Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Enhanced Sign Language Translation between American Sign Language (ASL) and Indian Sign Language (ISL) Using LLMs

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read A proposed three-stage pipeline—ASL recognition, LLM text correction, ISL alphabet mapping with frame interpolation—would translate ASL gestures into ISL gesture videos.

desk verdict The pipeline is a reasonable assembly of off-the-shelf parts, but the claimed ASL-to-ISL translation reduces to fingerspelled English and is never evaluated end to end. read the letter →

arxiv 2411.12685 v1 pith:QXULEYMI submitted 2024-11-19 cs.CL cs.AI

Malay Kumar , S. Sarvajit Visagan , Tanish Sarang Mahajan , Anisha Natarajan This is my paper · ORCID
classification cs.CLcs.AI
keywords signlanguagetranslationAmericanrecognitionIndianlargemodelsrandomforestclassifierconvolutionalneuralnetworktextcorrectionframeinterpolation
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

The paper proposes an end-to-end pipeline that translates American Sign Language (ASL) gestures into Indian Sign Language (ISL) gesture videos. The central claim is a three-stage chain: an ensemble recognizer maps ASL handshapes to text, a fine-tuned large language model corrects that text, and a mapping algorithm renders the corrected text as ISL letter gestures, with frame interpolation producing smooth video. The authors argue that the intermediate text representation lets the system preserve sentence intent and handle linguistic differences between ASL and ISL, and they report component-level results for recognition, text correction, and video synthesis. A sympathetic reading accepts the pipeline as a working demonstration of each stage in isolation, connected by the proposed architecture.

What carries the argument

The central device is the intermediate English text: recognition produces English, the LLM repairs it, and a predefined algorithm maps corrected text to ISL gesture frames. The paper treats ISL as a fingerspelling alphabet of 26 letters, so every English sentence becomes a sequence of ISL letter signs. Supporting machinery includes 42 hand-landmark features and silhouette images for classification, a fine-tuned LLM trained on 500 correction pairs, and frame duplication plus intermediate-flow interpolation to reach 60 FPS.

What would settle it

Take a complete ISL sentence such as 'The ball is on the table' and run it through the pipeline, then show the output video to a fluent ISL signer. If the signer reports the output is unrecognizable fingerspelling rather than natural ISL signs, the central translation claim fails. A second check: count the distinct signs in an ISL corpus; if ISL has thousands of whole-word signs beyond the 26 letters, the predefined alphabet mapping cannot express them.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that a hybrid Random Forest Classifier and convolutional network can recognize the 26 static ASL alphabet letters plus control classes, that a fine-tuned large language model can correct recognizer output into grammatical English, and that corrected English can be mapped to the 26-letter ISL alphabet with intermediate-frame interpolation yielding 60 FPS video. The asserted discovery is that chaining these modules creates a real-time ASL-to-ISL translation experience that automatically copes with gesture variability and linguistic differences.

Load-bearing premise

The framework assumes that Indian Sign Language can be produced by spelling out corrected English text letter by letter with the 26-letter ISL alphabet, but real ISL has its own vocabulary and grammar and is not an English cipher.

Editorial extensions

If this is right

  • An ASL user could gesture and, through the system, an ISL user could watch a corresponding gesture video without a human interpreter.
  • The fine-tuned LLM would absorb recognition errors such as 'HELOLO WRLD' and emit corrected English text before synthesis.
  • Frame interpolation would turn a sparse 1 FPS sign sequence into a 60 FPS video, making the output easier to follow.
  • The same recognition-correction-synthesis architecture could be retrained on other sign language datasets to support more language pairs.

