Pith. sign in

REVIEW 2 major objections 4 minor 61 references

Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval

T0 review · 2 major / 4 minor · reviewed 2026-07-13 · grok-4.5

Pith's one-line read Fine-grained sign retrieval fails because training never sees the true hard negatives: visually confusable signs.

desk verdict Solid, usable fix for a real mismatch in sign-language retrieval; the circularity in the fine-grained test is real but does not erase the comparative result against text-based mining. read the letter →

arxiv 2607.09263 v1 pith:NON4HLVZ submitted 2026-07-10 cs.CV

classification cs.CV
keywords signlanguageretrievalhardnegativeminingvisualconfusabilitycontrastivelearningfine-graineddiscriminationcross-modalPHOENIX-2014T
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

Sign language retrieval works reasonably for coarse matching but collapses when the model must tell apart signs that look almost the same yet mean different things. The authors argue the problem is not model size; it is a mismatch in the negatives used for training. Text-based hard-negative mining (word swaps, language models) produces negatives that are linguistically related but visually easy to separate, so the model never learns the subtle distinctions that actually matter. Their fix, Sign-Aware Hard Negative Mining (SAN), first finds reliable sign–word pairs, then mines other signs that sit close together in the visual embedding space, and finally substitutes those words into captions to create genuine hard negatives. On the standard weather-forecast sign dataset this raises fine-grained recall@1 by more than twenty points while leaving ordinary retrieval accuracy essentially intact. The result shows that supervision must be aligned with visual confusability rather than linguistic similarity if fine-grained sign retrieval is to succeed.

What carries the argument

Sign-Aware Hard Negative Mining (SAN): a three-step procedure that (1) extracts high-confidence sign–word alignments, (2) finds other signs whose embeddings exceed a cosine-similarity threshold yet carry different words, and (3) substitutes those words into captions to form hard negatives used in an auxiliary contrastive loss.

What would settle it

If a human study or phonological-feature analysis showed that the pairs SAN ranks as most similar are not the pairs signers actually confuse, while text-based pairs are more confusable, the claimed advantage of visual mining would disappear.

Watch

Extended reading notes

Core claim

Fine-grained retrieval failures arise from a negative distribution mismatch: semantically distinct yet visually confusable signs are almost never presented as hard negatives during contrastive training, and text-based mining strategies cannot reach those signs. Aligning hard-negative construction with visual proximity in the sign embedding space corrects the mismatch and unlocks the discrimination that existing models already possess.

Load-bearing premise

The paper treats cosine similarity of features from a pretrained sign encoder as a faithful stand-in for true visual confusability, both when mining training negatives and when building the fine-grained test set itself.

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

2 major / 4 minor

Summary. The paper argues that fine-grained failures in sign language retrieval (SLRet) arise from a negative-distribution mismatch in contrastive learning: semantically distinct but visually confusable signs are under-sampled as hard negatives, while text-based mining produces only linguistically hard (visually easy) negatives. It proposes Sign-Aware Hard Negative Mining (SAN), which (1) extracts high-confidence sign–word pairs from a pretrained retrieval model (threshold α), (2) mines visually similar but semantically distinct words via cosine similarity of sign embeddings (threshold β), and (3) generates hard-negative captions by keyword substitution. These are added to the CLCL objective via a weighted fine-grained InfoNCE term (Eqs. 10–11). On PHOENIX-2014T, SAN yields large gains on a custom fine-grained stress test (V2T R@1 +21.5 on CiCo, +32.3 on GFSLT-VLP) while largely preserving coarse-grained accuracy, outperforming FastText/RoBERTa/GPT-4o-mini mining.

