Pith. sign in

REVIEW 5 major objections 6 minor 42 references

DualAnchor shows that gloss-free sign translation improves when training couples a token-level language-prior anchor with an optimal-transport lexical anchor, reaching the best BLEU-4 among compared gloss-free methods on two benchmarks.

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 · deepseek-v4-flash

2026-08-01 04:35 UTC pith:SP6UDO5F

load-bearing objection Good, honest SLT method paper with two training-only regularizers; the headline SOTA claim is provisional because the main backbone is unnamed and no error bars are given. the 5 major comments →

arxiv 2607.27614 v1 pith:SP6UDO5F submitted 2026-07-30 cs.CL cs.AI

DualAnchor: Preserving Language Priors and Improving Lexical Fidelity in Gloss-Free Sign Language Translation

Hongbin Zhang , Junhao Liu , Xuefeng Bai , Youcheng Pan , Yang Xiang , Kehai Chen This is my paper
classification cs.CL cs.AI
keywords sign language translationgloss-freelarge language modelslanguage prior preservationlexical fidelityoptimal transportSinkhorntoken-level distillation
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 claims that the two persistent failure modes of gloss-free sign-language translation with LLM backbones—loss of fluent language structure during multimodal adaptation, and lexical errors that survive sentence-level video-text alignment—can be fixed by two complementary training-only objectives. The first, Token-level Prior Anchoring, regularizes each next-token distribution of the multimodal decoder toward the distribution of a frozen language model under the same autoregressive prefix, with a confidence gate that weakens the constraint where the prior is uncertain. The second, Optimal Transport Alignment, treats visual-to-content-word matching as entropy-regularized partial optimal transport, so each content word is softly grounded in visual evidence while unreliable tokens are absorbed by a dustbin. On PHOENIX-2014T and CSL-Daily the combined method reports the best BLEU-4 among compared gloss-free systems, and targeted analyses attribute the gains to TPA improving fluency and grammar and OTA reducing content-word errors. A sympathetic reader should care because it separates two failure axes that are usually conflated and addresses both without changing inference.

Core claim

DualAnchor establishes that preserving the pretrained language prior and grounding lexical choices can be done simultaneously in distribution space and representation space. TPA keeps the student model's next-token distribution close to the frozen LLM's distribution on the same prefix, scaled by confidence, so the decoder retains fluent linguistic form while still following visual evidence at uncertain positions. OTA solves an entropy-regularized partial transport problem between visual tokens and content-word tokens, learning soft many-to-many correspondences that survive granularity mismatch, with a dustbin absorbing unreliable mass. On the official test sets, DualAnchor reports the highes

What carries the argument

Token-level Prior Anchoring (TPA): a reverse-KL loss between the multimodal student's next-token distribution and a frozen LLM prior, both conditioned on the same gold prefix, weighted by a confidence gate ω = λ0·sigmoid(γ(H̄ − H)) built from the prior's entropy. Optimal Transport Alignment (OTA): an entropy-regularized partial optimal transport plan between ℓ2-normalized visual tokens and text content tokens, solved by masked Sinkhorn scaling, with dustbin slack absorbing unmatched mass. Together they carry the argument: TPA anchors linguistic form, OTA anchors lexical content.

Load-bearing premise

The load-bearing premise is that a low-entropy prediction from the frozen language model marks a position where visual evidence should not override it; if the language prior is confidently wrong about a word the video actually shows, the confidence gate applies the strongest pressure at exactly that position.

What would settle it

Run the same DualAnchor training with the confidence gate replaced by a constant weight or inverted; if BLEU-4 and IDF-weighted content-word recall on CSL-Daily do not drop, the gate is not doing the claimed work. More directly, collect gold examples where the frozen LLM's top token is wrong but the video-only student predicts correctly, and check whether TPA makes those positions worse.