Reading between the lines

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

  • The actual output is fingerspelled English in ISL letter shapes, so the translation step reduces to alphabet-to-alphabet mapping rather than rendering ISL vocabulary.
  • The LLM as deployed does no sign-language translation; it corrects English text, so any claim about 'converting ASL text to ISL' is carried entirely by the predefined alphabet mapping.
  • A testable extension would be to replace the letter mapping with a real ISL lexicon and grammar model, then measure whether the LLM correction layer still helps translation quality.
  • The intermediate-text design could generalize to a gloss pipeline, where recognition outputs signed glosses and the LLM orders them into ISL grammatical structure.
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. The paper proposes a three-stage ASL-to-ISL translation pipeline: an RFC+CNN ensemble recognizes static American Sign Language alphabet gestures from images and converts them to text; a fine-tuned LLM (Gemini-1.5 Flash) corrects the recognized text; and a synthesis stage maps the corrected text to frames of the 26-letter Indian Sign Language alphabet, with RIFE-Net used to interpolate smoother video. The authors report component-level results: an RFC confusion matrix on 20 test images per class, a CNN accuracy of 82.4% on 100 test images per class, an LLM correction accuracy of 94.2% on a 100-sample validation set, and one illustrative interpolated video example. The central claim is that this framework provides real-time end-to-end translation between ASL and ISL while overcoming the linguistic differences between the two languages.

Significance. A working end-to-end ASL-to-ISL translation system would be societally valuable and technically interesting, and the paper does attempt to combine several components that are individually relevant to sign-language technology. The authors also present some concrete engineering details, such as GridSearch-tuned random forest hyperparameters, an ensemble voting scheme, and a comparison of three LLMs for text correction. However, the significance of the contribution is not established by the current manuscript. The translation stage reduces ISL to orthographic fingerspelling of corrected English text, which is not a linguistically valid ISL translation, and no end-to-end evaluation of translation quality, intelligibility, or usability is reported. The component evaluations use small private datasets without external benchmarks, and no code or data is released, so the reported accuracies cannot be independently verified. The claimed central contribution therefore remains undemonstrated.

major comments (3)
  1. [Section III-C, Section IV step 3, and Section V-D] The load-bearing assumption of the paper is that a correct ISL utterance can be produced by mapping corrected English text, letter by letter, to frames of the 26-letter ISL alphabet using a predefined algorithm. This is stated in Section III-C, where the ISL dataset comprises visual manifestations of the 26 letters of the ISL alphabet, and in Section IV step 3, where corrected text is mapped to ISL gesture frames. Such a mapping produces fingerspelled English, not grammatical Indian Sign Language, because ISL has its own lexicon, syntax, use of space, and non-manual markers, and most content signs are not fingerspelled. The paper's own Future Scope, Sections VII-B and VII-E, concedes that the system does not handle dynamic gestures, emotional context, or idioms, all of which are core to sign-language communication. The central claim of overcoming linguistic differences between ASL and ISL is therefore unsupported, regardless of the accuracy of the recognition, correction, or interpolation components.
  2. [Section V and Section VI] No end-to-end evaluation is reported. The results section evaluates the RFC on 20 test records per class, the CNN on 100 test records per class, and the LLM text correction on a 100-sample validation set, but there is no measurement of whether the final ISL video output is intelligible, grammatical, or acceptable to ISL users. There is also no comparison against any external sign-language translation benchmark, no human evaluation, and no ablation to show that the proposed pipeline outperforms simpler baselines. The conclusion's claim of an integrated system that preserves context and intent and produces culturally sensitive translation is not supported by the experimental evidence in the manuscript.
  3. [Section V-C and Table V] The reported LLM correction accuracy of 94.2% is assessed on a 100-validation-sample set that appears to be drawn from the same distribution as the 500-example fine-tuning set used to train the model. This creates a risk that the accuracy reflects memorization rather than generalization, and no comparison is reported against the unfine-tuned model or against a held-out set derived from actual ASL recognizer output, which is the real use case. The claim that Gemini-1.5 Flash is the optimal choice would need to be supported by evaluation on a more realistic out-of-distribution benchmark.
minor comments (4)
  1. [Abstract and Section I] The abstract and parts of the introduction are written in an informal, non-standard style, including phrases such as a totally cool innovation, which is inappropriate for a research paper and should be revised throughout.
  2. [Figure 2 caption] The caption contains a typo, Silhoutte instead of Silhouette, and the figure itself should be polished for clarity.
  3. [References] The reference list has inconsistencies, including duplicate entries for the same work, such as references [32] and [38], and several listed references are not cited in the body of the paper, which should be corrected.
  4. [Section V] The recognition experiments lack detail about dataset splits, class-wise performance, and the optimization of the ensemble weights w_RFC and w_CNN; providing this information would be necessary for reproducibility.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'translation to ISL' step reduces by construction: corrected English text is fingerspelled through a 26-letter ISL alphabet, so the ISL output is the input re-encoded, not a linguistic translation.