Significance. If the central claim holds, the work supplies a concrete, easily implemented recipe for aligning hard-negative supervision with the articulatory constraints of sign language, a domain where visual confusability is known to be frequent. The consistent gains across two backbones, the clean ablations on Nswap/Nhard/λ, and the qualitative demonstration that SAN recovers phonologically plausible near-minimal pairs are useful contributions. The paper also makes the useful conceptual point that linguistic hardness ≠ visual hardness for SLRet. The main limitation on significance is that all evidence is confined to a single weather-forecast corpus and that the fine-grained evaluation itself is constructed with the same embedding-similarity proxy that SAN optimizes, so the generality of the claimed resolution of “true” visual ambiguity remains to be established.

major comments (2)
  1. Section 4.1 (“Fine-grained Stress Test Construction”) and Eqs. 8–9: both the training negatives and the fine-grained test set are defined by the identical criterion—cosine similarity of sign features from a pretrained retrieval model exceeding β = 0.7. Consequently the large R@1 gains in Table 1 partly measure recovery of the same proxy that SAN was trained to cover, rather than an independent measure of articulatory hardness. The I3D scores in Figure 3 and Appendix A.3 are post-hoc illustrations, not an independent construction of the test set. An external probe (phonological minimal pairs from a sign dictionary, human confusability ratings, or a held-out visual encoder) is needed to substantiate the claim that SAN resolves genuine visual ambiguity.
  2. All quantitative results (Table 1, ablations, Figures 4–6) come from a single dataset (PHOENIX-2014T) with no error bars, multiple random seeds, or cross-dataset validation. Given the free parameters α, β, λ, Nswap and Nhard, and the acknowledged domain specificity of weather-forecast signs (Limitations), the magnitude and stability of the reported gains cannot yet be assessed. At minimum, seed-averaged results and an experiment on a second corpus (or a non-weather subset) are required before the central claim can be considered robust.
minor comments (4)
  1. The reliability of the frozen sign–word alignments (Eqs. 4–7, α = 0.7) is asserted but never quantified (e.g., precision against gloss annotations). A short diagnostic would strengthen the pipeline description.
  2. Figure 4 shows a clear shift of SAN negatives toward higher cosine similarity, yet no statistical test or effect-size number is reported; a simple KS statistic or mean/median would make the distributional claim more precise.
  3. The final objective (Eq. 11) treats Lfine as an additive term with a single scalar λ; it is unclear whether the hard negatives are also used in the T2V direction or only V2T (Eq. 10). Clarifying the bidirectional formulation would improve reproducibility.
  4. Appendix A.2.3 notes that POS filtering is applied only at evaluation time for SAN/FastText; a brief statement of residual grammatical noise remaining after filtering would help readers interpret the fine-grained numbers.

Circularity Check

1 steps flagged · score 4.0 of 10

Fine-grained stress-test set is constructed with the identical cosine-similarity proxy (β) that SAN uses to mine training negatives, giving the method a mild circular advantage on its central claim.

  1. other [Section 4.1 (Fine-grained Stress Test Construction) + Implementation Details + Eqs. 8–9]
    "Target words are restricted to those exhibiting visual confusability—specifically, words whose corresponding signs have at least one visually similar counterpart (cosine similarity > β in the sign embedding space) among semantically distinct signs. ... we use sign embedding similarity as a proxy. ... we set α=0.7 and β=0.7 for mining negatives in both training and evaluation."

    The fine-grained evaluation instances are defined by the identical visual-similarity threshold and embedding space that SAN uses to construct its training hard negatives. Consequently the large R@1 lifts (Table 1) partly measure recovery of the proxy distribution that the method was designed to cover, rather than an independently annotated measure of articulatory confusability. The circularity is confined to evaluation construction; the training loss itself is not tautological.

full rationale

The paper's core claim—that SAN resolves fine-grained failures by aligning negatives with visual confusability—is evaluated on a diagnostic set whose target words are themselves selected by the same criterion (sign-embedding cosine > β) that defines SAN (Eqs. 8–9). Target words are held fixed across mining strategies and a merged pool of 40 negatives is used, so the comparison is controlled; coarse-grained results remain independent and SAN still shows a favorable trade-off. Nevertheless the hardness distribution of the fine-grained probe is not an external articulatory ground truth but a sample from the geometry SAN is optimized to cover. This is a genuine but limited circularity of evaluation design rather than a definitional collapse of the training objective or a self-citation chain. No other load-bearing steps reduce by construction; the method itself is a straightforward redefinition of hard-negative selection and the reported gains are real relative to text-based baselines on the chosen proxy.

