Generation of Indian Sign Language Letters, Numbers, and Words

Ajeet Kumar Yadav; Nishant Kumar; Rathna G N

REVIEW 4 major objections 4 minor 27 references

Generation of Indian Sign Language Letters, Numbers, and Words

T0 review · 4 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 A hybrid GAN that adds self-attention layers to progressive growing generates sharper Indian Sign Language images than ProGAN, with a new 247,500-image dataset behind the result.

desk verdict A sensible GAN combination and a potentially valuable dataset, but the semantic claim rests on unverified labels and the headline numbers don't match the table. read the letter →

arxiv 2508.09522 v1 pith:KD27Z5WU submitted 2025-08-13 cs.CV cs.AIcs.LG

Ajeet Kumar Yadav , Nishant Kumar , Rathna G N This is my paper

classification cs.CVcs.AIcs.LG

keywords IndianSignLanguageGANimagegenerationself-attentionprogressivegrowingWGAN-GPdatasettext-to-signclass-conditionalsynthesis

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 aims to make sign language generation practical for Indian Sign Language (ISL) by building a generative model that produces sharp, class-conditional images of letters, numbers, and words. Its proposed architecture grafts self-attention layers onto a progressively growing GAN (ProGAN) and trains with the WGAN-GP loss, with the claim that this hybrid yields finer finger articulation and clearer structure. On a dataset the authors introduce (247,500 images across 165 classes), the model improves Inception Score by 3.2 and Fréchet Inception Distance by 30.12 relative to ProGAN; on an existing 42,000-image ISL alphabet dataset the gains are up to 2.47 and 32.12. The paper also demonstrates a text-to-sign pipeline that splits sentences into words or letters and generates corresponding images. If the numbers hold, this would give hearing and hard-of-hearing users a workable image-generation route toward communication tools and, eventually, sign video generation.

What carries the argument

The load-bearing mechanism is the hybrid generator–critic architecture: progressive resolution growth from 8x8 to 512x512, with two self-attention layers inserted between convolution layers at 64x64 and 128x128; class labels are embedded and concatenated with latent noise in the generator, and similarly embedded and concatenated in the critic. Weight standardization, pixel-wise feature-vector normalization, minibatch standard deviation, and the Wasserstein gradient-penalty loss keep training stable. Attention lets the network preserve long-range dependencies and fine hand-gesture details while progressive growth keeps high resolution tractable.

What would settle it

Have a certified ISL interpreter label a random sample of 500 images from the released dataset: a high error rate would show the generated images do not carry the intended meanings. Also, retrain the hybrid and a vanilla ProGAN on an independently verified ISL dataset; if the reported IS/FID gains of 3.2 and 30.12 do not reappear, the improvement is an artifact of the self-recorded labels rather than of the attention architecture.

Watch

Extended reading notes

Core claim

The central claim is that a generative adversarial network combining ProGAN's progressive growing with self-attention layers at 64x64 and 128x128 resolution stages, class-label conditioning, and the WGAN-GP loss produces higher-quality Indian Sign Language images than the standard ProGAN. The authors report quantitative gains in Inception Score and FID on both their newly collected dataset and an existing 42,000-image ISL alphabet dataset, plus qualitative gains in finger definition and spatial structure. As part of the same contribution, they release a large dataset of high-resolution ISL images covering alphabets, numbers, and 129 words, and show a sentence-to-image generation scheme with

Load-bearing premise

The self-recorded 247,500-image dataset is assumed to correctly label the intended Indian Sign Language signs, yet no certified interpreter verified the labels, and the paper itself notes India has only around 300 certified interpreters.

Editorial extensions

If this is right

If the reported gains are correct, placing self-attention at medium resolutions is a reliable way to improve structured hand-gesture generation, not just for ISL but for other fine-grained sign languages.
The released 247,500-image dataset can serve as a benchmark for ISL recognition, generation, and translation, filling a gap in publicly available high-resolution ISL data.
The word/letter decomposition scheme means any sentence can be rendered as a sequence of sign images, even when the sentence contains out-of-vocabulary words.
Image-level generation combined with the paper's framing around frame interpolation suggests a direct path to sign-language video generation using existing video diffusion or interpolation models.
Using InceptionV3 to convert generated images back to text and scoring with BLEU gives an end-to-end measure of communicative utility, not just pixel fidelity.