  1. self definitional [Section III-C 'Video Synthesis' and Section IV, step 3]
    "'Corrected text is mapped to ISL gesture frames using a predefined algorithm to generate ISL videos.' 'The Indian Sign Language gesture dataset comprises visual manifestations of the 26 letters of the ISL alphabet in sequential order from 0 to 25.'"

    The final stage's output is defined as a per-character lookup: each letter of the corrected English text indexes one of the 26 ISL alphabet images. Therefore the produced 'ISL' video is the same English string re-encoded as fingerspelling; no ISL lexical, syntactic, or word-order information is introduced. The paper's claimed 'overcoming linguistic differences between ASL and ISL' thus reduces by construction to character transcription: the input text determines the output frames directly, so the 'translation' is equivalent to its own input under the predefined mapping.

full rationale

The recognition and LLM text-correction stages are empirical and not circular in isolation: the RFC/CNN are trained on gesture images and the LLM on typo-correction pairs. However, the paper's central claim is end-to-end ASL-to-ISL translation, and that claim collapses at the final stage: Section III-C defines the ISL dataset as the 26 letters of the ISL alphabet, and Section IV says corrected text is mapped to ISL gesture frames by a predefined algorithm. This makes the ISL output letter-for-letter fingerspelling of the corrected English text, so the 'translation' is the input by construction. The reported component accuracies (e.g., 94.2% LLM correction on a 100-record validation set) are measured without any external ISL benchmark and do not test linguistic adequacy; they are peripheral to the circular final step. No claims of uniqueness or author-imported theorems are used, and the only potentially self-referential citation ([41]) is not load-bearing. Score 6 because the central translation-to-ISL result reduces by construction, while the upstream recognition components retain independent empirical content.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The framework rests on a chain of hand-fitted components and private datasets. There are no external benchmarks and no released artifacts; the only 'invented entity' is the implicit claim that finger-spelling English with the ISL alphabet is an acceptable translation, which is a conceptual assumption rather than a physical entity.

free parameters (4)
  • RFC hyperparameters (n_estimators=200, max_depth=20, min_samples_split=5, min_samples_leaf=2, bootstrap=True) = n_estimators=200, max_depth=20, min_samples_split=5, min_samples_leaf=2, bootstrap=True
    Selected by GridSearch on the authors' own 2,800-image ASL alphabet dataset; no external validation.
  • Ensemble voting weights (w_RFC, w_CNN) = Not reported numerically
    Section III-A says the weights are 'optimized weights determined through validation performance', i.e., fit to the authors' data.
  • CNN architecture and training hyperparameters = 16/32/64 filters, 128 dense units, dropout 0.2, lr=1e-3, batch 32, 100 epochs
    Chosen by inspection and early stopping on the authors' silhouette dataset; no external benchmark.
  • LLM fine-tuning examples and validation set = 500 training pairs, 100 validation records
    Dataset is self-created; the reported 94.2% accuracy is measured on this private 100-record set.
assumptions (3)
  • domain assumption The 42 MediaPipe hand landmarks and static silhouettes are sufficient to encode ASL alphabet gestures.
    Section III-A builds recognition on 42 landmarks and 28/27 classes of static images; no dynamic gestures are considered, yet the framework claims to handle gesture variability.
  • ad hoc to paper Corrected English text can be mapped to valid ISL by a predefined letter-by-letter algorithm.
    Section III-C maps corrected text to the 26 ISL alphabet images; this ignores ISL grammar, lexical signs, and non-manual features, and is the load-bearing linguistic assumption.
  • domain assumption The small self-collected datasets are representative of real-world ASL and ISL use.
    All recognition and correction accuracies are measured on private datasets of 2,800 to 16,200 images and 500 to 100 text pairs, with no comparison to standard benchmarks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhanced Sign Language Translation between American Sign Language (ASL) and Indian Sign Language (ISL) Using LLMs." pith.science (2026). https://pith.science/paper/QXULEYMI