Assumptions & free parameters 5 free parameters · 3 assumptions · 1 invented entities

The central claim rests on a small set of free thresholds chosen by ablation, the domain assumption that embedding cosine is a valid confusability proxy, and the standard InfoNCE machinery. No new physical or mathematical entities are postulated.

free parameters (5)
  • α (sign–word reliability threshold) = 0.7
    Hand-chosen at 0.7 after ablation; controls which alignments enter the negative pool.
  • β (visual similarity threshold) = 0.7
    Hand-chosen at 0.7; defines both training negatives and the fine-grained test set.
  • λ (fine-grained loss weight) = 0.4
    Hand-chosen at 0.4 to balance coarse and fine objectives.
  • Nswap (words replaced per caption) = 2
    Selected as 2 after trade-off study.
  • Nhard (hard negatives per instance) = 5
    Selected as 5 after trade-off study.
assumptions (3)
  • domain assumption Cosine similarity of sign features from a pretrained retrieval model is a faithful proxy for human visual confusability of signs.
    Invoked for both mining (Eq. 8–9) and fine-grained test construction (Section 4.1).
  • standard math InfoNCE with temperature τ is an appropriate contrastive objective for video–text alignment.
    Standard CLCL formulation (Eqs. 1–3) taken from prior SLRet work.
  • ad hoc to paper High-confidence sign–word alignments (p > α) extracted from a frozen model are sufficiently clean to seed negative mining.
    Section 3.2, Reliable Sign–Word Pair Mining; no external validation of alignment quality is provided.
invented entities (1)
  • Sign-Aware Hard Negative Mining (SAN) pipeline
    purpose: Operational procedure that turns visual proximity into hard-negative captions.
    The three-stage algorithm is the paper’s main technical contribution; it has no independent existence outside this work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval." pith.science (2026). https://pith.science/paper/NON4HLVZ

@misc{pith2026260709263,
  author       = {Pith},
  title        = {Pith review of: Semantic Hardness Is Not Visual Hardness: Sign-Aware Hard Negative Mining for Sign Language Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NON4HLVZ}},
  note         = {Machine review of arXiv:2607.09263}
}
read the original abstract

Sign Language Retrieval (SLRet) enables efficient access to sign language content but remains fragile in fine-grained scenarios where visually similar signs must be distinguished. We show that this limitation does not stem from model capacity, but from ineffective hard negative supervision. Specifically, we formulate fine-grained retrieval failures as a negative distribution mismatch: semantically distinct yet visually confusable signs are rarely treated as hard negatives, while existing text-based mining strategies fail to capture such visual ambiguity. To address this issue, we propose Sign-Aware Hard Negative Mining (SAN), which constructs hard negatives based on visual confusability in the sign embedding space rather than linguistic similarity. Experiments on PHOENIX-2014T demonstrate that SAN substantially improves fine-grained retrieval performance while preserving coarse-grained accuracy, highlighting the importance of aligning negative supervision with visual ambiguity in sign language retrieval.

Figures

Figures reproduced from arXiv: 2607.09263 by the authors.

Figure 1
Figure 1. Illustration of fine-grained ambiguity in sign [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the Sign-Aware Hard Negative Mining (SAN) framework. SAN consists of three steps: [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. Distribution of cosine similarities between [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: Qualitative comparison of mined hard nega [PITH_FULL_IMAGE:figures/full_fig_p007_3.png]
Figure 6
Figure 6. Figure 6: Effect of λ on fine-grained and coarse-grained retrieval performance. drop does not indicate model failure, but rather con￾firms the high quality of our hard negatives. Unlike text-based baselines that produce easily distinguish￾able negatives, SAN successfully mines v…
Figure 7
Figure 7. Figure 7: Ablation study on hyperparameters α (top row) and β (bottom row). The left column shows fine￾grained retrieval results (V2T), while the right column shows coarse-grained results (V2T and T2V). A Appendix A.1 Impact of Threshold α and β in SAN framework We analyze the e…
Figure 8
Figure 8. Figure 8: An example of fine-grained retrieval in PHOENIX-2014T where only SAN correctly ranks the target at R@1, while language-model-based methods fail to correctly distinguish visually confusable nega￾tives. Target caption: “In the north, stormy gusts are also possible.” Hard…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 9 canonical work pages