Reading between the lines

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

The strongest unstated dependency is dataset label fidelity: the images were recorded by the authors following a standardized ISL dataset, but no certified interpreter is reported to have verified the 165 classes, and the paper notes India has only about 300 certified interpreters. If systematic mislabeling exists, the reported IS/FID improvements would not transfer to real communicative use.
A cheap test of the architecture's generality would be to train the same hybrid on a small, interpreter-verified ISL or American Sign Language dataset; a similar improvement over ProGAN would support the claim that the gains come from attention rather than from dataset idiosyncrasies.
The BLEU-based evaluation inherits InceptionV3's recognition limits, so a stronger falsifier would be a dedicated sign-language recognizer or human interpreter rating of the generated images, which the paper does not include.
The resolution-attention ablation (64x64 alone: FID 29.7; 128x128 alone: 28.8; both: 23.33) hints that adding further attention layers at even higher resolutions might yield additional gains, but at growing compute cost; the paper leaves this extrapolation untested.

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.

Desk Editor's Note

read the letter

The one thing worth remembering: this is not a new method, it's an engineering combination of ProGAN, self-attention, and WGAN-GP with two attention layers inserted at 64x64 and 128x128. That's a perfectly reasonable thing to try, and on the public Gunji dataset the FID improvements are consistently positive across resolutions. The dataset of 247,500 images across 165 classes could be a real resource for low-resource ISL work, provided it actually becomes public and the labels are trustworthy.

The soft spots are not minor. First, the abstract says IS and FID improved by 3.2 and 30.12, but Table III only reports Gunji numbers, where the max IS gain is 1.69, not 2.47 as claimed elsewhere. The own-dataset results are never shown in a table, so the headline numbers can't be checked. Second, the BLEU evaluation uses an ImageNet-pretrained InceptionV3 without any described fine-tuning on sign language, so the BLEU scores are more a measure of the classifier than the generator. Third, no error bars, no code, no dataset link, despite the claim that the dataset is 'available publicly here'.

But the load-bearing issue is the dataset labels. The images come from self-recorded videos, and while they follow ISL-CSLTR, there is no verification by certified interpreters. ISL has regional variation, and the authors themselves note India has only about 300 certified interpreters. Without label validation, the IS and FID numbers only say something about image fidelity and class separability as judged by an Inception network—they do not establish that the model generates the intended signs. That is a serious gap, because the paper's title and main contribution are about generating ISL letters, numbers, and words.

That said, the paper is not incoherent. The architecture is clearly described, the equations for weight standardization, pixel normalization, and the progressive growing fade-in are correct, and the comparison to ProGAN on Gunji is a legitimate baseline. The authors are honest about the limited prior work in ISL generation and cite the relevant SLP literature. The central problem is verification, not conceptual confusion.

Who is this for? Researchers working on sign language generation or low-resource class-conditional GANs could get value from the dataset and the attention-placement ablation, if the release actually happens. It deserves a serious referee, but the referee should insist on label validation, full results on the own dataset, and a corrected abstract. I'd send it to peer review with the expectation of major revision.

Referee Report

4 major / 4 minor

Summary. The paper proposes a class-conditional GAN for generating Indian Sign Language images by combining ProGAN's progressive growing with SAGAN-style self-attention and WGAN-GP loss. It reports IS and FID gains over ProGAN on the Gunji et al. dataset, presents BLEU scores for a sentence-to-image-to-text pipeline, and introduces a new self-recorded dataset of 247,500 images covering 26 letters, 10 digits, and 129 words. The claimed contributions are a modified architecture, a new dataset, and improved quantitative performance.

Significance. If the claims are substantiated, the architectural combination is a reasonable and potentially useful step for ISL image generation, and the dataset could support future research. The positive FID/IS trends are encouraging, and the use of standard external metrics avoids circularity in comparing generative quality. However, the semantic claims of generating correct ISL signs and the quantitative headline improvements are not currently supported by the evidence as presented: label correctness is unverified, the BLEU pipeline relies on an inappropriate recognizer, and the reported numbers are internally inconsistent. These issues are fixable but are central to the paper's main claims.

major comments (4)

