Pith. sign in

REVIEW 5 major objections 4 minor 1 cited by

Vision-Language Models for Automated Chest X-ray Interpretation: Leveraging ViT and GPT-2

T0 review · 5 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that SWIN-BART, pairing a SWIN Transformer image encoder with a BART text decoder, is the best of four configurations for generating chest X-ray reports on the IU-Xray dataset.

desk verdict A small, confounded parameter scan over off-the-shelf encoder-decoder pairs; the SWIN-BART claim is not supported as stated because BART models got more training and the tables contradict the 'superior across all' phrasing. read the letter →

arxiv 2501.12356 v1 pith:B5RFF6OX submitted 2025-01-21 cs.CV

Md. Rakibul Islam , Md. Zahid Hossain , Mustofa Ahmed , Most. Sharmin Sultana Samu This is my paper · ORCID
classification cs.CV
keywords chestX-rayreportgenerationvision-languagemodelSWINTransformerVisionBARTGPT-2IU-Xraydatasetradiologytextmetrics
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 evaluates four vision-language configurations for automated chest X-ray report generation: SWIN-BART, SWIN-GPT-2, ViT-B16-BART, and ViT-B16-GPT-2. On the IU-Xray dataset, it reports that SWIN-BART achieves the highest scores across ROUGE, BLEU, and BERTScore, making it the recommended encoder-decoder pairing among the four. If the claim holds, it identifies a practical architecture combination for generating radiology report drafts. The motivation is clinical workflow: manual report writing is slow and error-prone, and a reliable automated pairing could reduce that burden.

What carries the argument

The load-bearing object is the encoder-decoder pairing itself. The SWIN Transformer encodes images hierarchically through shifted windows, producing multi-scale visual features; BART then decodes text from a bidirectional encoder and autoregressive decoder. Cross-attention mechanisms feed the visual features into the language model. The comparison is carried by this shared machinery: every configuration uses the same dataset, optimizer (AdamW), learning rate, batch size, and evaluation pipeline, so the reported differences are attributed to the choice of encoder and decoder. The evaluation machinery is ROUGE, BLEU, and BERTScore, which measure n-gram overlap, precision, and contextual embedding similarity respectively.

What would settle it

Retrain all four models with identical epoch counts, learning-rate schedules, and multiple random seeds, and score the reports without removing stopwords; if the SWIN-BART lead disappears or falls within run-to-run variation, the claim that the architecture pairing is best is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that SWIN-BART—a SWIN Transformer image encoder coupled with a BART sequence-to-sequence text decoder—is the best of the four tested configurations for generating chest X-ray reports. On the IU-Xray test set, it reports the highest scores on ROUGE1 F1 (0.4134), ROUGE-L F1 (0.2935), BLEU (0.0648), and BERTScore F1 (0.8899), leading on nearly every metric in Tables 3 and 4. The authors interpret this as evidence that the hierarchical, shifted-window visual features of the SWIN Transformer align more effectively with BART's bidirectional encoding and autoregressive decoding than the other pairings do. This is an extension-style comparison: it takes existing pretrained encoders and decoders, fine-tunes all four cross products, and selects the winner.

Load-bearing premise

The paper's ranking assumes that the performance gap between SWIN-BART and the other three models comes from the choice of encoder and decoder, not from the unequal training budgets (8 epochs for BART models, 5 for GPT-2 models) or from removing stopwords before evaluation.

Editorial extensions

If this is right

  • Among the four configurations, SWIN-BART is the one reported as best on every metric family in Tables 3 and 4.
  • Because all four models share the same optimizer, learning rate, batch size, and evaluation setup, the paper treats the encoder-decoder choice as the variable that explains the score gaps.
  • The reported numbers provide a reproducible point of comparison for future report-generation work on the IU-Xray benchmark.
  • If the result is correct, practitioners building automated reporting pipelines would favor a SWIN encoder over a ViT-B16 encoder when paired with BART.

