed hierarchically — contrastive pulls between parallel sentences, token-level pulls on word-alignments, and consistency constraints

Load-bearing premise

The load-bearing premise is that a monolingual sentence with no parallel counterpart should be placed at equal semantic distance from the two sentences of every available parallel pair (the consistency losses in equations 8 and 9); this equality-of-distance assumption is likely false for many sentences and, if violated, the consistency loss could damage the very semantic geometry it is meant to clean.

Editorial extensions

If this is right

  • If the claim holds, any standard multilingual retriever can gain zero-shot cross-lingual ability simply by adding the two projection heads and the three auxiliary losses during English fine-tuning; no target-language retrieval labels are needed.
  • The consistent gains on languages that have no parallel data with English suggest indirect semantic consistency transfers alignment from parallel-covered languages to uncovered ones, not just within the covered set.
  • The measured saturation after roughly 40K parallel sentence pairs implies that moderate parallel resources suffice; additional alignment data gives diminishing returns.
  • Because both sentence-level and token-level alignment improve results, fine-grained token alignment is a complementary signal, not a redundancy — relevant for designing future multilingual retrievers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the entanglement story is general, the same two-subspace split should transfer to other cross-lingual semantic tasks such as retrieval-augmented generation or cross-lingual question answering — a testable extension the paper does not run.
  • Because the consistency loss assumes equal distance to both halves of a parallel pair, the method implicitly biases the semantic subspace toward a uniform-distance geometry; probing whether genuine translation distances are preserved would reveal how much of the gain is alignment versus repression of linguistic cues.
  • A practical extension would be to train a lightweight language classifier on the language subspace z_l as a cheap diagnostic of how completely the disentangler absorbs linguistic variation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. This paper proposes a disentangled contrastive learning (DCL) framework for zero-shot multilingual dense retrieval. The method splits a multilingual encoder's sentence representation into a semantic subspace and a linguistic subspace using two linear projections, imposes a batch-level orthogonality loss between the subspaces, and jointly optimizes a retrieval loss with disentanglement objectives. The disentanglement objectives include hierarchical semantic alignment (sentence-level contrastive learning with hard negatives, token-level alignment using word-alignments, and a non-parallel consistency loss over monolingual sentences) and a language debiasing contrastive loss that groups same-language representations. Experiments on mMARCO and MIRACL report consistent gains over English-supervised and multilingual-supervised baselines, with ablations and analyses across resource levels, language families, and parallel-data scales. The paper claims that separating semantic and linguistic features improves zero-shot cross-lingual transfer, particularly to languages without direct parallel data.

Significance. If the reported gains are robust, the paper would make a useful practical contribution: a training recipe for multilingual dense retrieval that requires no multilingual retrieval labels, while improving over strong baselines on both in-domain and out-of-domain benchmarks. The strengths of the paper are its clear overall framing, the inclusion of multiple ablations (Table 3, Table 9), the consistency of gains across two backbone models, the analysis by resource level and language family, and the supervision-scale experiment in Figure 2. The t-SNE visualization provides qualitative evidence that the semantic projection is more language-mixed than the encoder output. However, the paper's central mechanism for transferring alignment to languages without parallel data is not justified by the equations as written, several key hyperparameters are missing, and the evaluation protocol does not fully establish that the reported out-of-domain gains are not affected by hyperparameter selection on the evaluation set. These issues are load-bearing for the strongest transfer claims, although they are addressable in a revision.