[§III.A and Fig. 2] The central semantic claim—that the model generates intended ISL letters, numbers, and words—depends entirely on the correctness of the manually collected dataset labels. The text says images follow ISL-CSLTR standardized images, but no verification by certified interpreters or any independent validation is reported. The paper itself notes India has only ~300 certified interpreters, and ISL has regional variation. Since FID and IS are class-agnostic, they do not validate label correctness. The authors must provide a label-validation protocol (e.g., interpreter checks, inter-annotator agreement) or explicitly restrict the claims to 'images that resemble the authors' self-labeled signs.'
[§IV, Table I and Fig. 4] The BLEU evaluation uses a 'well-trained InceptionV3 classifier' [12] to convert generated images into text. InceptionV3 is trained on ImageNet's 1000 classes, which do not include ISL signs, and no fine-tuning on sign-language data or a custom classification head is described. Therefore the reported BLEU scores of 33.69 (BLEU-1) and 5.04 (BLEU-4) cannot be interpreted as measuring sign-language content. The authors must describe a recognizer that is actually capable of recognizing the generated signs, including its training data and accuracy, or remove the BLEU-based claim.
[Abstract and §IV, Table III] The abstract states IS and FID improvements of 3.2 and 30.12, respectively; §IV states a maximum improvement of 2.47 in IS and 32.12 in FID on the Gunji dataset. However, Table III shows the maximum IS gain over ProGAN is 1.69 (at 256×256) and the maximum FID gain is 32.12 (at 32×32). No table presents an IS gain of 3.2 or 2.47. The resolution and dataset for each claimed improvement must be specified, and the abstract, §IV, and Table III must be reconciled.
[Table II vs Table III] Table II reports FID scores for attention at 64×64 (29.701), 128×128 (28.803), and both (23.331), but it does not state the dataset or resolution at which these scores are measured. Table III lists the authors' final model FID at 64×64 as 30.5914, which differs from the 64×64 attention value in Table II, and no 23.331 value appears in Table III. The comparison across attention placements is therefore difficult to interpret. Clarify the experimental setup for Table II and explain how its best value relates to Table III.

minor comments (4)

[§IV] Typo: 'Seff-attention' should be 'Self-attention'.
[§III.B] A stray 'I' appears at the end of the architecture subsection, likely a formatting artifact.
[§III.A] The dataset link is only referred to as 'available publicly here' without a URL in the text. Provide a working link.
[Fig. 2 caption] The caption says the images are shown 'without class labels'; this means the reader cannot visually verify the claimed sign-to-label mapping, which is related to the label-validation concern.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: IS/FID/BLEU evaluations are external benchmarks and the architecture combines cited components without a self-referential derivation chain.

full rationale

The central claims are empirical: a modified GAN (ProGAN plus self-attention and WGAN-GP) is trained on sign-language image datasets, and its outputs are scored with Inception Score and FID, both computed from external Inception embeddings. These metrics are not defined in terms of the fitted generator or discriminator parameters, so the reported improvements over ProGAN are an experimental outcome rather than a consequence of construction. The BLEU evaluation uses a pre-trained InceptionV3 classifier to convert generated images into text; the paper does not describe fitting that classifier to the evaluation set in a way that would preordain the BLEU scores, so no fitted-input-called-prediction pattern is exhibited. The dataset is self-recorded and label accuracy is not independently verified by certified interpreters, but that is a data-validity and correctness concern, not a circularity of the derivation chain. The architecture is assembled from published components (ProGAN, SAGAN attention, WGAN-GP, weight standardization) with no load-bearing self-citations, no imported uniqueness theorem, and no renaming of a known result as a new derivation. Therefore the paper's derivation chain is self-contained with respect to circularity, and the appropriate score is 0.

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

The central claims rest on standard GAN training and evaluation assumptions. The most fragile are the suitability of ImageNet-trained InceptionV3 features for sign language imagery, the correctness of labels in the self-recorded dataset, and the comparability of the ProGAN baseline.

assumptions (4)

domain assumption InceptionV3 pre-trained on ImageNet provides meaningful features for computing IS and FID on sign language hand images.
Used in Section IV for IS and FID; no validation that these features correlate with sign quality.
domain assumption The labels in the author-recorded dataset correspond to standard Indian Sign Language signs as per ISL-CSLTR.
Section III.A states images follow ISL-CSLTR standardized images; no verification by certified interpreters is reported.
ad hoc to paper The InceptionV3 classifier used for BLEU can recognize generated sign language images without task-specific fine-tuning.
Section IV claims conversion of generated images to text via pre-trained InceptionV3; no training or fine-tuning on ISL data is described.
domain assumption The ProGAN baseline is re-implemented under training conditions comparable to the proposed model (same budget, hyperparameters, and schedule), so differences in IS/FID are attributable to the architecture.
Section IV states only 'trained on [20] dataset till the same epochs'; no code or configuration details for the baseline are given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Generation of Indian Sign Language Letters, Numbers, and Words." pith.science (2026). https://pith.science/paper/KD27Z5WU