@misc{pith2026241112685,
  author       = {Pith},
  title        = {Pith review of: Enhanced Sign Language Translation between American Sign Language (ASL) and Indian Sign Language (ISL) Using LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QXULEYMI}},
  note         = {Machine review of arXiv:2411.12685}
}
read the original abstract

We have come up with a research that hopes to provide a bridge between the users of American Sign Language and the users of spoken language and Indian Sign Language (ISL). The research enabled us to create a novel framework that we have developed for Learner Systems. Leveraging art of Large models to create key features including: - Real-time translation between these two sign languages in an efficient manner. Making LLM's capability available for seamless translations to ISL. Here is the full study showing its implementation in this paper. The core of the system is a sophisticated pipeline that begins with reclassification and recognition of ASL gestures based on a strong Random Forest Classifier. By recognizing the ASL, it is translated into text which can be more easily processed. Highly evolved natural language NLP (Natural Language Processing) techniques come in handy as they play a role in our LLM integration where you then use LLMs to be able to convert the ASL text to ISL which provides you with the intent of sentence or phrase. The final step is to synthesize the translated text back into ISL gestures, creating an end-to-end translation experience using RIFE-Net. This framework is tasked with key challenges such as automatically dealing with gesture variability and overcoming the linguistic differences between ASL and ISL. By automating the translation process, we hope to vastly improve accessibility for sign language users. No longer will the communication gap between ASL and ISL create barriers; this totally cool innovation aims to bring our communities closer together. And we believe, with full confidence in our framework, that we're able to apply the same principles across a wide variety of sign language dialects.

Figures

Figures reproduced from arXiv: 2411.12685 by the authors.

Figure 1
Figure 1. American Sign Language dataset sample Feature extraction: With the mediapipe framework we extract 42 prominent landmark points (L) placed on the hand which serve as the features for training the model. L = {(xi , yi , zi) | i ∈ [1, 42]} The coordinates obtained are then processed for normalization using StandardScaler. xnormalized = x − µx σx where µx and σx represent the mean and standard deviation of the feature d… view at source ↗
Figure 2
Figure 2. Silhoutte Image American Sign Language dataset [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. System Architecture V. RESULT AND OUTPUT This section provides the experimental results of various phases of development, which are performed and investigated to build a complete framework. The proposed framework functionalities are tested in different stages of the development cycle. In addition to that, we have shown the user interface screens of the final application. For the first phase sign lan￾guage recognitio… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Confusion Matrix Random Forest Classifier Model [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Accuracy and Loss Evaluation of the CNN Model [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix Convolution Neural Network Model [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 33 canonical work pages

  1. [1]

    P. S. Rao, et al., ”Multiple Languages to Sign Language Using NLTK,” International Journal of Scientific Research in Science and Technology , vol. 10, no. 2, pp. 12–17, Mar.–Apr. 2023. [Online]. Available: https: //doi.org/10.32628/IJSRST2310189

    work page doi:10.32628/ijsrst2310189 2023

  2. [2]

    Al-Qureshi, T

    M. Al-Qureshi, T. Khalid, and R. Souissi, ”Deep Learning for Sign Lan- guage Recognition: Current Techniques, Benchmarks, and Open Issues,” IEEE Access , vol. 9, 2021. [Online]. Available: https://ieeexplore.ieee. org/document/12345678

    arXiv 2021

  3. [3]

    Salian, et al., ”Proposed System for Sign Language Recognition,” in Proc

    S. Salian, et al., ”Proposed System for Sign Language Recognition,” in Proc. Int. Conf. Communication, Power, and Embedded Systems (ICCPEIC), 2017, pp. 058–062. doi: 10.1109/ICCPEIC.2017.8290339

    work page arXiv 2017

  4. [4]

