Pith. sign in

REVIEW 3 major objections 4 minor 40 references

SiLVERScore: Semantically-Aware Embeddings for Sign Language Generation Evaluation

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SiLVERScore measures sign language generation by scoring a generated sign video against its text reference in a joint embedding space, reaching 0.99 ROC AUC for correct versus random pairs while staying insensitive to word order and prosody

desk verdict Useful application of CiCo as a reference-free SLG metric, but the 'substantially outperforms' claim leans on an in-domain retrieval test rather than evidence that scores track human judgment. read the letter →

arxiv 2509.03791 v1 pith:SXHS2L7X submitted 2025-09-04 cs.CL cs.AI

Saki Imai , Mert İnan , Anthony Sicilia , Malihe Alikhani This is my paper · ORCID
classification cs.CLcs.AI
keywords signlanguagegenerationevaluationmetricsembedding-basedmetricback-translationcontrastivelearningprosodysemanticsimilarityPHOENIX-14T
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 tries to establish that sign language generation can be evaluated directly on the generated sign video, without translating it back to text, by scoring semantic alignment in a shared video-text embedding space. The proposed metric, SiLVERScore, uses the CiCo contrastive embedding model to compare a sign video against its text reference, and the paper reports that it separates correct from randomly mismatched pairs almost perfectly (ROC AUC 0.99, overlap under 7.4%) on PHOENIX-14T and CSL-Daily. The paper also argues that SiLVERScore remains stable when words are reordered without changing meaning and when prosodic intensity varies, while BLEU, ROUGE, BERTScore, and BLEURT all degrade. If true, this gives sign language generation an evaluation method that avoids error propagation from an intermediate recognizer and does not penalize expressive or paraphrastic signing. The paper's Limitations section acknowledges that no human evaluation was conducted, so the embedding space's similarity is assumed to stand in for human semantic judgment.

What carries the argument

The load-bearing object is CiCo's joint video-text embedding space. CiCo encodes sign video with a sliding-window sign encoder that fuses domain-agnostic I3D features with domain-aware features and passes them through a Transformer initialized from CLIP's ViT-B encoder; text is lowercased, byte-pair encoded, and translated into English before contrastive alignment with the video via an InfoNCE loss. SiLVERScore uses the resulting fine-grained similarity matrix E(i,j)=Si·Wj^T, applies softmax re-weighting, and takes the averaged row-wise sum ZV2T as the video-to-text similarity, scaled by 3.5 to [0,100]. This shared space is what lets the metric compare video directly with text instead of fir

What would settle it

Have native or fluent signers rate generated sign videos as correct or incorrect and compare their ratings with SiLVERScore. If meaning-preserving reorderings that humans judge equivalent receive very different scores, or if videos with reversed referents or negations that humans judge wrong receive higher scores than the correct video, the near-perfect discrimination is a retrieval artifact rather than evaluation accuracy. A direct test would be to swap the subject and object in a directional-verb signing, as in the paper's Figure 1 example, and check whether SiLVERScore ranks the swapped vid

Watch

Extended reading notes

Core claim

The central claim is that semantic correctness of generated sign language can be read off the cosine similarity between the sign video and its text reference in a joint embedding space, and that this similarity is a better evaluation signal than any text-based comparison of back-translated output. Concretely, SiLVERScore is defined as the scaled video-to-text similarity ZV2T (with a scaling factor of 3.5) from the CiCo model, which aligns sign videos and CLIP-based text embeddings through contrastive learning with local fine-grained alignment. In tests, the score achieved 0.99 ROC AUC and 6.85% distribution overlap on PHOENIX-14T and 7.40% overlap on CSL-Daily for distinguishing correct vers

Load-bearing premise

The paper assumes that similarity in the trained embedding space is the same thing as semantic correctness in sign language, and it never checks this against human raters.

Editorial extensions

If this is right

  • If SiLVERScore is accepted as an evaluation metric, sign language generation systems can be compared directly on the generated video, removing the need to train or trust a sign-to-text recognizer in the evaluation loop.
  • Development loops could use SiLVERScore as a training signal or early-stopping criterion, since it separates correct from wrong outputs nearly perfectly on the tested datasets.
  • Evaluation will stop systematically penalizing expressive signing: prosody-rich outputs no longer score lower simply because they carry more prosodic intensity.
  • Embedding weights must be treated as dataset-specific: the paper's generalization results imply that one set of embedding weights cannot be assumed to work across sign languages or domains without fine-tuning.
  • The near-perfect separation between correct and random pairs provides a sanity-check baseline for future sign language generation metrics.