Reading between the lines

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

  • The unequal epoch budgets (BART models trained for 8 epochs, GPT-2 models for 5) mean the ranking is confounded with training budget; an equal-epoch rerun is needed to credit the architecture.
  • Stopword removal before scoring likely inflates n-gram overlap scores for all models; keeping stopwords might change both the absolute numbers and the ordering.
  • Single-run scores with no variance estimate make the margins on BERTScore (roughly 0.02–0.06) hard to interpret; repeated seeds would show whether the lead is stable.
  • A clinically useful next test would be radiologist evaluation of the generated reports, since ROUGE, BLEU, and BERTScore do not measure diagnostic correctness.
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

5 major / 4 minor

Summary. The paper evaluates four encoder-decoder configurations for automated chest X-ray report generation on the IU-Xray dataset: SWIN Transformer or ViT-B16 as image encoders, and BART or GPT-2 as text decoders. All models are fine-tuned end-to-end with cross-entropy loss and AdamW, then compared via ROUGE-1/2/3/4/L, BLEU, and BERTScore. The central claim is that SWIN-BART is the best-performing combination, with the highest scores on most metrics (e.g., ROUGE-1 F1 0.4134, ROUGE-L F1 0.2935, BLEU 0.0648, BERTScore F1 0.8899 in Tables 3 and 4). The paper also presents preprocessing statistics, training/validation loss curves, and a qualitative example report.

Significance. If the comparison were clean, the paper would provide a useful empirical data point on pairing off-the-shelf vision encoders with autoregressive versus sequence-to-sequence decoders for medical report generation, and the use of a public dataset and multiple automatic metrics is a strength. However, the significance is substantially limited by experimental-design confounds: the four configurations are not trained under equal budgets, results come from a single run with no variance estimates, and the textual claims overstate what the tables show. These issues bear directly on the headline conclusion, so the paper's current contribution is more a preliminary observation than a robust architecture comparison.

major comments (5)
  1. [Table 2 and Section 5] The central comparison is confounded by training budget. Table 2 assigns 8 epochs to both BART configurations and only 5 epochs to both GPT-2 configurations, so every decoder comparison is also an epoch comparison. In particular, the headline result 'SWIN-BART is better than SWIN-GPT-2' could be entirely due to the extra 60% training budget rather than the decoder choice. The one epoch-matched comparison (SWIN-BART vs. ViT-B16-BART) does support the encoder-side conclusion, but the decoder-side conclusion is untested as reported. Please retrain with equal epoch budgets, or train each model to convergence with early stopping and report the actual epoch counts; otherwise the architecture ranking is not valid.
  2. [Tables 3 and 4] The paper reports a single run per configuration and provides no standard deviations, confidence intervals, or significance tests. Several claimed advantages are small enough that they could plausibly reverse under a different random seed: e.g., ROUGE-4 F1 is 0.0427 for SWIN-BART versus 0.0435 for ViT-B16-GPT-2, and BERTScore Recall is 0.8947 for SWIN-BART versus 0.9015 for ViT-B16-GPT-2. At minimum, repeat each configuration with multiple seeds and report mean and standard deviation; ideally also perform pairwise significance testing. Without this, the ranking in Section 5 is fragile.
  3. [Section 5, text accompanying Tables 3 and 4] The text overstates the results. It says SWIN-BART demonstrated 'superior performance across all ROUGE metrics' and 'superior performance across all BERTScore metrics,' but Table 3 shows SWIN-BART's ROUGE-4 F1 (0.0427) is lower than ViT-B16-GPT-2's (0.0435), and Table 4 shows its BERTScore Recall (0.8947) is lower than both ViT-B16-GPT-2 (0.9015) and SWIN-GPT-2 (0.8998). Please revise these claims to accurately describe the tables, e.g., 'best on the majority of metrics,' and discuss the exceptions explicitly.
  4. [Section 4.2] The preprocessing removes stopwords from radiology reports before evaluation. It is not stated whether stopwords are removed from references, generated text, or both, nor whether the reported ROUGE/BLEU/BERTScore numbers therefore measure stopword-stripped text. Removing stopwords changes the metric definitions and makes the absolute scores incomparable with prior work that evaluates on original reports. Please clarify the exact evaluation protocol and, ideally, also report metrics on the original, unmodified reports so readers can relate the numbers to the literature.
  5. [Section 3 and reproducibility] The architectural description is too underspecified to reproduce the central experiments. The paper says 'cross-attention mechanisms were employed to integrate the visual features into the language model,' but does not describe how cross-attention is implemented for either model, especially for decoder-only GPT-2, nor the projection layer, hidden sizes, number of decoder layers with cross-attention, or how the encoder features are fed to the decoder. No code or model configuration files are provided. Please give a precise model specification or release the implementation, since the claimed comparisons otherwise cannot be independently checked.
minor comments (4)
  1. [References] Reference [29] cites Mao et al., 'Towards Robust Vision Transformer,' as the source of the Vision Transformer, but the canonical ViT architecture is from Dosovitskiy et al., 'An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale.' Please cite the correct original work.
  2. [Various] There are several typographical issues, including 'T able 1' and 'T able 2' in the captions, inconsistent capitalization of 'SWIN' (SWIN vs. Swin) and 'BERTScore' (BER TScore in Table 4), and an inconsistent dataset name ('IU-Xray' vs. 'IU-XRay'). Please proofread carefully.
  3. [Figure 11] Figure 11 shows a generated report and ground truth, but the caption does not state which of the four models produced the example. If the intent is to illustrate SWIN-BART, that should be explicit.
  4. [Section 5] The paper claims 'remarkable results' and 'significantly outperforming' at several places, but no statistical significance testing is performed and no external baselines are included. Please temper the language or provide appropriate statistical support.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper reports empirical metrics for fixed model configurations, and selecting the top row of its own evaluation table is ordinary model comparison, not circular reasoning.