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

If this is right

  • If TPA works, LLM-based sign translation can fine-tune a decoder for visual input without sacrificing the fluency the LLM already has; the prior acts as a regularizer rather than a teacher at inference.
  • If OTA works, sentence-level contrastive alignment is not enough for content words; soft token-level transport is a practical way to ground rare actions, objects, and time expressions.
  • Because both losses are training-only, any existing LLM-based sign translation pipeline can adopt them without slower decoding or extra teacher models at test time.
  • The cross-backbone results imply the two anchors transfer across different language models, so the recipe is not tied to one backbone.
  • The diagnostic link between lower student-prior KL and lower external PPL suggests prior preservation is a measurable mechanism, not just a side effect.

Where Pith is reading between the lines

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

  • The confidence gate is the least tested piece: if the frozen LLM is confidently wrong about a word that the video shows, TPA applies its strongest KL pressure at exactly that position. A direct ablation removing or inverting the gate would test whether the gate helps or occasionally hurts rare-word recall.
  • OTA's transport plan doubles as an interpretability tool: the paper's occlusion analysis shows transport mass marks video spans that, when masked, hurt generation most, so the same plan could localize which video segment supports each translated word.
  • The partial-transport-with-dustbin formulation is a generic solution to granularity mismatch between continuous visual tokens and discrete content words; it could transfer to dense video captioning or instruction-following as a lexical grounding objective.
  • A natural extension is to make the confidence gate itself adaptive to visual evidence, so the model decides whether to trust the prior per position rather than relying on language-model entropy alone.

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

5 major / 6 minor

Summary. The paper identifies two problems in gloss-free LLM-based sign language translation: language-prior degradation, where adapting a multimodal decoder to video shifts next-token distributions away from the frozen LM's prior, and a lexical fidelity gap, where sentence-level video-text alignment leaves content words ungrounded. The proposed DualAnchor framework adds two training-only objectives: Token-level Prior Anchoring (TPA), a confidence-weighted reverse-KL penalty between the student and frozen LM next-token distributions under a shared gold prefix, and Optimal Transport Alignment (OTA), an entropy-regularized partial optimal transport loss that softly aligns visual tokens with content tokens. On PHOENIX-2014T and CSL-Daily, the paper reports the best BLEU-4 among compared gloss-free methods (27.60 and 24.21) and the highest arithmetic mean over the ten reported metrics; ablations and diagnostics attribute the gains to TPA improving fluency and OTA improving lexical grounding.

Significance. If the claims hold, DualAnchor is a clean and useful contribution: both auxiliary losses operate only at training time, so there is no test-time cost, and the two objectives target complementary dimensions of translation quality. The paper has genuine strengths: the preliminary diagnostics (external-LM PPL, drift in student-to-prior KL, and lexical-fidelity analysis) are well motivated; the cross-backbone experiments on NLLB, mBART, and mT0 provide some controlled evidence; and the analysis suite (hard-negative margins, category-wise recall, error-type rates, occlusion, and design ablations) is unusually thorough. The central idea is defensible. However, the headline SOTA claim is currently under-supported by the comparison protocol, and a core component of TPA, the confidence gate, is never ablated.