major comments (4)
  1. [Methodology, 'Hierarchical Semantic Alignment'; Eqs. (8)-(9)] The non-parallel consistency loss is claimed to 'transfer the aligned structure to languages without parallel data.' As written, however, it cannot provide any absolute semantic pull. For one triple, the two terms combine to -log[δ(s_k,s_i)δ(s_k,s_j)/(δ(s_k,s_i)+δ(s_k,s_j))^2]. If δ(s_k,s_i)=δ(s_k,s_j), the loss is identically log 4 and has zero gradient with respect to a common scaling of both similarities; more generally the gradient depends only on the ratio of δ_i to δ_j. The loss can equalize distances but cannot move a monolingual representation toward the semantic content of a parallel pair. The Table 2 improvements on languages without parallel data are therefore not explained by the stated mechanism. Moreover, the ablation in Table 3 is run on mMARCO, where every language has EN–X parallel data, and no experiment sets λ_q=0 for the MIRACL no-parallel group. Please either modify
  2. [Experiments, 'Implementation Details'; Eqs. (4)-(5)] The contrastive loss in Eq. (4) depends on a temperature τ, and the hard-negative mining in Eq. (5) uses TopK(...,K), but neither τ nor K is reported anywhere. The projection dimension d' of W_s and W_l is also omitted. Combined with the absence of released code, these omissions prevent reproduction of the central empirical claims. Please supply the missing hyperparameter values, and ideally release code or a full configuration file.
  3. [Experiments, 'Implementation Details' and 'Sensitivity Analysis' (Table 8)] The loss weights λ_d, λ_o, λ_q, λ_t are said to be set 'based on a simple grid search.' Table 8 reports sensitivity directly on MIRACL nDCG@10. If the grid search was performed on the MIRACL development set used for the out-of-domain evaluation, then the MIRACL results are not zero-shot in the hyperparameter-selection sense, and the claimed out-of-domain advantage is partly a selection artifact. Please state explicitly which dataset/split was used for hyperparameter selection and, if needed, use a separate validation split or nested selection protocol.
  4. [Experiments, Tables 1-3 and Appendix A] The text states that each reported result is averaged over three independent runs and marks statistical significance with paired t-tests, but no standard deviations, confidence intervals, or per-run values are given in any table. Because the headline gains are often only 0.4–1.5 points, the absence of variance information makes it hard to judge whether the marked significance is meaningful. Please add standard deviations or per-run results.
minor comments (5)
  1. [Methodology, Eq. (13)] The positive sample l_p in the language debiasing loss is not defined explicitly. Specify how p is chosen among same-language samples and how multiple same-language positives are handled.
  2. [Figure 2] The dashed 'N = 0 baseline' is not described in the text. Clarify what is trained when N=0—whether no parallel pairs are used and whether the non-parallel consistency loss is active in that setting.
  3. [Table 3] The abbreviations 'w/ Dis.', 'w/ HSA', and 'w/ LDCL' are not expanded in the caption. A sentence explaining the three components and the 'Δ' columns would improve readability.
  4. [Table 4] The construction of 'Random language pairs' versus 'English-centric pairs' is not defined precisely. Please describe how parallel pairs are sampled for each strategy, particularly whether random pairs are still semantically equivalent translations.
  5. [Appendix, 'Baseline Details'] BGEHN is a prior paper by the same research group. The comparison is legitimate, but the text should flag this explicitly when presenting it among baselines, especially because gains over BGEHN are a stated strength.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the method is trained on external corpora and evaluated on held-out retrieval benchmarks.

full rationale

The paper's core derivation is not circular. The disentangler (Eq. 1), orthogonal loss (Eq. 2), hierarchical semantic alignment (Eqs. 4–11), language debiasing contrastive loss (Eq. 13), and joint objective (Eq. 15) are explicit losses defined on external data: MS MARCO retrieval supervision, WikiMatrix parallel sentences, and Wikipedia monolingual text. No component is defined in terms of the mMARCO/MIRACL metrics reported later. The zero-shot generalization claims are evaluated against held-out external benchmarks and multiple strong baselines, so the results are not forced by construction. The self-cited BGEHN baseline (Huang et al. 2025) appears only as a comparison system and is not used to justify any methodological premise; it is therefore not load-bearing. The grid search over loss weights is a hyperparameter-selection concern, not a definitional reduction: the reported metrics are empirical outcomes, not identities. The skeptic's mechanistic critique of the non-parallel consistency loss (Eqs. 8–9) is a correctness/falsifiability concern rather than circularity: even if the loss only equalizes distances and cannot provide an absolute semantic pull, that does not make the derivation equivalent to its inputs. Because no equation reduces a predicted result to an input quantity and no load-bearing premise rests on a self-citation chain, no circularity is found.