full rationale

The paper makes no derivation from first principles and defines no quantity in terms of another predicted quantity. It trains four fixed encoder-decoder configurations (SWIN-BART, SWIN-GPT-2, ViT-B16-BART, ViT-B16-GPT-2) on a public dataset and reports ROUGE, BLEU, and BERTScore metrics. The claim that SWIN-BART is the best performer is an empirical ranking of the measured rows, not a prediction derived from a fitted parameter. There are no fitted inputs that are later renamed as predictions, no uniqueness theorem imported from the authors' prior work, and no self-citation used as load-bearing evidence; references [29]-[35] are standard external works for the component architectures and metrics. The main validity concerns are that the BART models received 8 epochs while the GPT-2 models received only 5, that only one run was performed per configuration with no variance estimates, and that the text overstates the tables in a few places (e.g., SWIN-BART's BERTScore Recall of 0.8947 is below ViT-B16-GPT-2's 0.9015 despite the claim of superior performance across all BERTScore metrics). These are correctness and experimental-design issues, not circularity: the ranking is not forced by definition, and nothing in the paper reduces to its own inputs. Therefore the appropriate circularity score is 0.

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

No new entities are postulated; the paper combines existing pretrained models and reports empirical metrics. The central claim rests on hand-chosen hyperparameters, the quality of the IU-Xray references, and the unstated assumption that text metrics capture clinical value.

free parameters (4)
  • Number of training epochs per model = 5 for GPT-2 models, 8 for BART models
    Chosen by hand; the unequal budgets confound the architecture comparison in Section 4.3.
  • Learning rate = 0.00005
    Fixed for all models without a sweep; affects convergence and therefore the reported ranking.
  • Weight decay = 0.01
    Fixed for all models; a standard regularizer choice that influences final metrics.
  • Batch size = 8
    Fixed for all models; not justified, and interacts with the number of epochs.
assumptions (4)
  • domain assumption IU-Xray ground-truth reports are accurate and complete enough to serve as references
    All metrics compare generated text against these reports; no radiologist verification is performed.
  • domain assumption Pretrained ViT, SWIN, BART, and GPT-2 features transfer to chest X-ray report generation after fine-tuning
    Assumed throughout Section 3; no failure analysis or feature quality evaluation is provided.
  • domain assumption ROUGE, BLEU, and BERTScore correlate with clinical quality
    Used as the sole evaluation; no clinical metrics such as CheXbert or RadGraph, and no human evaluation.
  • ad hoc to paper The implemented cross-attention between encoder and decoder is a correct and sufficient fusion mechanism
    Section 3 describes it in one sentence with no equations, tensor shapes, or ablation supporting this design.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Vision-Language Models for Automated Chest X-ray Interpretation: Leveraging ViT and GPT-2." pith.science (2026). https://pith.science/paper/B5RFF6OX