    Mubashira and A

    N. Mubashira and A. James, ”Transformer Network for Video-to-Text Translation,” in Proc. 2020 Int. Conf. Power, Instrumentation, Control and Computing (PICC) , 2020. [Online]. Available: https://doi.org/10. 1109/picc51425.2020.9362374

    work page arXiv 2020

  5. [5]

    Halder and A

    A. Halder and A. Tayade, ”Real-time Vernacular Sign Language Recog- nition Using MediaPipe and Machine Learning,” Int. J. Res. Proj. Res. , vol. 2, no. 3, 2021. [Online]. Available: www.ijrpr.com

    work page 2021

  6. [6]

    Shin, et al., ”Korean Sign Language Recognition Using Transformer- Based Deep Neural Network,” Applied Sciences, vol

    J. Shin, et al., ”Korean Sign Language Recognition Using Transformer- Based Deep Neural Network,” Applied Sciences, vol. 13, no. 5, p. 3029,

    work page

  7. [7]

    Wang, et al., ”Improved 3D-ResNet Sign Language Recognition Algorithm with Enhanced Hand Features,” Scientific Reports , vol

    S. Wang, et al., ”Improved 3D-ResNet Sign Language Recognition Algorithm with Enhanced Hand Features,” Scientific Reports , vol. 12, no. 1, 2022. doi: 10.1038/s41598-022-21636-z

    work page doi:10.1038/s41598-022-21636-z 2022

  8. [8]

    Sahoo, G

    A. Sahoo, G. Mishra, and K. Ravulakollu, ”Sign Language Recognition: State of the Art,” ARPN J. Eng. Appl. Sci. , vol. 9, 2014, pp. 116–134

    work page 2014

Show all 42 references
  1. [9]

    R. K. Pathan, et al., ”Sign Language Recognition Using the Fusion of Image and Hand Landmarks Through Multi-Headed Convolutional Neural Network,” Scientific Reports , vol. 13, 2023, p. 16975. doi: 10. 1038/s41598-023-43852-x

    2023

  2. [10]

    Shenoy, et al., ”Real-time Indian Sign Language (ISL) Recognition,” in Proc

    K. Shenoy, et al., ”Real-time Indian Sign Language (ISL) Recognition,” in Proc. 2018 9th Int. Conf. Comput., Commun. and Netw. Technol. (ICCCNT), Bengaluru, India, 2018, pp. 1–9. doi: 10.1109/ICCCNT. 2018.8493808

    2018

  3. [11]

    Sharma, et al., ”Translating Speech to Indian Sign Language Using Natural Language Processing,” Future Internet, vol

    P. Sharma, et al., ”Translating Speech to Indian Sign Language Using Natural Language Processing,” Future Internet, vol. 14, no. 9, p. 253,

  4. [12]

    P. C. Badhe and V . Kulkarni, ”Indian Sign Language Translator Using Gesture Recognition Algorithm,” in Proc. 2015 IEEE Int. Conf. Comput. Graph., Vis. and Inf. Security (CGVIS) , Bhubaneswar, India, 2015, pp. 195–200. doi: 10.1109/CGVIS.2015.7449921

    2015

  5. [13]

    M. K. Das, et al., ”Real-time Gesture Recognition for American Sign Language Using Convolutional Neural Network,” Int. J. Comput. Appl. , vol. 200, no. 8, pp. 6–10, Jul. 2019. [Online]. Available: https://doi.org/ 10.5120/ijca2019919038

    2019 doi

  6. [14]

    Gupta and R

    A. Gupta and R. Kumar, ”Sign Language Recognition Using Hand Gesture Recognition Techniques: A Review,” Int. J. Comput. Appl. , vol. 195, no. 9, Sep. 2018, pp. 18–23. [Online]. Available: https: //doi.org/10.5120/ijca2018917833

    2018 doi

  7. [15]

    Jain and R

    S. Jain and R. Patel, ”Deep Learning-Based Sign Language Recognition: A Review,” Journal of Computer Science , vol. 17, no. 2, pp. 138–150,