major comments (5)
  1. [Implementation details / Table 1] The main runs never name the language backbone or visual encoder used for DualAnchor. The baselines in Table 1 are quoted from their original papers, and each uses a different backbone and training protocol. With BLEU-4 margins of only 1.60 and 0.96 over SCL-SLT, the reported gains cannot be attributed to TPA/OTA rather than to a stronger backbone or configuration difference. Please specify the backbone and visual encoder for the main runs, and either retrain SCL-SLT under the same protocol or provide same-backbone comparisons that include SCL-SLT. Table 2 is a step in this direction but is limited to CSL-Daily and omits SCL-SLT.
  2. [Tables 1 and 2 / Figure 8] No confidence intervals, multiple seeds, or significance tests are reported. With test sets of 642 and 1,176 sentences and early stopping on validation BLEU-4, the headline margins may fall within sampling noise. Please report at least three seeds with standard deviations/confidence intervals, or a bootstrap significance test for the pairwise comparisons against SCL-SLT and the closest baselines.
  3. [Eqs. (10)-(11), Table 4(a)] The confidence gate omega_{b,t} is a distinctive element of TPA, but it is never ablated. Table 4(a) varies only KL direction and context prefix; no run removes the gate or fixes omega=1. Since the paper claims the gate lets visual evidence influence uncertain positions, its contribution needs direct evidence. Add an ablation with a constant gate and, ideally, an inverted gate.
  4. [Preliminary Analysis / Figure 4 / RQ1] The 'external' LM used to measure fluency may be identical to the frozen prior used by TPA. The main backbone is not stated; if the decoder is Baichuan2-7B on CSL-Daily or Qwen2.5-7B on PHOENIX-2014T, then the reported PPL improvement is simply the quantity TPA optimizes, not independent evidence of fluency. Please identify the backbone and the evaluator; if they are the same model, use a genuinely held-out evaluator or explicitly qualify the claim.
  5. [Methodology / Equations (12)-(16); Implementation details] Several components needed for reproducibility are not specified: the content-selector rule for excluding 'weakly semantic function tokens', the values of lambda_0, gamma, rho, tau, alpha_TPA, alpha_OTA, the Sinkhorn iteration count, and the LoRA rank/alpha. These are not cosmetic; OTA's behavior depends directly on rho and tau, and TPA's gate depends on lambda_0 and gamma. Please report the values or a hyperparameter table.
minor comments (6)
  1. [Main Results] The 'highest arithmetic mean over the ten reported metrics' claim is not a standard comparison: BLEU-1..BLEU-4 are highly correlated and no variance is reported. Please report a pre-specified primary metric or a multi-metric significance test.
  2. [Figure 6] The caption says 'At the analyzed checkpoint' but does not state which epoch/checkpoint is used. Specify this for reproducibility.
  3. [Table 2] The sentence 'Rows share cached visual features, decoding settings, and the evaluation script' is unclear. Which visual encoder produced the cached features, and how do the per-backbone projections differ?
  4. [Section 3.3 OT Alignment] The content selector is only described verbally. Provide an algorithmic or thresholds-based specification, since it affects which tokens receive alignment mass.
  5. [Implementation details] 'Validation rounds' in the early-stopping description is ambiguous (epochs? evaluation steps?). Please clarify. Also, no code release is mentioned; consider adding an anonymous repository.
  6. [Conclusion / General] The paper has no limitations section. A brief discussion of when TPA's confidence gate or OTA's partial transport might fail would strengthen the presentation.

Circularity Check

0 steps flagged

No circularity found: the auxiliary losses are training objectives, and the headline results are measured with external metrics (BLEU/ROUGE and separate frozen LMs).

full rationale

The derivation chain is self-contained. TPA (Eq. 11) and OTA (Eq. 16) are training regularizers; neither is defined in terms of the evaluation metrics. The BLEU/ROUGE numbers in Table 1 are computed on official test sets against references, so the SOTA ranking is an empirical claim rather than a consequence of the objective definitions. The fluency attribution uses a frozen LM (Baichuan2/Qwen2.5) as an external PPL evaluator, and the lexical-error analyses use IDF-weighted content-word overlap, which is not the same quantity as the cosine-cost transport objective. The paper's own matched ablations (Table 4, Figures 4-8) support the attribution. The only self-citations (Chen et al. 2020a; Sun et al. 2020) are related-work references for IDF weighting and knowledge distillation; they are not load-bearing for the central result, and no uniqueness theorem from the authors is invoked. The confidence-gate under-ablation and lack of a named default backbone are correctness/completeness concerns, not evidence of circularity.

Axiom & Free-Parameter Ledger

7 free parameters · 5 axioms · 0 invented entities

DualAnchor introduces no new physical or conceptual entities; the 'dustbin' is an algorithmic device within partial OT, and 'language-prior degradation' / 'lexical fidelity gap' are names for observed phenomena, not invented entities. The central claim relies on at least six hand-specified hyperparameters (lambda_0, gamma, alpha_TPA, alpha_OTA, rho, tau) whose default values are not fully reported, plus several domain assumptions about the suitability of frozen-LM distributions, cosine cost, and external-LM evaluation.