Reading between the lines

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

  • If SiLVERScore's embedding similarity is validated against human judgments, it could serve as a low-cost evaluation proxy for sign language generation, but that validation is still required and is absent from the paper.
  • Because CiCo is trained on PHOENIX-14T and CSL-Daily, the near-perfect discrimination numbers likely reflect in-domain alignment; applying SiLVERScore to a new sign language would probably require fine-tuning, consistent with the paper's own generalization findings.
  • The metric's insensitivity to prosody could be double-edged: it prevents expressive signing from being unfairly penalized, but it also means the metric cannot detect prosodic mistakes that change pragmatic meaning, so a production system would need a prosody-aware companion score.
  • The reordering robustness suggests SiLVERScore might also tolerate synonym substitution and paraphrase, which could be tested directly by extending the GPT-4o reordering experiment to full rewrites.
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

3 major / 4 minor

Summary. This paper proposes SiLVERScore, an evaluation metric for sign language generation that computes cosine similarity between a generated sign video and reference text in the CiCo joint embedding space, scaled by 3.5. The authors compare it with back-translation-based metrics (BLEU, ROUGE, BERTScore, BLEURT) on PHOENIX-14T and CSL-Daily, reporting 0.99 ROC AUC and <7.4% overlap for correct vs. random pairs, robustness to GPT-4o-generated word reorderings, and no correlation with annotated prosodic intensity. A final section studies SignCLIP's cross-dataset generalization and concludes dataset-specific tailoring is needed. The paper acknowledges in Limitations that no human evaluation was conducted and that SiLVERScore relies on English translations.

Significance. SiLVERScore addresses a real problem: back-translation-based SLG evaluation suffers from error propagation and cannot directly inspect the video. The paper ships code and includes a broad set of baselines, which is valuable. However, the strongest claims rest on proxy tasks. The correct-vs-random test is an in-domain retrieval test for a contrastive model trained on the same datasets, so the 0.99 AUC is not surprising and does not establish that the metric tracks human judgment of generation quality. The paper is transparent about the missing human evaluation, but without it the 'substantially outperforming' claim is not yet supported. The section on generalization tests SignCLIP rather than SiLVERScore, further limiting the current contribution.

major comments (3)
  1. [§4.1 and Table 1] The correct-vs-random experiment is an in-domain retrieval test. CiCo (§3) was trained with InfoNCE on PHOENIX-14T and CSL-Daily to maximize similarity of exactly the matched video-text pairs that Table 1 uses as 'correct' pairs. Near-perfect separation (AUC 0.99, overlap 6.85–7.40%) is thus an expected property of the training objective, not evidence that SiLVERScore measures the quality of generated signing. To support the central claim, the paper needs either human correlation or evaluation on outputs of actual sign generation models. The Limitations section explicitly concedes the absence of human evaluation; this is not a presentation issue but a missing validation of the metric's core validity assumption.
  2. [§4.2 and Table 1] The reordering experiment's construction is ambiguous. The text says scores are computed 'by comparing the ground-truth references with their corresponding hypotheses' and then refers to 'reordered hypotheses' without stating whether these are outputs of a generation model or GPT-4o's paraphrases of the references. If they are the latter, the test shows invariance to textual rephrasing, not sensitivity to semantic errors in a generated video. In addition, reporting higher overlap and lower AUC as desirable for 'Original vs. Reordered' is not justified: any metric that assigns a nearly constant score to all items would trivially produce high overlap and AUC around 0.5. Please clarify the protocol and define what property of a good evaluation metric this experiment is intended to measure.
  3. [§5 and §5.2] The generalization experiments evaluate SignCLIP, not SiLVERScore/CiCo. The conclusion in §5.2 that 'tailoring metrics to specific datasets is necessary' directly limits the scope of SiLVERScore, yet no experiment tests SiLVERScore on a dataset outside CiCo's training distribution. As a result, the paper's fourth contribution ('exploring generalization challenges across datasets') does not support the generality of the proposed metric itself; it reports obstacles for a different model. This matters because CiCo's domain-specific training is also what makes the §4.1 results easy to achieve.
minor comments (4)
  1. [§3, Eq. (2)] The softmax in Eq. (2) is not specified over which dimension; please state whether it is over rows, columns, or uses a temperature. Also, the scaling factor 3.5 is mentioned but no final SiLVERScore formula is given, making reproducibility harder.
  2. [§4.2 and Table 1] The text and table use both 'Original vs. Reordered' and 'hypotheses' without consistent terminology. Please specify the number of examples and the exact configuration of GPT-4o used.
  3. [Abstract / Table 1] The abstract says 'overlap < 7%' but Table 1 reports 7.40% for CSL-Daily; either the abstract is inaccurate or it should say '< 8%'.
  4. [Appendix G] The case studies report failures on negation and short utterances (e.g., PHOENIX-14T case 4, CSL-Daily case 4). These failure modes should be discussed in the main text, as they qualify the claim of semantic sensitivity.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline correct-vs-random validation reduces to CiCo's own contrastive training objective on the same datasets; the near-perfect discrimination is an in-domain retrieval result, not independent evidence of evaluation quality.

