Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Phonological Representation Learning for Isolated Signs Improves Out-of-Vocabulary Generalization

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

Pith's one-line read Phonological inductive biases help a vector-quantized autoencoder generalize to unseen sign language words.

desk verdict Sensible idea, honest ablations, but the headline reconstruction claim is not comparable across conditions as reported — the recognition result is the more defensible finding. read the letter →

arxiv 2509.04745 v1 pith:ECX3BFB6 submitted 2025-09-05 cs.CL cs.CV

classification cs.CLcs.CV
keywords signlanguagevectorquantizationout-of-vocabularygeneralizationphonologicalrepresentationisolatedrecognitionASL-LEXdisentangledlearningsemi-supervision
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 show that a sign-language model can generalize to words it never saw during training if its learned representations are shaped by linguistic structure. It builds a vector-quantized autoencoder that encodes pose data as discrete codes, then adds two phonological biases: separating the input into parallel streams for handshape, location, and movement, and weakly supervising the codebooks with expert phonological labels from ASL-LEX 2.0. The central finding is that these two biases work in complementary ways: the architectural separation improves reconstruction of novel signs, while the phonological supervision improves the discriminability of the codes for recognizing unseen signs. Combined, they reduce reconstruction error on out-of-vocabulary signs and raise recognition accuracy relative to a controlled baseline. The paper argues that purely data-driven discrete sign representations tend to capture spurious correlations, and that explicit, linguistically-motivated priors are a principled remedy.

What carries the argument

The central object is VQ-ASL, a vector-quantized variational autoencoder with discrete codebooks, augmented by two phonological mechanisms. Parameter Disentanglement is an architectural bias that partitions the pose input and latent space into separate streams for the right/left hands, movement, body, and non-manual markers, each with its own codebook and with left/right codebooks shared. Phonological Semi-Supervision is a regularization technique that pre-assigns some codebook entries to expert phonological labels from ASL-LEX 2.0 and, with some probability, forces the quantizer to select the label's code, aligning the representation with linguistic features. The VQ bottleneck, the forced c

What would settle it

Train the full VQ-ASL model with the same forced-code procedure but with phonological labels randomly permuted across signs, keeping the same codebook size and forcing probability. If the out-of-vocabulary recognition and reconstruction gains persist under shuffled labels, the reported improvements are not due to genuine phonological structure but to the supervision regularizer itself.

Watch

Extended reading notes

Core claim

The paper's central claim is that two phonological inductive biases, Parameter Disentanglement and Phonological Semi-Supervision, improve out-of-vocabulary generalization of a Transformer-based VQ-VAE for isolated ASL signs. Parameter Disentanglement restructures the model into parallel encoder-decoder-codebook streams, one for each major articulatory parameter, while Phonological Semi-Supervision forces a subset of quantized codes to align with expert-annotated ASL phonological features during training. On the Sem-Lex benchmark's unseen-gloss split, the full model reduces reconstruction MSE from 0.058 (baseline) to 0.041 and raises out-of-vocabulary sign-recognition MRR from .381 to .452. T

Load-bearing premise

The success of Phonological Semi-Supervision rests on the belief that forcing the quantizer to pick the expert-labeled code during training teaches a visual-to-phonology mapping that transfers to unseen signs, rather than merely injecting the label information as a shortcut that inflates the reported gains.

Editorial extensions

If this is right

  • If the central claim is right, vector-quantized sign representations learned under phonological biases can serve as pre-trained tokenizers for downstream continuous sign recognition and translation, where vocabulary coverage is limited.
  • The complementary effects of architectural and semantic biases suggest that purely structural inductive biases are not sufficient for discriminative tasks; semantic supervision is needed to make discrete codes useful for recognition.
  • The positive correlation between phonological-feature recognition and isolated-sign recognition on unseen signs supports the hypothesis that learning phonology helps generalization, not just memorization of training vocabulary.
  • The reconstruction-recognition trade-off identified in the paper implies that models trained only for generative fidelity may not automatically produce representations that are discriminative for unseen signs, and vice versa.
  • The success of the forced-code supervision suggests that expert linguistic labels can be injected into discrete codebook training without destroying reconstruction quality, opening a path for using structured linguistic knowledge in other low-resource gesture and motion domains.