  1. [1]

    Papadopoulos and Vassia Zacharopoulou and George J

    Nikolas Adaloglou and Theocharis Chatzis and Ilias Papastratis and Andreas Stergioulas and Georgios Th. Papadopoulos and Vassia Zacharopoulou and George J. Xydopoulos and Klimnis Atzakas and Dimitris Papazachariou and Petros Daras , title =. 2022 , url =. doi:10.1109/TMM.2021.3070438 , timestamp =

  2. [2]

    Neurocomputing , volume =

    Kun Hu and Fengxiang He and Adam Schembri and Zhiyong Wang , title =. Neurocomputing , volume =. 2025 , url =. doi:10.1016/J.NEUCOM.2025.130077 , timestamp =

  3. [3]

    CoRR , volume =

    Han Fang and Pengfei Xiong and Luhui Xu and Yu Chen , title =. CoRR , volume =. 2021 , url =. 2106.11097 , timestamp =

  4. [4]

    Learning Transferable Visual Models From Natural Language Supervision , booktitle =

    Alec Radford and Jong Wook Kim and Chris Hallacy and Aditya Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever , editor =. Learning Transferable Visual Models From Natural Language Supervision , booktitle =. 2021 , url =

  5. [5]

    Neurocomputing , volume =

    Huaishao Luo and Lei Ji and Ming Zhong and Yang Chen and Wen Lei and Nan Duan and Tianrui Li , title =. Neurocomputing , volume =. 2022 , url =. doi:10.1016/J.NEUCOM.2022.07.028 , timestamp =

  6. [6]

    VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding , booktitle =

    Hu Xu and Gargi Ghosh and Po. VideoCLIP: Contrastive Pre-training for Zero-shot Video-Text Understanding , booktitle =. 2021 , url =. doi:10.18653/V1/2021.EMNLP-MAIN.544 , timestamp =

  7. [7]

    The Thirteenth International Conference on Learning Representations,

    Garrett Tanzer and Biao Zhang , title =. The Thirteenth International Conference on Learning Representations,. 2025 , url =

  8. [8]

    YouTube-ASL:

    David Uthus and Garrett Tanzer and Manfred Georg , editor =. YouTube-ASL:. Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023 , year =

Show all 61 references
  1. [9]

    BBC-Oxford British Sign Language Dataset , journal =

    Samuel Albanie and G. BBC-Oxford British Sign Language Dataset , journal =. 2021 , url =. 2111.03635 , timestamp =

  2. [10]

    Open-Domain Sign Language Translation Learned from Online Video , booktitle =

    Bowen Shi and Diane Brentari and Gregory Shakhnarovich and Karen Livescu , editor =. Open-Domain Sign Language Translation Learned from Online Video , booktitle =. 2022 , url =. doi:10.18653/V1/2022.EMNLP-MAIN.427 , timestamp =

  3. [11]

    Neural Sign Language Translation , booktitle =

    Necati Cihan Camg. Neural Sign Language Translation , booktitle =. 2018 , url =. doi:10.1109/CVPR.2018.00812 , timestamp =

  4. [12]

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

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

  5. [13]

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

    Benjia Zhou and Zhigang Chen and Albert Clap. Gloss-free Sign Language Translation: Improving from Visual-Language Pretraining , booktitle =. 2023 , url =. doi:10.1109/ICCV51070.2023.01908 , timestamp =

  6. [14]

    Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation , booktitle =

    Zhigang Chen and Benjia Zhou and Jun Li and Jun Wan and Zhen Lei and Ning Jiang and Quan Lu and Guoqing Zhao , editor =. Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation , booktitle =. 2024 , url =

  7. [15]

    Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation , journal =

    Ryan Wong and Necati Cihan Camg. Sign2GPT: Leveraging Large Language Models for Gloss-Free Sign Language Translation , journal =. 2024 , url =. doi:10.48550/ARXIV.2405.04164 , eprinttype =. 2405.04164 , timestamp =

  8. [16]

    Park , editor =

    Eui Jun Hwang and Sukmin Cho and Junmyeong Lee and Jong C. Park , editor =. An Efficient Gloss-Free Sign Language Translation Using Spatial Configurations and Motion Dynamics with LLMs , booktitle =. 2025 , url =. doi:10.18653/V1/2025.NAACL-LONG.197 , timestamp =

  9. [17]

    2024 , url =

    Jia Gong and Lin Geng Foo and Yixuan He and Hossein Rahmani and Jun Liu , title =. 2024 , url =. doi:10.1109/CVPR52733.2024.01738 , timestamp =

  10. [18]

    2023 , url =

    Aoxiong Yin and Tianyun Zhong and Li Tang and Weike Jin and Tao Jin and Zhou Zhao , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.00251 , timestamp =

  11. [19]

    Visual Alignment Pre-training for Sign Language Translation , booktitle =

    Peiqi Jiao and Yuecong Min and Xilin Chen , editor =. Visual Alignment Pre-training for Sign Language Translation , booktitle =. 2024 , url =. doi:10.1007/978-3-031-72946-1\_20 , timestamp =

  12. [20]

    Hezhen Hu and Weichao Zhao and Wengang Zhou and Yuechen Wang and Houqiang Li , title =. 2021. 2021 , url =. doi:10.1109/ICCV48922.2021.01090 , timestamp =

  13. [21]

    2023 , url =

    Ronglai Zuo and Fangyun Wei and Brian Mak , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.01430 , timestamp =

  14. [22]

    2021 , url =

    Songyao Jiang and Bin Sun and Lichen Wang and Yue Bai and Kunpeng Li and Yun Fu , title =. 2021 , url =. doi:10.1109/CVPRW53098.2021.00380 , timestamp =

  15. [23]

    Thirty-Seventh

    Weichao Zhao and Hezhen Hu and Wengang Zhou and Jiaxin Shi and Houqiang Li , editor =. Thirty-Seventh. 2023 , url =. doi:10.1609/AAAI.V37I3.25470 , timestamp =

  16. [24]

    Computer Vision -

    Samuel Albanie and G. Computer Vision -. 2020 , url =. doi:10.1007/978-3-030-58621-8\_3 , timestamp =

  17. [25]

    Learn to Understand Negation in Video Retrieval , booktitle =

    Ziyue Wang and Aozhu Chen and Fan Hu and Xirong Li , editor =. Learn to Understand Negation in Video Retrieval , booktitle =. 2022 , url =. doi:10.1145/3503161.3547968 , timestamp =

  18. [26]

    Piyush Bagad and Makarand Tapaswi and Cees G. M. Snoek , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.00247 , timestamp =

  19. [27]

    Hazel Doughty and Cees G. M. Snoek , title =. 2022 , url =. doi:10.1109/CVPR52688.2022.01346 , timestamp =

  20. [28]

    2023 , url =

    Liliane Momeni and Mathilde Caron and Arsha Nagrani and Andrew Zisserman and Cordelia Schmid , title =. 2023 , url =. doi:10.1109/ICCV51070.2023.01428 , timestamp =

  21. [29]

    Aozhu Chen and Hazel Doughty and Xirong Li and Cees G. M. Snoek , editor =. Beyond Coarse-Grained Matching in Video-Text Retrieval , booktitle =. 2024 , url =. doi:10.1007/978-981-96-0908-6\_2 , timestamp =

  22. [30]

    Proceedings of the 32nd

    Longtao Jiang and Min Wang and Zecheng Li and Yao Fang and Wengang Zhou and Houqiang Li , editor =. Proceedings of the 32nd. 2024 , url =. doi:10.1145/3664647.3681237 , timestamp =

  23. [31]