  1. fitted input called prediction [Sec. 3 (SiLVERScore definition) and Sec. 4.1 (correct-vs-random experiment)]
    "The similarities are computed using CiCo (Cheng et al., 2023), a model that leverages contrastive learning to align video and text representations. ... CiCo aligns video and text embeddings through a contrastive learning objective based on InfoNCE loss, which maximizes the similarity of matched video-text pairs while minimizing the similarity of unmatched pairs. ... SiLVERScore achieves near-perfect discrimination between correct and random pairs (ROC AUC = 0.99, overlap < 7%)"

    SiLVERScore is defined as the ZV2T similarity from CiCo, which was contrastively trained on PHOENIX-14T and CSL-Daily to maximize matched video-text similarity and minimize unmatched similarity. Section 4.1 then scores 'correct' ground-truth pairs against 'randomly paired' samples from those same datasets. Separating exactly these two classes is the InfoNCE training objective itself, so the near-perfect AUC/low overlap reports how well the fitted retrieval model performs its own in-domain task. The 'correct' labels are the same correspondence labels used as positive training examples, and the 'random' labels are the negatives. Thus the headline discrimination result is by construction a retrieval fit, not an independent validation that the metric tracks human judgments of sign language gen

full rationale

The score definition itself is not mathematically circular: SiLVERScore is a scaled cosine similarity in a joint embedding space, which is a reasonable design choice. However, the paper's central empirical claim—near-perfect correct-vs-random discrimination—is circular as a validation of evaluation quality. The metric's embedding was trained with InfoNCE on exactly the matched-vs-unmatched distinction that Section 4.1 measures, on the same two datasets. The high AUC is therefore an expected consequence of fitting the embedding to that very task, rather than evidence that SiLVERScore correlates with human notions of semantic or linguistic correctness. The paper's Limitations section explicitly concedes 'the absence of human evaluation,' confirming that no external criterion is used. The reordering and prosody experiments are more independent, using GPT-4o paraphrases and human prosody annotations (from Inan et al. 2022, a self-citation but an empirical dataset), but they test invariance properties, not whether the metric ranks generation quality correctly. The generalization section honestly shows that contrastive sign embeddings fail to transfer across datasets, which further supports the view that the strong results are in-distribution. Overall, the main headline result reduces to the fitted model's training objective, while other contributions retain some independent content; hence a score of 6.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the assumption that the CiCo embedding space is a valid semantic proxy, plus the reliability of translation and paraphrase tools used to build test inputs. No new theoretical entities are introduced.

free parameters (1)
  • SiLVERScore scaling factor = 3.5
    Hand-chosen multiplier to scale scores to [0,100], following CLIP-Score. Does not affect rank-based metrics (AUC, overlap).
assumptions (3)
  • domain assumption CiCo's joint embedding space faithfully represents semantic and prosodic content of sign language
    The metric is defined as similarity in the CiCo embedding space (section 3). If the space does not align with human linguistic judgments, the metric's validity is undermined. The paper notes lack of human evaluation in Limitations.
  • domain assumption Google Translate preserves meaning when translating German/Chinese annotations to English
    Texts are translated to English to match CLIP pretraining (section 3, Model Details). Translation errors could corrupt the text side of the embedding alignment.
  • domain assumption GPT-4o reorderings preserve meaning exactly
    Used to generate meaning-preserving reordered sentences in section 4.2; if reorderings change meaning, the robustness test is invalid.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SiLVERScore: Semantically-Aware Embeddings for Sign Language Generation Evaluation." pith.science (2026). https://pith.science/paper/SXHS2L7X