Reading between the lines

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

  • An untested but natural extension: varying the forcing probability during Phonological Semi-Supervision may reveal a trade-off curve between reconstruction fidelity and OOV discriminability, letting a practitioner choose the operating point for a target task.
  • The shared left/right codebook symmetry constraint suggests a testable hypothesis that signs performed with the non-dominant hand should exhibit the same OOV generalization benefits, which could be verified on mirrored or left-handed signing data if available.
  • Because the paper's phonological labels come from a single sign-language database, the framework could be stress-tested by transferring the same codebook-alignment procedure to another sign language with its own phonological inventory, which would indicate whether the benefit comes from general phonological structure or ASL-specific label statistics.
  • The phonotactic hypothesis sketched in the appendix could be operationalized as a concrete falsifier: feeding random permutations of learned codes to the decoder and measuring reconstruction error would test whether the codebooks implicitly encode legal versus illegal feature combinations.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes a vector-quantized autoencoder (VQ-VAE) for isolated sign language pose data, augmented with two phonological inductive biases: Parameter Disentanglement (PD), an architectural multi-stream factorization, and Phonological Semi-Supervision (PSS), a regularization that forces quantizer codes to align with ASL-LEX 2.0 phonological labels. The authors evaluate four configurations (Baseline, VQ-ASL-PD, VQ-ASL-PSS, and the full VQ-ASL) on the Sem-Lex benchmark using a vocabulary-disjoint split, reporting that the full model improves OOV reconstruction MSE from 0.058 to 0.041 and OOV sign-recognition MRR from .381 to .452. The paper interprets these results as evidence that explicit linguistic priors improve compositional generalization for unseen signs.

Significance. If the central claims hold, the paper makes a useful contribution to sign-language representation learning: it provides a controlled ablation of two linguistically motivated biases, uses a vocabulary-disjoint split that directly targets OOV generalization, and reports codebook-utilization strategies to mitigate VQ collapse. The connection to Locatello et al.'s impossibility result is well motivated. However, the strength of the conclusions is currently limited by a metric-comparability problem in the reconstruction evaluation, a partially circular PFR probe, an overstated abstract claim about one-shot reconstruction, and the absence of any uncertainty quantification. These issues are fixable within the paper's scope, so the central idea remains defensible but needs revision.