free parameters (7)
  • lambda_0
    Maximum anchoring scale in TPA's confidence gate (Eq. 10-11); value not reported in text, presumably tuned on validation set.
  • gamma
    Gate sharpness in the sigmoid for TPA; value not reported, hand-chosen/tuned.
  • alpha_TPA = best around 0.3-0.4 (from Fig 8)
    TPA loss weight; Figure 8 shows a sweep from 0 to 0.5 but the default used in main experiments is not explicitly stated.
  • alpha_OTA = best around 0.1-0.2 (from Fig 8)
    OTA loss weight; Figure 8 shows a sweep but default is not explicitly stated.
  • rho
    Mass reserved for genuine matching in partial OT (Eq. 13); value not reported.
  • tau
    Entropy regularization temperature for Sinkhorn (Eq. 14); value not reported.
  • Content selector threshold/rule
    How content tokens are selected (excluding padding, punctuation, function tokens) is not specified beyond the description; the exact rule is an unstated free choice.
axioms (5)
  • domain assumption The frozen LLM's next-token distribution under the gold autoregressive prefix is a suitable linguistic prior target for sign-language translation.
    TPA (Eq. 11) assumes that matching the student to this frozen prior improves fluency without harming visual fidelity; the paper provides empirical correlation but no proof.
  • domain assumption Cosine similarity between projected visual tokens and textual content-token embeddings is a valid cross-modal match cost.
    OTA (Eq. 12) relies on this geometric comparability; there is no independent verification that the projected spaces are commensurate.
  • domain assumption External language models (Baichuan2-7B, Qwen2.5-7B) are reliable automatic fluency evaluators for Chinese and German.
    Diagnostics in Figure 2 and Figure 4 use these models' perplexity as a proxy for translation fluency.
  • domain assumption BLEU and IDF-weighted content-word F1/recall are adequate metrics for translation quality and lexical fidelity.
    The paper uses BLEU for the headline results and IDF-based metrics for diagnostics; it does not justify that these capture the relevant notion of fidelity.
  • domain assumption Uniform mass distributions over visual and textual tokens are an appropriate prior for partial OT matching.
    OTA sets a = 1/n and b = 1/m (Eq. 13), assuming each token carries equal mass; this is a modeling choice that is not empirically validated.

pith-pipeline@v1.3.0-daily-deepseek · 13971 in / 12688 out tokens · 131099 ms · 2026-08-01T04:35:08.310604+00:00 · methodology

0 comments
read the original abstract

Recent advances in large language models (LLMs) have led sign language translation (SLT), the task of converting sign-language videos into spoken-language text, to increasingly adopt LLMs as textual backbones. However, despite their strong language modeling capabilities, existing LLM-based SLT methods often undermine rather than exploit this language prior, producing disfluent translations, a failure we term language-prior degradation. Meanwhile, existing methods typically align videos and text at the sentence level, which does not ensure accurate lexical details and creates a lexical fidelity gap. To address both issues, we propose DualAnchor, a gloss-free LLM-based SLT training framework that couples two complementary anchors for linguistically fluent and visually faithful generation. Token-level Prior Anchoring (TPA) preserves the LLM's language prior by regularizing the multimodal decoder at each decoding step toward the next-token distribution of a frozen LLM conditioned on the same autoregressive prefix. Optimal Transport Alignment (OTA) improves lexical fidelity by formulating visual-textual matching as entropy-regularized partial optimal transport, with Sinkhorn optimization inducing a soft alignment between visual tokens and textual content tokens under a cosine cost. DualAnchor achieves strong overall performance on both PHOENIX-2014T and CSL-Daily. Targeted analyses attribute these gains to the complementary effects of the two anchors: TPA improves fluency, whereas OTA reduces fine-grained lexical errors.

Figures

Figures reproduced from arXiv: 2607.27614 by Hongbin Zhang, Junhao Liu, Kehai Chen, Xuefeng Bai, Yang Xiang, Youcheng Pan.