@misc{pith2026250112356,
  author       = {Pith},
  title        = {Pith review of: Vision-Language Models for Automated Chest X-ray Interpretation: Leveraging ViT and GPT-2},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B5RFF6OX}},
  note         = {Machine review of arXiv:2501.12356}
}
read the original abstract

Radiology plays a pivotal role in modern medicine due to its non-invasive diagnostic capabilities. However, the manual generation of unstructured medical reports is time consuming and prone to errors. It creates a significant bottleneck in clinical workflows. Despite advancements in AI-generated radiology reports, challenges remain in achieving detailed and accurate report generation. In this study we have evaluated different combinations of multimodal models that integrate Computer Vision and Natural Language Processing to generate comprehensive radiology reports. We employed a pretrained Vision Transformer (ViT-B16) and a SWIN Transformer as the image encoders. The BART and GPT-2 models serve as the textual decoders. We used Chest X-ray images and reports from the IU-Xray dataset to evaluate the usability of the SWIN Transformer-BART, SWIN Transformer-GPT-2, ViT-B16-BART and ViT-B16-GPT-2 models for report generation. We aimed at finding the best combination among the models. The SWIN-BART model performs as the best-performing model among the four models achieving remarkable results in almost all the evaluation metrics like ROUGE, BLEU and BERTScore.

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. Privacy-Preserving Chest X-ray Report Generation via Multimodal Federated Learning with ViT and GPT-2

    eess.IV 2025-05 conditional novelty 3.0 of 10

    Krum aggregation produced the highest automatic text metrics for a federated ViT-GPT-2 chest X-ray report generator on IU-Xray, but margins over FedAvg and centralized training are tiny and no privacy mechanism backs ...

Reference graph

Works this paper leans on

35 extracted references · 21 canonical work pages · cited by 1 Pith paper

  1. [21]

    Heliyon 10(7) (2024)

    Magalh˜ aes, G.V., Santos, R.L.d.S., Vogado, L.H., Paiva, A.C., Santos Neto, P.d.A.: Xrayswingen: Automatic medical reporting for x-ray exams with multi- modal model. Heliyon 10(7) (2024)

    2024

  2. [1]

    arXiv preprint arXiv:2204.13258 (2022)

    Chen, Z., Shen, Y., Song, Y., Wan, X.: Cross-modal memory networks for radiology report generation. arXiv preprint arXiv:2204.13258 (2022)

    arXiv 2022

  3. [2]

    arXiv preprint arXiv:1711.08195 (2017)

    Jing, B., Xie, P., Xing, E.: On the automatic generation of medical imaging reports. arXiv preprint arXiv:1711.08195 (2017)

    arXiv 2017

  4. [3]

    In: Proceedings of the AAAI Conference on Artificial Intelligence, vol

    Li, C.Y., Liang, X., Hu, Z., Xing, E.P.: Knowledge-driven encode, retrieve, paraphrase for medical image report generation. In: Proceedings of the AAAI Conference on Artificial Intelligence, vol. 33, pp. 6666–6673 (2019)

    2019

  5. [4]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Tanida, T., M¨ uller, P., Kaissis, G., Rueckert, D.: Interactive and explainable region-guided radiology report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 7433–7442 (2023)

    2023

  6. [5]