@misc{pith2026250809522,
  author       = {Pith},
  title        = {Pith review of: Generation of Indian Sign Language Letters, Numbers, and Words},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KD27Z5WU}},
  note         = {Machine review of arXiv:2508.09522}
}

read the original abstract

Sign language, which contains hand movements, facial expressions and bodily gestures, is a significant medium for communicating with hard-of-hearing people. A well-trained sign language community communicates easily, but those who don't know sign language face significant challenges. Recognition and generation are basic communication methods between hearing and hard-of-hearing individuals. Despite progress in recognition, sign language generation still needs to be explored. The Progressive Growing of Generative Adversarial Network (ProGAN) excels at producing high-quality images, while the Self-Attention Generative Adversarial Network (SAGAN) generates feature-rich images at medium resolutions. Balancing resolution and detail is crucial for sign language image generation. We are developing a Generative Adversarial Network (GAN) variant that combines both models to generate feature-rich, high-resolution, and class-conditional sign language images. Our modified Attention-based model generates high-quality images of Indian Sign Language letters, numbers, and words, outperforming the traditional ProGAN in Inception Score (IS) and Fr\'echet Inception Distance (FID), with improvements of 3.2 and 30.12, respectively. Additionally, we are publishing a large dataset incorporating high-quality images of Indian Sign Language alphabets, numbers, and 129 words.

Figures

Figures reproduced from arXiv: 2508.09522 by the authors.

**Figure 2.** This figure exhibits some samples from our dataset. The 2x5 grid [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Some random samples from the dataset provided by [20]. All images [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: This figure shows the generated output for the sentence, ”Welcome to class where you serve,” by our model trained on our dataset. A given sentence [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: A qualitative comparison of our model with ProGAN output. Both [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 20 canonical work pages

[12]

Szegedy, C., Vanhoucke, V ., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceed- ings of the IEEE conference on computer vision and pattern recognition (pp. 2818-2826)

work page 2016
[1]

”Slowfast Network for Continuous Sign Language Recognition.” ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Ahn, Junseok, Youngjoon Jang, and Joon Son Chung. ”Slowfast Network for Continuous Sign Language Recognition.” ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024

work page 2024
[2]

Progressive transformers for end- to-end sign language production

Saunders B, Camgoz NC, Bowden R. Progressive transformers for end- to-end sign language production. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16 2020 (pp. 687-705). Springer International Publishing

work page 2020
[3]

Selvaraj, P., Nc, G., Kumar, P., & Khapra, M. (2021). Openhands: Making sign language recognition accessible with pose-based pre- trained models across languages. arXiv preprint arXiv:2110.05877

arXiv 2021
[4]

Singh, A., Wadhawan, A., Rakhra, M., Mittal, U., Al Ahdal, A., & Jha, S. K. (2022, October). Indian Sign Language Recognition system for dynamic signs. In 2022 10th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO) (pp. 1-6). IEEE

work page 2022
[5]

Abdullahi, T., Singh, R., & Eickhoff, C. (2024). Retrieval Augmented Zero-Shot Text Classification. arXiv preprint arXiv:2406.15241

work page Pith review arXiv 2024
[6]

(2013, October)

Kalchbrenner, N., & Blunsom, P. (2013, October). Recurrent continuous translation models. In Proceedings of the 2013 conference on empirical methods in natural language processing (pp. 1700-1709)

work page 2013
[7]

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30

2017

Show all 27 references

[8]

Karras, T., Aila, T., Laine, S., & Lehtinen, J. (2017). Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196

2017 arXiv
[9]

Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V ., & Courville, A. C. (2017). Improved training of Wasserstein gans. Advances in neural information processing systems, 30

2017
[10]

Zuo, R., & Mak, B. (2022). C2slr: Consistency-enhanced continuous sign language recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 5131-5140)

2022
[11]

C., Hadfield, S., Koller, O., Ney, H., & Bowden, R

Camgoz, N. C., Hadfield, S., Koller, O., Ney, H., & Bowden, R. (2018). Neural sign language translation. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7784-7793)

2018
[13]

& Bengio, Y

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y . (2014). Generative adversarial nets. Advances in neural information processing systems, 27

2014
[14]

”Conditional generative adversarial nets.” arXiv preprint arXiv:1411.1784 (2014)

Mirza, Mehdi, and Simon Osindero. ”Conditional generative adversarial nets.” arXiv preprint arXiv:1411.1784 (2014)

2014 arXiv
[15]

(2019, May)

Zhang, H., Goodfellow, I., Metaxas, D., & Odena, A. (2019, May). Self- attention generative adversarial networks. In International conference on machine learning (pp. 7354-7363). PMLR

2019
[16]

T., Schicktanz, F., Stankowski, A., & Avramidis, E

Nguyen, L. T., Schicktanz, F., Stankowski, A., & Avramidis, E. (2021, August). Automatic generation of a 3D sign language avatar on AR glasses given 2D videos of human signers. In Proceedings of the 1st international workshop on automatic translation for signed and spoken lang...