Figure 1
Figure 1. Figure 1: Motivating examples of the two issues in LLM [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Preliminary diagnostics of prior drift and lexical fidelity. (a) On CSL-Daily, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of DualAnchor. TPA anchors the student next-token distribution to a frozen language prior under the same autoregressive prefix, while OTA learns partial correspondences between visual tokens and textual content tokens. Both modules operate only during training. maximum anchoring scale, γ controls gate sharpness, and σ(·) is the sigmoid function. Low-entropy positions receive stronger prior guidanc… view at source ↗
Figure 4
Figure 4. Figure 4: CSL-Daily TPA diagnostics: (a) frozen-Baichuan2- [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: OTA grounding diagnostics on CSL-Daily: (a) lex [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: CSL-Daily OTA diagnostics: (a) category-wise [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Absolute CSL-Daily IDF-weighted recall (left) and [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: CSL-Daily best-checkpoint BLEU-4 versus (a) [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] 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

42 extracted references · 8 canonical work pages · 1 internal anchor

  1. [1]

    Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Camgoz, Necati Cihan and Hadfield, Simon and Koller, Oscar and Ney, Hermann and Bowden, Richard , title =. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2018 , pages =

    2018

  2. [2]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Zhou, Hao and Zhou, Wengang and Qi, Weizhen and Pu, Junfu and Li, Houqiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2021 , pages =

    2021

  3. [3]

    Gloss-Free Sign Language Translation: Improving from Visual-Language Pretraining , booktitle =

    Zhou, Benjia and Chen, Zhigang and Clap. Gloss-Free Sign Language Translation: Improving from Visual-Language Pretraining , booktitle =. 2023 , pages =

    2023

  4. [4]

    36th British Machine Vision Conference 2025,

    Asasi, Sobhan and Lakhal, Mohamed Ilyes and Sincan, Ozge Mercanoglu and Bowden, Richard , title =. 36th British Machine Vision Conference 2025,. 2025 , url =

    2025

  5. [5]

    The Twelfth International Conference on Learning Representations , year =

    Wong, Ryan and Camgoz, Necati Cihan and Bowden, Richard , title =. The Twelfth International Conference on Learning Representations , year =

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Gong, Jia and Foo, Lin Geng and He, Yixuan and Rahmani, Hossein and Liu, Jun , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  7. [7]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , month =

    Chen, Zhigang and Zhou, Benjia and Li, Jun and Wan, Jun and Lei, Zhen and Jiang, Ning and Lu, Quan and Zhao, Guoqing , title =. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024) , month =. 2024 , address =

    2024

  8. [8]

    , title =

    Hwang, Eui Jun and Cho, Sukmin and Lee, Junmyeong and Park, Jong C. , title =. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , month =. 2025 , address =. doi:10.18653/v1/2025.naacl-long.197 , url =

    doi:10.18653/v1/2025.naacl-long.197 2025

  9. [9]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Yin, Aoxiong and Zhong, Tianyun and Tang, Li and Jin, Weike and Jin, Tao and Zhao, Zhou , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    2023

  10. [10]

    Advances in Neural Information Processing Systems , volume =

    Ye, Jinhui and Wang, Xing and Jiao, Wenxiang and Liang, Junwei and Xiong, Hui , title =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  11. [11]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =

    Kim, Jungeun and Jeon, Hyeongwoo and Bae, Jongseong and Kim, Ha Young , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , month =. 2025 , pages =

    2025

  12. [12]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month =

    Think in Latent Thoughts: A New Paradigm for Gloss-Free Sign Language Translation , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month =. 2026 , address =. doi:10.18653/v1/2026.acl-long.454 , url =

    work page doi:10.18653/v1/2026.acl-long.454 2026

  13. [13]

    and Leike, Jan and Lowe, Ryan , title =

    Ouyang, Long and Wu, Jeffrey and Jiang, Xu and Almeida, Diogo and Wainwright, Carroll and Mishkin, Pamela and Zhang, Chong and Agarwal, Sandhini and Slama, Katarina and Ray, Alex and Schulman, John and Hilton, Jacob and Kelton, Fraser and Miller, Luke and Simens, Maddie and Askell, Amanda and Welinder, Peter and Christiano, Paul F. and Leike, Jan and Lowe...