  8. [16]

    Kumar and B

    A. Kumar and B. Sahoo, ”Sign Language Recognition System Using Hand Gestures: A Survey,” Int. J. Comput. Sci. Inf. Technol. , vol. 5, no. 4, pp. 5609–5613, 2014

    2014

  9. [18]

    Patel and P

    K. Patel and P. Verma, ”Sign Language Recognition Using Deep Learning Techniques: A Review,” Int. J. Comput. Sci. Mobile Comput. , vol. 10, no. 2, pp. 52–58, Feb. 2021

    2021

  10. [19]

    Sharma and S

    N. Sharma and S. Chaudhary, ”A Survey on Deep Learning Techniques for Sign Language Recognition,” Int. J. Adv. Res. Comput. Sci. Softw. Eng., vol. 10, no. 2, pp. 452–459, Feb. 2020

    2020

  11. [20]

    Singh and R

    S. Singh and R. Chellappan, ”A Comprehensive Review on Sign Language Recognition: Datasets, Techniques, and Challenges,” Pattern Recognition Letters , vol. 143, Dec. 2020, pp. 88–100. doi: 10.1016/j. patrec.2020.12.004

    2020 doi

  12. [21]

    Singh and A

    V . Singh and A. Sharma, ”A Review on Sign Language Recognition Techniques,” Journal of Computer Science and Applications , vol. 19, no. 1, pp. 22–30, 2021

    2021

  13. [23]

    Aggarwal, et al., ”A Survey on Sign Language Recognition Tech- niques: Challenges and Future Directions,” Journal of Ambient Intelli- gence and Humanized Computing , vol

    Y . Aggarwal, et al., ”A Survey on Sign Language Recognition Tech- niques: Challenges and Future Directions,” Journal of Ambient Intelli- gence and Humanized Computing , vol. 12, no. 2, pp. 2181–2195, 2021. doi: 10.1007/s12652-020-02069-x

    2021 doi

  14. [24]

    Bansal and S

    R. Bansal and S. Bansal, ”A Review on Sign Language Recognition Techniques,” International Journal of Advanced Research in Computer Science, vol. 11, no. 2, pp. 123-128, Feb. 2020

    2020

  15. [25]

    Chauhan and A

    T. Chauhan and A. Kumar, ”Sign Language Recognition: A Review,” Int. J. Comput. Sci. Inf. Security , vol. 16, no. 5, pp. 1–8, May 2018

    2018

  16. [26]

    Chaudhary and S

    V . Chaudhary and S. Chaudhary, ”A Review on Sign Language Recog- nition Techniques and Approaches,” Int. J. Comput. Appl. , vol. 212, no. 13, pp. 40–45, May 2020. doi: https://doi.org/10.5120/ijca2020918781

    2020 doi

  17. [27]

    Goyal and S

    P. Goyal and S. Goyal, ”A Comprehensive Review on Sign Language Recognition Techniques,” Int. J. Adv. Res. Comput. Sci. , vol. 11, no. 2, pp. 123–128, Feb

  18. [28]

    Kumar, et al., ”Sign Language Recognition: A Review,” Int

    S. Kumar, et al., ”Sign Language Recognition: A Review,” Int. J. Comput. Sci. Mobile Comput. , vol. 9, no. 9, pp. 50–57, Sep. 2020

    2020

  19. [29]

    N. E.-d. M. Salem, A. Al-Atabi, and M. Sarhan, ”A Review of Sign Language Translation Systems,” Artificial Intelligence Review , vol. 57, no. 2, pp. 721–758, 2022. doi: 10.1007/s10472-021-09945-w

    2022 doi

  20. [30]

    Ko, et al., ”Sign Language Translation with Multimodal Neural Machine Translation,” in Proc

    B. Ko, et al., ”Sign Language Translation with Multimodal Neural Machine Translation,” in Proc. 2019 Conf. Empirical Methods in Natu- ral Language Processing and 9th Int. Joint Conf. Natural Language Processing (EMNLP-IJCNLP) , Hong Kong, China, Nov. 2019, pp. 5802–5813. [Onli...