Assumptions & free parameters 5 free parameters · 5 assumptions · 2 invented entities

The paper's central claim rests on a set of modeling choices that are not derived from first principles: linear disentanglement, orthogonality as independence, and a questionable equal-distance consistency objective. Several important hyperparameters are unreported or tuned on the evaluation benchmarks, and the two latent subspaces are internal constructs with no independent falsifiable evidence beyond the paper's own experiments.

free parameters (5)
  • Loss weights lambda_d, lambda_o, lambda_q, lambda_t = 0.05, 0.2, 0.1, 0.2
    Chosen by 'simple grid search' and is likely tuned on the evaluation benchmarks; sensitivity analysis is reported on MIRACL dev. The exact validation split is not disclosed.
  • Temperature tau in contrastive losses (Eq. 4) = not reported
    Appears in all contrastive losses but its value is never specified, which affects the hardness of the contrastive signal and is necessary for replication.
  • Hard-negative count K (TopK in Eq. 5) = not reported
    The paper defines TopK hard negative mining but does not state K, a parameter that directly controls the contrastive loss.
  • Projection dimension d' for Ws and Wl = not reported
    The dimensions of the semantic and language subspaces are not given; without this, the architecture cannot be reproduced exactly.
  • Parallel corpus sample size per language = 40,000
    The number of WikiMatrix pairs sampled per language is a manually chosen training data quantity; performance saturates around 40K but the choice is ad hoc.
assumptions (5)
  • ad hoc to paper Shared multilingual encoder representations can be decomposed into semantic and linguistic factors by two linear projection matrices (Eq. 1).
    This linear separability assumption is the core of the method but is not derived or independently verified; if false, the disentangler cannot achieve its aim.
  • ad hoc to paper Minimizing the batch-level Frobenius norm of Z_s^T Z_l (Eq. 2) is sufficient to separate semantic and linguistic subspaces.
    Orthogonality of projected batch vectors is used as a proxy for semantic independence; this is a standard but unproven approximation.
  • ad hoc to paper The non-parallel consistency loss (Eqs. 8-9) correctly teaches language-agnostic semantics by encouraging equal distances to parallel pairs.
    This assumption is not justified and is the weakest modeling choice in the paper; a monolingual sentence need not be equally similar to both sides of any arbitrary parallel pair.
  • domain assumption Awesome-align (Dou and Neubig 2021) provides token alignments of sufficient quality to train token-level cross-lingual alignment on WikiMatrix pairs.
    The token-level losses rely on automatic word alignments, which can be noisy for low-resource languages and may propagate errors.
  • domain assumption In-batch negatives and TopK hard negatives approximate the true hard-negative distribution for retrieval.
    This is a standard assumption in dense retrieval training, but it is not validated for the multilingual zero-shot setting here.
invented entities (2)
  • Semantic subspace Z_s (with projection head W_s)
    purpose: Captures retrieval-relevant, language-invariant semantic features used for query-document matching.
    The subspace is a new latent construct introduced by the paper. The only evidence for its semantic nature is t-SNE visualization and downstream retrieval gains, both generated inside the paper, not an independent falsifiable handle.
  • Linguistic subspace Z_l (with projection head W_l)
    purpose: Absorbs language-specific variations so they do not interfere with semantic matching.
    A new latent construct; its language-discriminative behavior is enforced by the language debiasing contrastive loss by construction, so this does not constitute independent evidence of true language-feature separation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangled Contrastive Learning for Zero-Shot Multilingual Dense Retrieval." pith.science (2026). https://pith.science/paper/4UN5XNQ7

@misc{pith2026260802189,
  author       = {Pith},
  title        = {Pith review of: Disentangled Contrastive Learning for Zero-Shot Multilingual Dense Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UN5XNQ7}},
  note         = {Machine review of arXiv:2608.02189}
}
read the original abstract