    In: Findings of the Association for Computational Linguistics: EMNLP 2020, pp

    Lovelace, J., Mortazavi, B.: Learning to generate clinically coherent chest x-ray reports. In: Findings of the Association for Computational Linguistics: EMNLP 2020, pp. 1235–1243 (2020)

    work page 2020

  7. [6]

    IEEE Access 9, 21236– 21250 (2021)

    Hou, D., Zhao, Z., Liu, Y., Chang, F., Hu, S.: Automatic report generation for chest x-ray images via adversarial reinforcement learning. IEEE Access 9, 21236– 21250 (2021)

    work page 2021

  8. [7]

    Plos one 17(1), 0262209 (2022)

    Sirshar, M., Paracha, M.F.K., Akram, M.U., Alghamdi, N.S., Zaidi, S.Z.Y., Fatima, T.: Attention based automated radiology report generation using cnn and lstm. Plos one 17(1), 0262209 (2022)

    work page 2022

Show all 35 references
  1. [8]

    Applied Sciences 12(22), 11750 (2022)

    Ahmed, S.B., Solis-Oba, R., Ilie, L.: Explainable-ai in automated medical report generation using chest x-ray images. Applied Sciences 12(22), 11750 (2022)

    2022

  2. [9]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp

    Li, J., Li, S., Hu, Y., Tao, H.: A self-guided framework for radiology report generation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 588–598 (2022). Springer

    2022

  3. [10]

    Scientific data 6(1), 317 (2019)

    Johnson, A.E., Pollard, T.J., Berkowitz, S.J., Greenbaum, N.R., Lungren, M.P., Deng, C.-y., Mark, R.G., Horng, S.: Mimic-cxr, a de-identified publicly available database of chest radiographs with free-text reports. Scientific data 6(1), 317 (2019)

    2019

  4. [11]

    Journal of the American Medical Informatics Association 23(2), 304–310 (2016)

    Demner-Fushman, D., Kohli, M.D., Rosenman, M.B., Shooshan, S.E., Rodriguez, L., Antani, S., Thoma, G.R., McDonald, C.J.: Preparing a collection of radiology 14 examinations for distribution and retrieval. Journal of the American Medical Informatics Association 23(2), 304–310 (2016)

    2016

  5. [12]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp

    Wang, Z., Tang, M., Wang, L., Li, X., Zhou, L.: A medical semantic-assisted transformer for radiographic report generation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention, pp. 655–664 (2022). Springer

    2022

  6. [13]

    In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp

    Li, Q.: Harnessing the power of pre-trained vision-language models for effi- cient medical report generation. In: Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, pp. 1308–1317 (2023)

    2023

  7. [14]

    Artificial intelligence in medicine 144, 102633 (2023)

    Nicolson, A., Dowling, J., Koopman, B.: Improving chest x-ray report genera- tion by leveraging warm starting. Artificial intelligence in medicine 144, 102633 (2023)

    2023

  8. [15]

    In: ECAI 2023, pp

    Voutharoja, B.P., Wang, L., Zhou, L.: Automatic radiology report generation by learning with increasingly hard negatives. In: ECAI 2023, pp. 2427–2434. IOS Press, ??? (2023)

    2023

  9. [16]

    In: 2023 3rd Inter- national Conference on Advanced Research in Computing (ICARC), pp

    Nimalsiri, W., Hennayake, M., Rathnayake, K., Ambegoda, T.D., Meedeniya, D.: Automated radiology report generation using transformers. In: 2023 3rd Inter- national Conference on Advanced Research in Computing (ICARC), pp. 90–95 (2023). IEEE

    2023

  10. [17]

    In: 2023 Third International Conference on Artificial Intelligence and Smart Energy (ICAIS), pp

    Kumar, M.A., Panitini, M., Vemulapalli, S., Sai, M.J.N.V.: Deep learning based automatic radiology report generation. In: 2023 Third International Conference on Artificial Intelligence and Smart Energy (ICAIS), pp. 1521–1526 (2023). IEEE