    2022

  14. [14]

    Advances in Neural Information Processing Systems , volume =

    Wang, Weihan and Lv, Qingsong and Yu, Wenmeng and Hong, Wenyi and Qi, Ji and Wang, Yan and Ji, Junhui and Yang, Zhuoyi and Zhao, Lei and Song, Xixuan and Xu, Jiazheng and Chen, Keqin and Xu, Bin and Li, Juanzi and Dong, Yuxiao and Ding, Ming and Tang, Jie , title =. Advances in Neural Information Processing Systems , volume =. 2024 , doi =

    2024

  15. [15]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Lin, Ji and Yin, Hongxu and Ping, Wei and Molchanov, Pavlo and Shoeybi, Mohammad and Han, Song , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  16. [16]

    LinguDistill: Recovering Linguistic Ability in Vision-Language Models via Selective Cross-Modal Distillation

    Irawan, Patrick Amadeus and Fuadi, Erland Hilman and Kumar, Shanu and Aji, Alham Fikri and Kementchedjhieva, Yova , year =. doi:10.48550/ARXIV.2604.00829 , url =. 2604.00829 , archivePrefix =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2604.00829

  17. [17]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Cheng, Yiting and Wei, Fangyun and Bao, Jianmin and Chen, Dong and Zhang, Wenqiang , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    2023

  18. [18]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =

    Low, Jian He and Sincan, Ozge Mercanoglu and Bowden, Richard , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops , month =. 2025 , pages =

    2025

  19. [19]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month =

    Selective Contrastive Learning For Gloss Free Sign Language Translation , author =. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , month =. 2026 , address =. doi:10.18653/v1/2026.acl-long.2116 , url =

    work page doi:10.18653/v1/2026.acl-long.2116 2026

  20. [20]

    doi:10.48550/ARXIV.2309.10305 , url =

    Yang, Aiyuan and Xiao, Bin and Wang, Bingning and Zhang, Borong and Bian, Ce and Yin, Chao and Lv, Chenxu and Pan, Da and Wang, Dian and Yan, Dong and Yang, Fan and Deng, Fei and Wang, Feng and Liu, Feng and Ai, Guangwei and Dong, Guosheng and Zhao, Haizhou and Xu, Hang and Sun, Haoze and Zhang, Hongda and Liu, Hui and Ji, Jiaming and Xie, Jian and Dai, J...

    Pith/arXiv arXiv doi:10.48550/arxiv.2309.10305

  21. [21]

    doi:10.48550/ARXIV.2412.15115 , url =

    Yang, An and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Li, Chengyuan and Liu, Dayiheng and Huang, Fei and Wei, Haoran and Lin, Huan and Yang, Jian and Tu, Jianhong and Zhang, Jianwei and Yang, Jianxin and Yang, Jiaxi and Zhou, Jingren and Lin, Junyang and Dang, Kai and Lu, Keming and Bao, Keqin and Yang, Kexin and Y...

    Pith/arXiv arXiv doi:10.48550/arxiv.2412.15115

  22. [22]

    2002 , address =

    Papineni, Kishore and Roukos, Salim and Ward, Todd and Zhu, Wei-Jing , booktitle =. 2002 , address =. doi:10.3115/1073083.1073135 , pages =

    arXiv 2002

  23. [23]

    2004 , address =

    Lin, Chin-Yew , booktitle =. 2004 , address =

    2004

  24. [24]

    Unsupervised Sign Language Translation and Generation

    Guo, Zhengsheng and He, Zhiwei and Jiao, Wenxiang and Wang, Xing and Wang, Rui and Chen, Kehai and Tu, Zhaopeng and Xu, Yong and Zhang, Min. Unsupervised Sign Language Translation and Generation. Findings of the Association for Computational Linguistics: ACL 2024. 2024. doi:10.18653/v1/2024.findings-acl.835

    work page doi:10.18653/v1/2024.findings-acl.835 2024

  25. [25]

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , url =

    Cuturi, Marco , booktitle =. Sinkhorn Distances: Lightspeed Computation of Optimal Transport , url =

  26. [26]

    Sign Language Transformers: Joint End-to-End Sign Language Recognition and Translation , booktitle =

    Camg. Sign Language Transformers: Joint End-to-End Sign Language Recognition and Translation , booktitle =. 2020 , pages =. doi:10.1109/CVPR42600.2020.01004 , url =

    arXiv 2020

  27. [27]

    Proceedings of the 28th International Conference on Computational Linguistics , year =