major comments (4)
  1. [§3.2.1 / Table 2] The reconstruction comparison is not apples-to-apples. The baseline operates on raw absolute pose, while PD streams are preprocessed: hands/face are re-centered to wrist/nose, movement streams contain wrist keypoints at all frames, and 'the total reconstruction loss is the sum of the reconstruction losses from each individual stream.' Thus the reported 'Overall MSE' of 0.046 for VQ-ASL-PD and 0.041 for the full model is computed on transformed targets, not the same target as the baseline's 0.058. The 21% reconstruction improvement claimed in §5.1 is therefore not established. Please recompute all models' MSE after inverting the per-stream preprocessing, or report per-stream errors in a common coordinate frame.
  2. [Abstract / §4.3] The abstract claims the learned representations are 'more effective for one-shot reconstruction of unseen signs,' but no one-shot evaluation protocol appears anywhere in the paper. The reconstruction metric in Table 2 is standard held-out MSE after training, not one-shot reconstruction from a single example. Either implement and report a genuine one-shot protocol (e.g., reconstructing a novel sign from one exposure) or revise the abstract to remove 'one-shot.'
  3. [§4.3 / Table 3] The PFR probe is partially circular for the supervised models. In §3.2.2, PSS forces the quantizer to select the code vector corresponding to the ASL-LEX ground-truth label with a specified probability. Consequently, high PFR accuracy for VQ-ASL-PSS and VQ-ASL (Full) largely measures whether the model reproduced the injected label signal, not whether phonological structure emerged independently. The sentence in §4.3 calling the PFR probe 'confirmation that the ASL-LEX 2.0 features generalize to unseen signs' overstates this. Please report PFR on codebook entries that were never forced, or ablate the forcing probability; at minimum, the paper should explicitly describe PFR for supervised models as verification of label injection rather than independent evidence of phonological alignment.
  4. [§5 / Tables 2 and 3] All results are single-run point estimates with no variance or significance testing. The differences that support the central claims are small (e.g., OOV MRR .452 vs .435; OOV MSE 0.041 vs 0.046) and may be within noise. Please add repeated runs with multiple seeds (e.g., 5 or more), report confidence intervals or standard deviations, and perform a significance test for the headline comparisons. The hyperparameter-search procedure on the baseline only is transparent, but without seed variance it is hard to assess whether the reported improvements are robust.
minor comments (4)
  1. [Table 1 / Appendix B] The PSS forcing probability, the PSS auxiliary loss weight, and the Gumbel temperature schedule are not reported in Table 1 or in the implementation details. These are important for reproducibility and for assessing the label-shortcut risk; please include their final values.
  2. [§4.3] The first probe is sometimes called 'MISR' (with a leading M) in the text and 'ISR' in the tables. Use consistent notation throughout.
  3. [Throughout] Minor typographical/style issues: 'VQ-V AEs' should be 'VQ-VAEs' or 'VQ-VAEs'; the spacing between 'VQ-V' and 'AEs' is inconsistent. Also, Table 2's 'Channel-wise Test MSE' columns are only populated for the PD models; consider adding a note explaining why the baseline and PSS rows have no channel-wise breakdown.
  4. [§6.3] The limitations section is honest and appropriately scoped, but it does not mention the metric-comparability issue with reconstruction loss or the absence of multiple seeds. A brief addition acknowledging these evaluation limitations would align the discussion with the actual evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the core OOV reconstruction and recognition evaluations are self-contained against a held-out vocabulary split.

full rationale

The paper's central claims (reconstruction MSE reduction and OOV sign-recognition MRR improvement) are evaluated on a disjoint 'unseen gloss' split, so the predictions are not equivalent to training inputs by construction. Phonological Semi-Supervision does force the quantizer to select pre-assigned ASL-LEX label codes during training, but the paper explicitly labels the PFR probe on supervised models as 'an intrinsic evaluation of how successfully the PSS objective was learned' rather than as evidence of emergent phonological structure; the OOV PFR and ISR scores still require the encoder to generalize to signs never seen in training. The reconstruction gain attributed to Parameter Disentanglement may involve a preprocessing comparability concern (per-stream targets are re-centered or use wrist-only keypoints), but that is a measurement-validity issue, not a circular reduction of the prediction to the input. Self-citations to the Sem-Lex benchmark and prior phonology work provide dataset and motivation, but the controlled ablation against a baseline and the held-out OOV evaluation supply independent content. No load-bearing step reduces to a fit, a self-citation chain, or a definitional identity.

Assumptions & free parameters 13 free parameters · 7 assumptions · 0 invented entities

The central claim rests on two engineered training signals, PD stream factorization and PSS label forcing, plus a benchmark split. Those interventions add several hand-chosen hyperparameters and domain assumptions; the model contributes the architecture and evaluation, not a new derived law.