    2019 doi

  21. [31]

    Wang, et al., ”Towards Sign Language Translation: A Review,” IEEE Trans

    H. Wang, et al., ”Towards Sign Language Translation: A Review,” IEEE Trans. Syst., Man, Cybern.: Syst. , vol. 52, no. 1, pp. 72–84, 2022. doi: 10.1109/TSMC.2021.3111522

    2022

  22. [32]

    Gong, et al., ”LLMs are good sign language translators,” in Proc

    J. Gong, et al., ”LLMs are good sign language translators,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. , 2024

    2024

  23. [33]

    Chen, et al., ”Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation,” arXiv preprint arXiv:2403.12556, 2024

    Z. Chen, et al., ”Factorized Learning Assisted with Large Language Model for Gloss-free Sign Language Translation,” arXiv preprint arXiv:2403.12556, 2024

    2024 arXiv

  24. [34]

    Fang, et al., ”SignLLM: Sign Languages Production Large Language Models,” arXiv preprint arXiv:2405.10718, 2024

    S. Fang, et al., ”SignLLM: Sign Languages Production Large Language Models,” arXiv preprint arXiv:2405.10718, 2024

    2024 arXiv

  25. [35]

    M. I. Saleem, et al., ”A Machine Learning Based Full Duplex System Supporting Multiple Sign Languages for the Deaf and Mute,” Applied Sciences, vol. 13, no. 13, p. 3114, 2023. doi: 10.3390/app13053114

    2023 doi

  26. [36]

    ”Relevant Intrinsic Feature Enhancement Network for Few-Shot Semantic Segmentation,” arXiv preprint , 2023.[Online]

    Xiaoyi Bao, et al. ”Relevant Intrinsic Feature Enhancement Network for Few-Shot Semantic Segmentation,” arXiv preprint , 2023.[Online]. Available: https://arxiv.org/abs/2312.06474v1

    2023 arXiv

  27. [37]

    ”A survey of large language models.” arXiv preprint, 2023

    Zhao, Wayne Xin, et al. ”A survey of large language models.” arXiv preprint, 2023. [Online]. Available: https://arxiv.org/abs/2303.18223

    2023 arXiv

  28. [38]

    ”LLMs are Good Sign Language Translators,” arXiv preprint 2024

    Jia Gong, et al. ”LLMs are Good Sign Language Translators,” arXiv preprint 2024. [Online]. Available: https://arxiv.org/abs/2404.00925

    2024 arXiv

  29. [39]

    ”Textually Pretrained Speech Lan- guage Models,” arXiv preprint , 2024

    Michael Hassid, et al. ”Textually Pretrained Speech Lan- guage Models,” arXiv preprint , 2024. [Online]. Available: https://arxiv.org/abs/2305.13009

    2024 arXiv

  30. [40]

    Hertz, ”Comparison of LLMs: Evaluating Large Lan- guage Models,” Leeway Hertz , 2024 [Online]

    L. Hertz, ”Comparison of LLMs: Evaluating Large Lan- guage Models,” Leeway Hertz , 2024 [Online]. Available: https://www.leewayhertz.com/comparison-of-llms/#evaluating-large- language-models

    2024

  31. [41]

    Natarajan et al., ”Development of an End-to-End Deep Learning Framework for Sign Language Recognition, Translation, and Video Generation,” in IEEE Access, vol

    B. Natarajan et al., ”Development of an End-to-End Deep Learning Framework for Sign Language Recognition, Translation, and Video Generation,” in IEEE Access, vol. 10, pp. 104358-104374, 2022, doi: 10.1109/ACCESS.2022.3210543. [Online]. Available: https://ieeexplore.ieee.org/do...

    2022

  32. [2021]

    doi: https://doi.org/10.3844/jcssp.2021.138.150

    2021 doi

  33. [2022]

    doi: 10.3390/fi14090253

    doi

  34. [2023]

    doi: 10.3390/app13053029

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.