Multilingual dense retrieval aims to handle queries and documents across different languages based on a unified retriever model. The challenge lies in enabling robust retrieval transfer to low-resource languages where annotated retrieval data is often scarce. Although previous studies transfer high-resource supervision to low-resource languages in multilingual semantic representation learning, the shared representation often entangles semantic and linguistic features, which may interfere with optimizing semantic relevance for retrieval. Different from existing methods that focus on learning language-agnostic semantic features under such entanglement, we propose a disentangled contrastive learning~(DCL) method for multilingual dense retrieval by separating multilingual representations into semantic and linguistic subspaces. Specifically, we design disentangled optimization objectives based on hierarchical semantic alignment and language debiasing contrastive learning. By aligning retrieval-relevant semantics across languages at both sentence and token levels while capturing language-specific variations in the linguistic subspace, these objectives reduce language-induced interference in semantic matching. We jointly optimize them with the retrieval objective to facilitate stable zero-shot transfer from English supervision to multilingual dense retrieval. Extensive experiments on mMARCO and MIRACL show that our method consistently outperforms several strong baselines, demonstrating its effectiveness and generalization ability.

Figures

Figures reproduced from arXiv: 2608.02189 by the authors.

Figure 1
Figure 1. Overview of our disentangled framework, including (i) a language-semantic disentangler separating semantic and [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Performance on MIRACL with average nDCG@10 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. t-SNE visualization of different representation spaces. The subfigure titles identify the model modules, whose outputs [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: • MIRACL (Zhang et al. 2023b): a multilingual retrieval benchmark dataset covering 18 languages, with 726k manual relevance judgments and a collection size of over 100 million documents. Each query is provided with an average of 10 manually verified relevance labels. T…
Figure 4
Figure 4. Figure 4: Number of EN–X parallel sentence pairs in WikiMatrix across the 13 non-English languages used for cross-lingual [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 9 linked inside Pith

  1. [1]

    arXiv preprint arXiv:1611.09268 , year=

    MS MARCO: A human generated machine reading comprehension dataset , author=. arXiv preprint arXiv:1611.09268 , year=

    arXiv

  2. [2]

    Proceedings of the 16th conference of the European Chapter of the Association for Computational Linguistics: Main volume , pages=

    Wikimatrix: Mining 135m parallel sentences in 1620 language pairs from wikipedia , author=. Proceedings of the 16th conference of the European Chapter of the Association for Computational Linguistics: Main volume , pages=

  3. [3]

    arXiv preprint arXiv:2108.13897 , year=

    mmarco: A multilingual version of the ms marco passage ranking dataset , author=. arXiv preprint arXiv:2108.13897 , year=

    arXiv

  4. [4]

    Transactions of the Association for Computational Linguistics , volume=

    Miracl: A multilingual retrieval dataset covering 18 diverse languages , author=. Transactions of the Association for Computational Linguistics , volume=. 2023 , publisher=

    2023

  5. [5]

    arXiv preprint arXiv:2203.05765 , year=

    Tevatron: An efficient and flexible toolkit for dense retrieval , author=. arXiv preprint arXiv:2203.05765 , year=

    arXiv

  6. [6]

    2009 , publisher=

    The probabilistic relevance framework: BM25 and beyond , author=. 2009 , publisher=

    2009

  7. [7]

    arXiv preprint arXiv:2112.09118 , year=

    Unsupervised dense information retrieval with contrastive learning , author=. arXiv preprint arXiv:2112.09118 , year=

    arXiv

  8. [8]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    Colbert-xm: A modular multi-vector representation model for zero-shot multilingual information retrieval , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

Show all 40 references
  1. [9]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    mAggretriever: A simple yet effective approach to zero-shot multilingual dense retrieval , author=. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing , pages=

    2023

  2. [10]

    arXiv preprint arXiv:2212.03533 , year=

    Text embeddings by weakly-supervised contrastive pre-training , author=. arXiv preprint arXiv:2212.03533 , year=

    arXiv

  3. [11]

    arXiv preprint arXiv:2402.03216 , volume=

    Bge m3-embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation , author=. arXiv preprint arXiv:2402.03216 , volume=

    arXiv

  4. [12]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Boosting data utilization for multilingual dense retrieval , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  5. [13]

    arXiv preprint arXiv:1912.01703 , year=

    Pytorch: An imperative style, high-performance deep learning library , author=. arXiv preprint arXiv:1912.01703 , year=