free parameters (13)
  • Commitment cost beta = 3e-6
    Tuned on baseline validation set via Optuna; sets codebook commitment loss weight (Table 1).
  • Diversity weight gamma = 3.0
    Tuned on baseline; weight for codebook diversity loss (Table 1).
  • Learning rate = 8.61e-5
    Tuned on baseline; Adam optimizer learning rate (Table 1).
  • Transformer hidden dim = 256
    Tuned on baseline; hidden dimension of Transformer layers (Table 1).
  • Transformer layers = 5
    Tuned on baseline; number of encoder/decoder layers (Table 1).
  • Latent dim Lc = 32
    Tuned on baseline; dimensionality of each code vector (Table 1).
  • Num latent vecs Np = 30
    Tuned on baseline; number of vectors in the bottleneck (Table 1).
  • Codebook size K = 200
    Tuned on baseline; entries per codebook (Table 1).
  • Dropout = 0.2
    Tuned on baseline; dropout probability in Transformers (Table 1).
  • PSS forcing probability = not reported
    Probability that the quantizer is forced to the ASL-LEX label code; value not given in Section 3.2.2 or Appendix B, but central to PSS behavior.
  • Per-stream codebook and latent allocation = not specified
    Allocation of Np and K across streams is 'proportional to descriptive complexity' (Appendix B), not exact values; affects all PD results.
  • PSS auxiliary loss weight = not reported
    No explicit weight for the phonological auxiliary loss is given; if it is integrated through forced code selection, the forcing schedule still needs specification.
  • Gumbel temperature schedule = not reported
    Temperature annealing schedule for Gumbel-Softmax in Appendix B.1 is not specified; affects codebook utilization and training dynamics.
assumptions (7)
  • domain assumption Locatello et al. (2019) impossibility theorem motivates the need for inductive biases and applies to this VQ setting.
    The paper cites the theorem as central motivation (Section 2.2) and assumes it transfers to a discrete, partially supervised OOV setting.
  • domain assumption Brentari's Prosodic Model factorization of signs into articulators, movement, and place of articulation is a valid and sufficient decomposition for pose-based OOV generalization.
    Used in Section 3.2.1 to define the streams and codebooks; if the factorization is wrong, the PD intervention's inductive bias is misplaced.
  • domain assumption ASL-LEX 2.0 phonological labels are accurate and complete enough to supervise codebooks and to probe OOV signs.
    PSS and the PFR probe depend on these labels (Sections 3.2.2 and 4.3); label noise or missing values would weaken supervision.
  • domain assumption The Sem-Lex unseen-gloss split ensures the test signs are genuinely out-of-vocabulary and that no signer or recording artifacts leak between splits.
    Section 4.1 defines the split only by vocabulary disjointness; no signer-disjoint or clip-level leakage check is reported.
  • domain assumption Pose keypoints from MediaPipe capture the phonological contrasts needed; the reduction is a phonetic bias, not a fatal information loss.
    Section 2.1 treats pose as a useful phonetic abstraction; if keypoints drop handshape or location distinctions, reconstruction and recognition gains are capped.
  • standard math Straight-through estimator and Gumbel-Softmax relaxation for VQ training are valid and produce usable gradients.
    The model relies on these standard VQ-VAE mechanisms (Section 3.1 and Appendix B.1).
  • ad hoc to paper Arbitrary pre-assignment of a subset of codes to ASL-LEX feature values is a meaningful initialization for PSS.
    Section 3.2.2 says codes are 'arbitrarily pre-assigned' to expert features; no stability or coverage analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Phonological Representation Learning for Isolated Signs Improves Out-of-Vocabulary Generalization." pith.science (2026). https://pith.science/paper/ECX3BFB6

@misc{pith2026250904745,
  author       = {Pith},
  title        = {Pith review of: Phonological Representation Learning for Isolated Signs Improves Out-of-Vocabulary Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ECX3BFB6}},
  note         = {Machine review of arXiv:2509.04745}
}
read the original abstract

Sign language datasets are often not representative in terms of vocabulary, underscoring the need for models that generalize to unseen signs. Vector quantization is a promising approach for learning discrete, token-like representations, but it has not been evaluated whether the learned units capture spurious correlations that hinder out-of-vocabulary performance. This work investigates two phonological inductive biases: Parameter Disentanglement, an architectural bias, and Phonological Semi-Supervision, a regularization technique, to improve isolated sign recognition of known signs and reconstruction quality of unseen signs with a vector-quantized autoencoder. The primary finding is that the learned representations from the proposed model are more effective for one-shot reconstruction of unseen signs and more discriminative for sign identification compared to a controlled baseline. This work provides a quantitative analysis of how explicit, linguistically-motivated biases can improve the generalization of learned representations of sign language.