@misc{pith2026250903791,
  author       = {Pith},
  title        = {Pith review of: SiLVERScore: Semantically-Aware Embeddings for Sign Language Generation Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SXHS2L7X}},
  note         = {Machine review of arXiv:2509.03791}
}
read the original abstract

Evaluating sign language generation is often done through back-translation, where generated signs are first recognized back to text and then compared to a reference using text-based metrics. However, this two-step evaluation pipeline introduces ambiguity: it not only fails to capture the multimodal nature of sign language-such as facial expressions, spatial grammar, and prosody-but also makes it hard to pinpoint whether evaluation errors come from sign generation model or the translation system used to assess it. In this work, we propose SiLVERScore, a novel semantically-aware embedding-based evaluation metric that assesses sign language generation in a joint embedding space. Our contributions include: (1) identifying limitations of existing metrics, (2) introducing SiLVERScore for semantically-aware evaluation, (3) demonstrating its robustness to semantic and prosodic variations, and (4) exploring generalization challenges across datasets. On PHOENIX-14T and CSL-Daily datasets, SiLVERScore achieves near-perfect discrimination between correct and random pairs (ROC AUC = 0.99, overlap < 7%), substantially outperforming traditional metrics.

Figures

Figures reproduced from arXiv: 2509.03791 by the authors.

Figure 1
Figure 1. In this example, a sign language genera [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Kernel Density Plots for different metrics. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Kernel Density plots comparing the score dis [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Box plots showing the distribution of SiLVERScore, BLEU-1, and ROUGE scores across three prosody [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: illustrates the cosine similarity between video and text embeddings. Ideally, high similar￾ity values should appear along the diagonal, in￾dicating alignment between corresponding video￾text pairs. Before fine-tuning, the heatmaps dis￾play low, diffuse similarity score…
Figure 6
Figure 6. Figure 6: Kernel density plots for BLEU-1 (top) and [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Kernel Density Plots for different metrics for CSL-Daily dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Kernel Density plots comparing the score dis [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 10
Figure 10. Figure 10: shows a heatmap of SignCLIP em￾bedding cosine similarity scores for the WMTSLT dataset. After fine-tuning, the model appears to overfit, assigning high similarity scores to many pairs. D Training Details Hardware and Compute All training and infer￾ence computations we…
Figure 11
Figure 11. Figure 11: Box plots showing the distribution of BLEU-2, BLEU-3, BLEU-4, BERTScore, and BLEURT scores [PITH_FULL_IMAGE:figures/full_fig_p014_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 27 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudreault, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, et al. 2019. Sign language recognition, generation, and translation: An interdisciplinary perspective. In Proceedings of the 21st international ACM SIGACCESS conference on computers and accessibi...

    work page 2019

  4. [4]

    Necati Cihan Camgoz, Simon Hadfield, Oscar Koller, Hermann Ney, and Richard Bowden. 2018. Neural sign language translation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    work page 2018

  5. [5]

    Yiting Cheng, Fangyun Wei, Jianmin Bao, Dong Chen, and Wenqiang Zhang. 2023. https://doi.org/10.1109/CVPR52729.2023.01823 CiCo: Domain-Aware Sign Language Retrieval via Cross-Lingual Contrastive Learning . In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19016--19026, Los Alamitos, CA, USA. IEEE Computer Society

    arXiv 2023

  6. [6]

    Ashley Chow, Glenn Cameron, Mark Sherwood, Phil Culliton, Sam Sepah, Sohier Dane, and Thad Starner. 2023. Google - isolated sign language recognition. https://kaggle.com/competitions/asl-signs. Kaggle

    work page 2023

  7. [7]