    1912 arXiv

  6. [14]

    arXiv preprint arXiv:1910.03771 , year=

    Huggingface's transformers: State-of-the-art natural language processing , author=. arXiv preprint arXiv:1910.03771 , year=

    1910 arXiv

  7. [15]

    Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

    Colbert: Efficient and effective passage search via contextualized late interaction over bert , author=. Proceedings of the 43rd International ACM SIGIR conference on research and development in Information Retrieval , pages=

  8. [16]

    ACM Transactions on Information Systems , volume=

    Toward best practices for training multilingual dense retrieval models , author=. ACM Transactions on Information Systems , volume=. 2023 , publisher=

    2023

  9. [17]

    Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

    Soft prompt decoding for multilingual dense retrieval , author=. Proceedings of the 46th international ACM SIGIR conference on research and development in information retrieval , pages=

  10. [18]

    Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    Leveraging LLMs for synthesizing training data across many languages in multilingual dense retrieval , author=. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers...

    2024

  11. [19]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Boosting zero-shot cross-lingual retrieval by training on artificially code-switched data , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  12. [20]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  13. [21]

    Advances in neural information processing systems , volume=

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

  14. [22]

    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=

  15. [23]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    InfoXLM: An information-theoretic framework for cross-lingual language model pre-training , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  16. [24]

    Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Language-agnostic BERT sentence embedding , author=. Proceedings of the 60th annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  17. [25]

    Proceedings of the 2022 conference on empirical methods in natural language processing , pages=

    English contrastive learning can learn universal cross-lingual sentence embeddings , author=. Proceedings of the 2022 conference on empirical methods in natural language processing , pages=

    2022

  18. [26]

    Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    ERNIE-M: Enhanced multilingual representation by aligning cross-lingual semantics with monolingual corpora , author=. Proceedings of the 2021 conference on empirical methods in natural language processing , pages=

    2021

  19. [27]

    arXiv preprint arXiv:2002.03518 , year=

    Multilingual alignment of contextual word representations , author=. arXiv preprint arXiv:2002.03518 , year=

    2002 arXiv

  20. [28]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Disentangling semantics and syntax in sentence embeddings with pre-trained language models , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2021

  21. [29]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Learning disentangled semantic representations for zero-shot cross-lingual transfer in multilingual machine reading comprehension , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  22. [30]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Learning disentangled semantic spaces of explanations via invertible neural networks , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  23. [31]

    2010 , publisher=

    Cross-language information retrieval , author=. 2010 , publisher=

    2010

  24. [32]

    Cross-lingual alignment of contextual word embeddings, with applications to zero-shot dependency parsing , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long an...

    2019

  25. [33]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Multilingual LLMs are better cross-lingual in-context learners with alignment , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  26. [34]

    Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    Understanding cross-lingual Alignment—A survey , author=. Findings of the Association for Computational Linguistics: ACL 2024 , pages=

    2024

  27. [35]

    Advances in Neural Information Processing Systems , volume=

    Non-linguistic supervision for contrastive learning of sentence embeddings , author=. Advances in Neural Information Processing Systems , volume=

  28. [36]

    European Conference on Information Retrieval , pages=

    Teaching a new dog old tricks: Resurrecting multilingual retrieval using zero-shot learning , author=. European Conference on Information Retrieval , pages=. 2020 , organization=

    2020

  29. [37]

    A multi-task approach for disentangling syntax and semantics in sentence representations , author=. Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) , pages=

    2019

  30. [38]

    IEEE Transactions on Knowledge and Data Engineering , volume=

    Causal disentanglement for semantic-aware intent learning in recommendation , author=. IEEE Transactions on Knowledge and Data Engineering , volume=. 2022 , publisher=

    2022

  31. [39]

    IEEE Transactions on image processing , volume=

    Semantics disentangling for cross-modal retrieval , author=. IEEE Transactions on image processing , volume=. 2024 , publisher=

    2024

  32. [40]

    Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume , pages=

    Word alignment by fine-tuning embeddings on parallel corpora , author=. Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume , pages=

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.