Figures

Figures reproduced from arXiv: 2509.04745 by the authors.

Figure 1
Figure 1. We show that out-of-vocabulary signs are [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Brentari’s Prosodic Model (Brentari, 1998) is hierarchical; the boxed variables represent disentangled parameters in this work. function. A value of β > 1 imposes a stronger constraint on the information capacity of the latent bottleneck, forcing the model to learn a more effi￾cient, and therefore more disentangled, representa￾tion by balancing reconstruction accuracy against the statistical independence of the late… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Targeted Linguistic Analysis of Sign Language Models with Minimal Translation Pairs

    cs.CL 2026-04 unverdicted novelty 7.0 of 10

    Introduces ASL-MTP benchmark and shows a state-of-the-art ASL-to-English model relies strongly on manual cues while missing non-manual cues.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    Artem Abzaliev and Rada Mihalcea. 2024. Unsupervised discrete representations of american sign language. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 19786--19793

  2. [2]

    Takuya Akiba, Sota Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. 2019. Optuna : A next-generation hyperparameter optimization framework. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining

  3. [3]

    Danielle Bragg, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, and Meredith Ringel Morris. 2020. Exploring collection of sign language datasets: Privacy, participation, and model performance. In The 22nd International ACM SIGACCESS Conference on Computers and Accessibility, pages 1--14

  4. [4]

    Danielle Bragg, Oscar Koller, Mary Bellard, Larwan Berke, Patrick Boudrealt, Annelies Braffort, Naomi Caselli, Matt Huenerfauth, Hernisa Kacorri, Tessa Verhoef, Christian Vogler, and Meredith Ringel Morris. 2019. Sign language recognition, generation, and translation: An interdisciplinary perspective. In The 21st International ACM SIGACCESS Conference on ...

  5. [5]

    Diane Brentari. 1998. A Prosodic Model of Sign Language Phonology. MIT Press

  6. [6]

    Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. 2019. OpenPose : Realtime multi-person 2D pose estimation using part affinity fields. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(1):172--186

  7. [7]

    Mirella De Sisto, Vincent Vandeghinste, Santiago Egea G \' o mez, Mathieu De Coster, Dimitar Shterionov, and Horacio Saggion. 2022. Challenges with sign language datasets for sign language recognition and translation. In Proceedings of the Language Resources and Evaluation Conference ( LREC ) , pages 2711--2718

  8. [8]

    Manas Gaur, Kaushik Roy, Ugur Kursuncu, Ziad Al-Halah, Krishnaprasad Thirunarayan, and Amit Sheth. 2022. Advancing neuro-symbolic AI with knowledge-infused learning. IEEE Intelligent Systems, 37(5):103--110

Show all 32 references
  1. [9]

    Yutong Hao, Zunshu Liu, Minsi Li, Jiahui Wu, Chen Li, Shidong Zhang, Yidong Ge, and Ying Shan. 2021. Self-supervised learning on pose data for sign language recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops ( ICCVW )

  2. [10]

    Irina Higgins, Loic Matthey, Arka Pal, Christopher Burgess, Xavier Glorot, Matthew Botvinick, Shakir Mohamed, and Alexander Lerchner. 2017. beta-vae: Learning basic visual concepts with a constrained variational framework. In International Conference on Learning Representation...

  3. [11]

    Eric Jang, Shixiang Gu, and Ben Poole. 2017. Categorical reparameterization with gumbel-softmax. In International Conference on Learning Representations ( ICLR )

  4. [12]

    Songyao Jiang, Bin Sun, Lichen Wang, Yue Bai, Kunpeng Li, and Yun Fu. 2021. Skeleton aware multi-modal sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3432--3441

  5. [13]

    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...

  6. [14]

    Lee Kezar, Elana Pontecorvo, Adele Daniels, Connor Baer, Ruth Ferster, Lauren Berger, Jesse Thomason, Zed Sevcikova Sehyr, and Naomi Caselli. 2023 a . The Sem-Lex benchmark: Modeling ASL signs and their phonemes. In The 25th International ACM SIGACCESS Conference on Computers ...