    Yin, Kayo and Read, Jesse , title =. Proceedings of the 28th International Conference on Computational Linguistics , year =. doi:10.18653/v1/2020.coling-main.525 , url =

    doi:10.18653/v1/2020.coling-main.525 2020

  28. [28]

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

    Lin, Kezhou and Wang, Xiaohan and Zhu, Linchao and Sun, Ke and Zhang, Bang and Yang, Yi , title =. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. doi:10.18653/v1/2023.acl-long.722 , url =

    doi:10.18653/v1/2023.acl-long.722 2023

  29. [29]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , year =

    Jiang, Zifan and Sant, Gerard and Moryossef, Amit and M. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , year =. doi:10.18653/v1/2024.emnlp-main.518 , url =

    work page doi:10.18653/v1/2024.emnlp-main.518 2024

  30. [30]

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

    Chen, Kehai and Wang, Rui and Utiyama, Masao and Sumita, Eiichiro , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =. doi:10.18653/v1/2020.acl-main.34 , url =

    work page doi:10.18653/v1/2020.acl-main.34 2020

  31. [31]

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

    Sun, Haipeng and Wang, Rui and Chen, Kehai and Utiyama, Masao and Sumita, Eiichiro and Zhao, Tiejun , title =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics , year =. doi:10.18653/v1/2020.acl-main.324 , url =

    work page doi:10.18653/v1/2020.acl-main.324 2020

  32. [32]

    , title =

    Kim, Yoon and Rush, Alexander M. , title =. Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , year =. doi:10.18653/v1/D16-1139 , url =

    doi:10.18653/v1/d16-1139 2016

  33. [33]

    and Gasso, Gilles , title =

    Chapel, Laetitia and Alaya, Mokhtar Z. and Gasso, Gilles , title =. Advances in Neural Information Processing Systems , volume =. 2020 , url =

    2020

  34. [34]

    Proceedings of the 37th International Conference on Machine Learning , series =

    Chen, Liqun and Gan, Zhe and Cheng, Yu and Li, Linjie and Carin, Lawrence and Liu, Jingjing , title =. Proceedings of the 37th International Conference on Machine Learning , series =. 2020 , pages =

    2020

  35. [35]

    and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =

    Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , title =. International Conference on Learning Representations , year =

  36. [36]

    International Conference on Learning Representations , year =

    Loshchilov, Ilya and Hutter, Frank , title =. International Conference on Learning Representations , year =

  37. [37]

    2024 , pages =

    Scaling Neural Machine Translation to 200 Languages , journal =. 2024 , pages =. doi:10.1038/s41586-024-07335-x , url =

    doi:10.1038/s41586-024-07335-x 2024

  38. [38]

    Transactions of the Association for Computational Linguistics , volume =

    Liu, Yinhan and Gu, Jiatao and Goyal, Naman and Li, Xian and Edunov, Sergey and Ghazvininejad, Marjan and Lewis, Mike and Zettlemoyer, Luke , title =. Transactions of the Association for Computational Linguistics , volume =. 2020 , pages =. doi:10.1162/tacl_a_00343 , url =

    doi:10.1162/tacl_a_00343 2020

  39. [39]

    Muennighoff, Niklas and Wang, Thomas and Sutawika, Lintang and Roberts, Adam and Biderman, Stella and Le Scao, Teven and Bari, M. Saiful and Shen, Sheng and Yong, Zheng Xin and Schoelkopf, Hailey and Tang, Xiangru and Radev, Dragomir and Aji, Alham Fikri and Almubarak, Khalid and Albanie, Samuel and Alyafeai, Zaid and Webson, Albert and Raff, Edward and R...

    2023

  40. [40]

    International Conference on Learning Representations , year =

    Zhang, Biao and M. International Conference on Learning Representations , year =

  41. [41]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

    Yao, Huijie and Zhou, Wengang and Feng, Hao and Hu, Hezhen and Zhou, Hao and Li, Houqiang , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision , year =

  42. [42]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , year =

    Tan, Sihan and Miyazaki, Taro and Nakadai, Kazuhiro , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , year =. doi:10.18653/v1/2025.acl-short.43 , url =

    work page doi:10.18653/v1/2025.acl-short.43 2025