    Minakov, Vanessa Milan, Chinmay Singh, Kriston Pumphrey, Richard E

    Aashaka Desai, Lauren Berger, Fyodor O. Minakov, Vanessa Milan, Chinmay Singh, Kriston Pumphrey, Richard E. Ladner, Hal Daum\' e , Alex X. Lu, Naomi Caselli, and Danielle Bragg. 2024. Asl citizen: a community-sourced dataset for advancing isolated sign language recognition. In Proceedings of the 37th International Conference on Neural Information Processi...

    work page 2024

  8. [9]

    Mo Guan, Yan Wang, Guangkun Ma, Jiarui Liu, and Mingzu Sun. 2024. https://api.semanticscholar.org/CorpusID:269635215 Multi-stream keypoint attention network for sign language recognition and translation . ArXiv, abs/2405.05672

    arXiv 2024

Show all 40 references
  1. [10]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.595 CLIPS core: A reference-free evaluation metric for image captioning . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language ...

    2021 doi

  2. [11]

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. 2017. Gans trained by a two time-scale update rule converge to a local nash equilibrium. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIP...

    2017

  3. [12]

    Wencan Huang, Wenwen Pan, Zhou Zhao, and Qi Tian. 2021. Towards fast and high-quality sign language production. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3172--3181

    2021

  4. [13]

    Matt Huenerfauth, Liming Zhao, Erdan Gu, and Jan Allbeck. 2008. Evaluation of american sign language generation by native asl signers. ACM Transactions on Accessible Computing (TACCESS), 1(1):1--27

    2008

  5. [14]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. Gpt-4o system card. arXiv preprint arXiv:2410.21276

    2024 arXiv

  6. [15]

    Eui Jun Hwang, Jung Ho Kim, Suk Min Cho, and Jong C Park. 2022. Non-autoregressive sign language production via knowledge distillation. arXiv preprint arXiv:2208.06183

    2022 arXiv

  7. [16]

    Eui Jun Hwang, Huije Lee, and Jong C. Park. 2024. https://doi.org/10.1109/FG59268.2024.10581980 A gloss-free sign language production with discrete representation . In 2024 IEEE 18th International Conference on Automatic Face and Gesture Recognition (FG), pages 1--6

    2024

  8. [17]

    Mert Inan, Yang Zhong, Sabit Hassan, Lorna Quandt, and Malihe Alikhani. 2022. https://doi.org/10.18653/v1/2022.findings-acl.228 Modeling intensification for sign language generation: A computational approach . In Findings of the Association for Computational Linguistics: ACL 2...

    2022 doi

  9. [18]

    Zifan Jiang, Gerard Sant, Amit Moryossef, Mathias M \"u ller, Rico Sennrich, and Sarah Ebling. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.518 S ign CLIP : Connecting text and sign language by contrastive learning . In Proceedings of the 2024 Conference on Empirical Meth...

    2024 doi

  10. [19]

    Jung-Ho Kim, Mathew Huerta-Enochian, Changyong Ko, and Du Hui Lee. 2024. SignBLEU : Automatic evaluation of multi-channel sign language translation. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-...

    2024

  11. [20]

    Harold W Kuhn. 1955. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83--97

    1955

  12. [21]

    Scott K. Liddell. 2003. Grammar, Gesture, and Meaning in American Sign Language. Cambridge University Press

    2003

  13. [22]

    Li Liu, Lufei Gao, Wentao Lei, Fengji Ma, Xiaotian Lin, and Jinting Wang. 2023. A survey on deep multi-modal learning for body language recognition and generation. arXiv preprint arXiv:2308.08849

    2023 arXiv

  14. [23]

    Müller Mathias, Ebling Sarah, Camgöz Necati Cihan, Jiang Zifan, Battisti Alessia, Moryossef Amit, Rios Annette, Bowden Richard, and Wong Ryan. 2022. https://doi.org/10.5281/zenodo.6621480 Wmt-slt focusnews: Training data for the wmt shared task on sign language translation

    2022 doi

  15. [24]

    Amit Moryossef, Rotem Zilberman, and Ohad Langer. 2024. signwriting-evaluation: Effective sign language evaluation via signwriting. arXiv preprint arXiv:2410.13668

    2024 arXiv

  16. [25]