  7. [15]

    Lee Kezar, Jesse Thomason, and Zed Sevcikova Sehyr. 2023 b . Improving sign recognition with phonology. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics

  8. [16]

    Pang Wei Koh, Thao Nguyen, Yew Siang Tang, Stephen Mussmann, Emma Pierson, Been Kim, and Percy Liang. 2020. Concept bottleneck models. In Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings of Machine Learning Research, pages 5338--5...

  9. [17]

    Francesco Locatello, Stefan Bauer, Mario Lucic, Gunnar Raetsch, Sylvain Gelly, Bernhard Sch \"o lkopf, and Olivier Bachem. 2019. Challenging common assumptions in the unsupervised learning of disentangled representations. In Proceedings of the 36th International Conference on ...

  10. [18]

    Camillo Lugaresi, Jiuqiang Tang, Hadon Nash, Chris McClanahan, Esha Uboweja, Michael Hays, Fan Zhang, Chuo-Ling Chang, Ming Guang Yong, Juhyun Lee, Wan-Teh Chang, Wei Hua, Manfred Georg, and Matthias Grundmann. 2019. MediaPipe : A framework for building perception pipelines. a...

  11. [19]

    Maddison, Andriy Mnih, and Yee Whye Teh

    Chris J. Maddison, Andriy Mnih, and Yee Whye Teh. 2017. The concrete distribution: A continuous relaxation of discrete random variables. In International Conference on Learning Representations ( ICLR )

  12. [20]

    Amit Moryossef, Artem Abzaliev, Rada Mihalcea, and Yoav Goldberg. 2021. Unsupervised discrete representations of american sign language. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 4531--4537

  13. [21]

    J. Pu, W. Zhou, and H. Li. 2021. Learning to sign by watching tv. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition ( CVPR )

  14. [22]

    Ali Razavi, Aaron van den Oord, and Oriol Vinyals. 2019. Generating diverse high-fidelity images with VQ-VAE-2 . In Advances in Neural Information Processing Systems

  15. [23]

    Ben Saunders, Necati Cihan Camgoz, and Richard Bowden. 2020. Everybody sign now: Translating spoken language to photo-realistic sign language video. arXiv preprint arXiv:2011.09846

  16. [24]

    Cohen-Goldberg, and Karen Emmorey

    Zed Sehyr, Naomi Caselli, Ariel M. Cohen-Goldberg, and Karen Emmorey. 2021. The ASL-LEX 2.0 project: A database of lexical and phonological properties for 2,723 signs in american sign language. Journal of Deaf Studies and Deaf Education, 26(2):263--277

  17. [25]

    Stokoe, Jr

    William C. Stokoe, Jr. 1960. Sign Language Structure: An Outline of the Visual Communication Systems of the American Deaf, volume 8 of Studies in Linguistics: Occasional Papers. Linstok Press

  18. [26]

    Valerie Sutton. 1990. Sign writing for everyday use. Deaf Action Committee for Sign Writing

  19. [27]

    Sumeyye Meryem Tasyurek, Tugce Kiziltepe, and Hacer Yalim Keles. 2025. Disentangle and regularize: Sign language production with articulator-based disentanglement and channel-aware regularization. arXiv preprint arXiv:2504.06610

  20. [28]

    Aaron Van Den Oord, Oriol Vinyals, and Koray Kavukcuoglu. 2017. Neural discrete representation learning. In Advances in Neural Information Processing Systems ( NIPS )

  21. [29]

    Gomez, ukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, pages 5998--6008

  22. [30]

    Sijie Yan, Yuanjun Xiong, and Dahua Lin. 2018. Spatial temporal graph convolutional networks for skeleton-based action recognition. In Proceedings of the AAAI Conference on Artificial Intelligence

  23. [31]

    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...

  24. [32]

    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 gl...

Pith tools

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