    2023 , url =

    Yiting Cheng and Fangyun Wei and Jianmin Bao and Dong Chen and Wenqiang Zhang , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.01823 , timestamp =

  24. [32]

    Sign Language Video Retrieval with Free-Form Textual Queries , booktitle =

    Amanda Cardoso Duarte and Samuel Albanie and Xavier Gir. Sign Language Video Retrieval with Free-Form Textual Queries , booktitle =. 2022 , url =. doi:10.1109/CVPR52688.2022.01370 , timestamp =

  25. [33]

    Uncertainty-Aware Sign Language Video Retrieval with Probability Distribution Modeling , booktitle =

    Xuan Wu and Hongxiang Li and Yuanjiang Luo and Xuxin Cheng and Xianwei Zhuang and Meng Cao and Keren Fu , editor =. Uncertainty-Aware Sign Language Video Retrieval with Probability Distribution Modeling , booktitle =. 2024 , url =. doi:10.1007/978-3-031-72784-9\_22 , timestamp =

  26. [34]

    Contrastive Learning with Hard Negative Samples , booktitle =

    Joshua David Robinson and Ching. Contrastive Learning with Hard Negative Samples , booktitle =. 2021 , url =

  27. [35]

    Teaching Structured Vision

    Sivan Doveh and Assaf Arbelle and Sivan Harary and Eli Schwartz and Roei Herzig and Raja Giryes and Rog. Teaching Structured Vision. 2023 , url =. doi:10.1109/CVPR52729.2023.00261 , timestamp =

  28. [36]

    When and Why Vision-Language Models Behave like Bags-Of- rs, and What to Do About It? , booktitle =

    Mert Y. When and Why Vision-Language Models Behave like Bags-Of- rs, and What to Do About It? , booktitle =. 2023 , url =

  29. [37]

    2023 , url =

    Roni Paiss and Ariel Ephrat and Omer Tov and Shiran Zada and Inbar Mosseri and Michal Irani and Tali Dekel , title =. 2023 , url =. doi:10.1109/ICCV51070.2023.00294 , timestamp =

  30. [38]

    2024 , url =

    Le Zhang and Rabiul Awal and Aishwarya Agrawal , title =. 2024 , url =. doi:10.1109/CVPR52733.2024.01307 , timestamp =

  31. [39]

    Hard Negative Mixing for Contrastive Learning , booktitle =

    Yannis Kalantidis and Mert B. Hard Negative Mixing for Contrastive Learning , booktitle =. 2020 , url =

  32. [40]

    Improving Gloss-free Sign Language Translation by Reducing Representation Density , booktitle =

    Jinhui Ye and Xing Wang and Wenxiang Jiao and Junwei Liang and Hui Xiong , editor =. Improving Gloss-free Sign Language Translation by Reducing Representation Density , booktitle =. 2024 , url =

  33. [41]

    Towards Privacy-Aware Sign Language Translation at Scale , booktitle =

    Phillip Rust and Bowen Shi and Skyler Wang and Necati Cihan Camg. Towards Privacy-Aware Sign Language Translation at Scale , booktitle =. 2024 , url =. doi:10.18653/V1/2024.ACL-LONG.467 , timestamp =

  34. [42]

    Scaling Sign Language Translation , booktitle =

    Biao Zhang and Garrett Tanzer and Orhan Firat , editor =. Scaling Sign Language Translation , booktitle =. 2024 , url =

  35. [43]

    30th British Machine Vision Conference 2019,

    Hamid Reza Vaezi Joze and Oscar Koller , title =. 30th British Machine Vision Conference 2019,. 2019 , url =

  36. [44]

    2021 , url =

    Razieh Rastgoo and Kourosh Kiani and Sergio Escalera and Mohammad Sabokrou , title =. 2021 , url =. doi:10.1109/CVPRW53098.2021.00384 , timestamp =

  37. [45]

    CoRR , volume =

    Yinhan Liu and Myle Ott and Naman Goyal and Jingfei Du and Mandar Joshi and Danqi Chen and Omer Levy and Mike Lewis and Luke Zettlemoyer and Veselin Stoyanov , title =. CoRR , volume =. 2019 , url =. 1907.11692 , timestamp =