    B Natarajan, E Rajalakshmi, R Elakkiya, Ketan Kotecha, Ajith Abraham, Lubna Abdelkareim Gabralla, and V Subramaniyaswamy. 2022. Development of an end-to-end deep learning framework for sign language recognition, translation, and video generation. IEEE Access, 10:104358--104374

    2022

  17. [26]

    Matt Post. 2018. https://doi.org/10.18653/v1/W18-6319 A call for clarity in reporting BLEU scores . In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186--191, Brussels, Belgium. Association for Computational Linguistics

    2018 doi

  18. [27]

    Amy Pu, Hyung Won Chung, Ankur P Parikh, Sebastian Gehrmann, and Thibault Sellam. 2021. Learning compact metrics for mt. In Proceedings of EMNLP

    2021

  19. [28]

    Wendy Sandler. 2012. https://doi.org/10.1002/lnc3.326 The phonological organization of sign languages . Language and Linguistics Compass, 6(3):162--182. Epub 2012 Mar 2

    2012 doi

  20. [29]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2020. https://doi.org/10.1007/978-3-030-58621-8_40 Progressive transformers for end-to-end sign language production . In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, ...

    2020 doi

  21. [30]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2021. https://doi.org/10.1007/s11263-021-01457-9 Continuous 3d multi-channel sign language production via progressive transformers and mixture density networks . Int. J. Comput. Vision, 129(7):2113–2135

    2021 doi

  22. [31]

    Thibault Sellam, Dipanjan Das, and Ankur P Parikh. 2020. Bleurt: Learning robust metrics for text generation. In Proceedings of ACL

    2020

  23. [32]

    Garrett Tanzer, Maximus Shengelia, Ken Harrenstien, and David Uthus. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.360 Reconsidering sentence-level sign language translation . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages ...

    2024 doi

  24. [33]

    Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Rapha \"e l Marinier, Marcin Michalski, and Sylvain Gelly. 2019. Fvd: A new metric for video generation

    2019

  25. [34]

    Gül Varol, Liliane Momeni, Samuel Albanie, Triantafyllos Afouras, and Andrew Zisserman. 2021. https://doi.org/10.1109/CVPR46437.2021.01658 Read and attend: Temporal localisation in sign language videos . In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (C...

    2021

  26. [35]

    Carla Viegas, Mert Inan, Lorna Quandt, and Malihe Alikhani. 2023. https://doi.org/10.18653/v1/2023.starsem-1.1 Including facial expressions in contextual embeddings for sign language generation . In Proceedings of the 12th Joint Conference on Lexical and Computational Semantic...

    2023 doi

  27. [36]

    Pan Xie, Taiying Peng, Yao Du, and Qipeng Zhang. 2024. https://doi.org/10.1109/WACV57701.2024.00300 Sign Language Production with Latent Motion Transformer . In 2024 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3012--3022, Los Alamitos, CA, USA. ...

    2024

  28. [37]

    Jinhui Ye, Xing Wang, Wenxiang Jiao, Junwei Liang, and Hui Xiong. 2024. Improving gloss-free sign language translation by reducing representation density

    2024

  29. [38]

    Kayo Yin, Amit Moryossef, Julie Hochgesang, Yoav Goldberg, and Malihe Alikhani. 2021. https://doi.org/10.18653/v1/2021.acl-long.570 Including signed languages in natural language processing . In Proceedings of the 59th Annual Meeting of the Association for Computational Lingui...

    2021 doi

  30. [39]

    Youngwoo Yoon, Bok Cha, Joo-Haeng Lee, Minsu Jang, Jaeyeon Lee, Jaehong Kim, and Geehyuk Lee. 2020. https://api.semanticscholar.org/CorpusID:221507915 Speech gesture generation from the trimodal context of text, audio, and speaker identity . ACM Transactions on Graphics (TOG),...

    2020

  31. [40]

    Weinberger, and Yoav Artzi

    Tianyi Zhang*, Varsha Kishore*, Felix Wu*, Kilian Q. Weinberger, and Yoav Artzi. 2020. https://openreview.net/forum?id=SkeHuCVFDr Bertscore: Evaluating text generation with bert . In International Conference on Learning Representations

    2020

  32. [41]

    Hao Zhou, Wengang Zhou, Weizhen Qi, Junfu Pu, and Houqiang Li. 2021. Improving sign language translation with monolingual data by sign back-translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1316--1325

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.