2021
[17]

& Salimans, T

Ho, J., Chan, W., Saharia, C., Whang, J., Gao, R., Gritsenko, A., ... & Salimans, T. (2022). Imagen video: High-definition video generation with diffusion models. arXiv preprint arXiv:2210.02303

2022 arXiv
[18]

C., Hadfield, S., & Bowden, R

Stoll, S., Camg ¨oz, N. C., Hadfield, S., & Bowden, R. (2018, September). Sign language production using neural machine translation and gener- ative adversarial networks. In Proceedings of the 29th British Machine Vision Conference (BMVC 2018). British Machine Vision Association

2018
[19]

Jain, S., Watson, D., Tabellion, E., Hoły ´nski, A., Poole, B., & Kontka- nen, J. (2024). Video interpolation with diffusion models. arXiv preprint arXiv:2404.01203

2024 arXiv
[20]

M., Bhargav, N

Gunji, B. M., Bhargav, N. M., Dey, A., Mohammed, I. K. Z., & Sathyajith, S. (2021). Recognition of sign language based on hand gestures. Journal of Advances in Applied & Computational Mathematics, 8, 21-32

2021
[21]

Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V ., Radford, A., & Chen, X. (2016). Improved techniques for training gans. Advances in neural information processing systems, 29

2016
[22]

Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., & Hochreiter, S. (2017). Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30

2017
[23]

& Dolan, B

Faltings, F., Galley, M., Peng, B., Brantley, K., Cai, W., Zhang, Y ., ... & Dolan, B. (2023). Interactive text generation. arXiv preprint arXiv:2303.00908

2023 arXiv
[24]

Papineni, K., Roukos, S., Ward, T., & Zhu, W. J. (2002, July). Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics (pp. 311-318)

2002
[25]

R, Elakkiya; B, NATARAJAN (2021), ”ISL-CSLTR: Indian Sign Lan- guage Dataset for Continuous Sign Language Translation and Recogni- tion”, Mendeley Data, V1, doi: 10.17632/kcmpdxky7p.1

2021 doi
[26]

Qiao, S., Wang, H., Liu, C., Shen, W., & Yuille, A. (2019). Micro-batch training with batch-channel normalization and weight standardization. arXiv preprint arXiv:1903.10520

2019 arXiv
[27]

(2017, July)

Odena, A., Olah, C., & Shlens, J. (2017, July). Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning (pp. 2642-2651). PMLR

2017

Pith tools

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

[1] [12]

Szegedy, C., Vanhoucke, V ., Ioffe, S., Shlens, J., & Wojna, Z. (2016). Rethinking the inception architecture for computer vision. In Proceed- ings of the IEEE conference on computer vision and pattern recognition (pp. 2818-2826)

work page 2016

[2] [1]

”Slowfast Network for Continuous Sign Language Recognition.” ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

Ahn, Junseok, Youngjoon Jang, and Joon Son Chung. ”Slowfast Network for Continuous Sign Language Recognition.” ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024

work page 2024

[3] [2]

Progressive transformers for end- to-end sign language production

Saunders B, Camgoz NC, Bowden R. Progressive transformers for end- to-end sign language production. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XI 16 2020 (pp. 687-705). Springer International Publishing

work page 2020

[4] [3]

Selvaraj, P., Nc, G., Kumar, P., & Khapra, M. (2021). Openhands: Making sign language recognition accessible with pose-based pre- trained models across languages. arXiv preprint arXiv:2110.05877

arXiv 2021

[5] [4]

Singh, A., Wadhawan, A., Rakhra, M., Mittal, U., Al Ahdal, A., & Jha, S. K. (2022, October). Indian Sign Language Recognition system for dynamic signs. In 2022 10th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions)(ICRITO) (pp. 1-6). IEEE

work page 2022

[6] [5]

Abdullahi, T., Singh, R., & Eickhoff, C. (2024). Retrieval Augmented Zero-Shot Text Classification. arXiv preprint arXiv:2406.15241

work page Pith review arXiv 2024

[7] [6]

(2013, October)

Kalchbrenner, N., & Blunsom, P. (2013, October). Recurrent continuous translation models. In Proceedings of the 2013 conference on empirical methods in natural language processing (pp. 1700-1709)

work page 2013

[8] [7]

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30

2017