  38. [46]

    2024 , url =

    Junseok Ahn and Youngjoon Jang and Joon Son Chung , title =. 2024 , url =. doi:10.1109/ICASSP48485.2024.10445841 , timestamp =

  39. [47]

    2023 , url =

    Lianyu Hu and Liqing Gao and Zekang Liu and Wei Feng , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.00249 , timestamp =

  40. [48]

    2022 , url =

    Yutong Chen and Fangyun Wei and Xiao Sun and Zhirong Wu and Stephen Lin , title =. 2022 , url =. doi:10.1109/CVPR52688.2022.00506 , timestamp =

  41. [49]

    Aiming Hao and Yuecong Min and Xilin Chen , title =. 2021. 2021 , url =. doi:10.1109/ICCV48922.2021.01111 , timestamp =

  42. [50]

    Pattern Recognit

    Mo Guan and Yan Wang and Guangkun Ma and Jiarui Liu and Mingzu Sun , title =. Pattern Recognit. , volume =. 2025 , url =. doi:10.1016/J.PATCOG.2025.111602 , timestamp =

  43. [51]

    The Eleventh International Conference on Learning Representations,

    Biao Zhang and Mathias M. The Eleventh International Conference on Learning Representations,. 2023 , url =

  44. [52]

    Two-Stream Network for Sign Language Recognition and Translation , booktitle =

    Yutong Chen and Ronglai Zuo and Fangyun Wei and Yu Wu and Shujie Liu and Brian Mak , editor =. Two-Stream Network for Sign Language Recognition and Translation , booktitle =. 2022 , url =

  45. [53]

    Noise-contrastive estimation:

    Michael Gutmann and Aapo Hyv. Noise-contrastive estimation:. Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics,. 2010 , url =

  46. [54]

    CoRR , volume =

    OpenAI , title =. CoRR , volume =. 2023 , url =. doi:10.48550/ARXIV.2303.08774 , eprinttype =. 2303.08774 , timestamp =

  47. [55]

    Advances in Pre-Training Distributed Word Representations , booktitle =

    Tom. Advances in Pre-Training Distributed Word Representations , booktitle =. 2018 , url =

  48. [56]

    2023 , url =

    Yuxin Chen and Zongyang Ma and Ziqi Zhang and Zhongang Qi and Chunfeng Yuan and Ying Shan and Bing Li and Weiming Hu and Xiaohu Qie and Jianping Wu , title =. 2023 , url =. doi:10.1109/CVPR52729.2023.01060 , timestamp =

  49. [57]

    VideoCon: Robust Video-Language Alignment via Contrast Captions , booktitle =

    Hritik Bansal and Yonatan Bitton and Idan Szpektor and Kai. VideoCon: Robust Video-Language Alignment via Contrast Captions , booktitle =. 2024 , url =. doi:10.1109/CVPR52733.2024.01321 , timestamp =

  50. [58]

    TripletCLIP: Improving Compositional Reasoning of

    Maitreya Patel and Abhiram Kusumba and Sheng Cheng and Changhoon Kim and Tejas Gokhale and Chitta Baral and Yezhou Yang , editor =. TripletCLIP: Improving Compositional Reasoning of. Advances in Neural Information Processing Systems 38: Annual Conference on Neural Information ...

  51. [59]

    FastText.zip: Compressing text classification models , journal =

    Armand Joulin and Edouard Grave and Piotr Bojanowski and Matthijs Douze and Herv. FastText.zip: Compressing text classification models , journal =. 2016 , url =. 1612.03651 , timestamp =

  52. [60]

    Read and Attend: Temporal Localisation in Sign Language Videos , booktitle =

    G. Read and Attend: Temporal Localisation in Sign Language Videos , booktitle =. 2021 , url =. doi:10.1109/CVPR46437.2021.01658 , timestamp =

  53. [61]

    Wendy Sandler , title =. Lang. Linguistics Compass , volume =. 2012 , url =. doi:10.1002/LNC3.326 , timestamp =

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.