    2023

  11. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Deria, A., Kumar, K., Chakraborty, S., Mahapatra, D., Roy, S.: Inverge: Intelli- gent visual encoder for bridging modalities in report generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 2028–2038 (2024)

    2024

  12. [19]

    Scientific data 9(1), 122 (2022)

    Khaled, R., Helal, M., Alfarghaly, O., Mokhtar, O., Elkorany, A., El Kassas, H., Fahmy, A.: Categorized contrast enhanced mammography dataset for diagnostic and artificial intelligence research. Scientific data 9(1), 122 (2022)

    2022

  13. [20]

    In: CEUR Workshop Proceedings (2024)

    Hoque, M., Hasan, M.R., Emon, M., Khalifa, F., Rahman, M.: Medical image interpretation with large multimodal models. In: CEUR Workshop Proceedings (2024). CEUR Workshop Proceedings 3740, CEUR-WS. org 2024

    2024

  14. [22]

    arXiv preprint arXiv:2410.17357 (2024)

    Dawidowicz, G., Hirsch, E., Tal, A.: Image-aware evaluation of generated medical 15 reports. arXiv preprint arXiv:2410.17357 (2024)

    2024 arXiv

  15. [23]

    Yu, F., Endo, M., Krishnan, R., Pan, I., Tsai, A., Reis, E.P., Fonseca, E., Lee, H., Shakeri, Z., Ng, A., et al.: Radiology report expert evaluation (rexval) dataset (2023)

    2023

  16. [24]

    arXiv preprint arXiv:2410.00379 (2024)

    Wang, X., Wang, F., Li, Y., Ma, Q., Wang, S., Jiang, B., Li, C., Tang, J.: Cxpmrg- bench: Pre-training and benchmarking for x-ray medical report generation on chexpert plus dataset. arXiv preprint arXiv:2410.00379 (2024)

    2024 arXiv

  17. [25]

    arXiv preprint arXiv:2405.19538 (2024)

    Chambon, P., Delbrouck, J.-B., Sounack, T., Huang, S.-C., Chen, Z., Varma, M., Truong, S.Q., Chuong, C.T., Langlotz, C.P.: Chexpert plus: Hundreds of thousands of aligned radiology texts, images and patients. arXiv preprint arXiv:2405.19538 (2024)

    2024 arXiv

  18. [26]

    arXiv preprint arXiv:2411.01153 (2024)

    Singh, S.: Designing a robust radiology report generation system. arXiv preprint arXiv:2411.01153 (2024)

    2024 arXiv

  19. [27]

    In: 2024 11th International Conference on Wireless Networks and Mobile Communications (WINCOM), pp

    Cheddi, F., Habbani, A., Nait-Charif, H.: A multi-modal feature fusion-based approach for chest x-ray report generation. In: 2024 11th International Conference on Wireless Networks and Mobile Communications (WINCOM), pp. 1–7 (2024). IEEE

    2024

  20. [28]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp

    Wang, X., Peng, Y., Lu, L., Lu, Z., Bagheri, M., Summers, R.M.: Chestx-ray8: Hospital-scale chest x-ray database and benchmarks on weakly-supervised classi- fication and localization of common thorax diseases. In: Proceedings of the IEEE Conference on Computer Vision and Patte...

    2017

  21. [29]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Mao, X., Qi, G., Chen, Y., Li, X., Duan, R., Ye, S., He, Y., Xue, H.: Towards robust vision transformer. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 12042–12051 (2022)

    2022

  22. [30]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp

    Liu, Z., Lin, Y., Cao, Y., Hu, H., Wei, Y., Zhang, Z., Lin, S., Guo, B.: Swin trans- former: Hierarchical vision transformer using shifted windows. In: Proceedings of the IEEE/CVF International Conference on Computer Vision, pp. 10012–10022 (2021)

    2021

  23. [31]

    OpenAI blog 1(8), 9 (2019)

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al.: Language models are unsupervised multitask learners. OpenAI blog 1(8), 9 (2019)

    2019

  24. [32]

    arXiv preprint arXiv:1910.13461 (2019)

    Lewis, M.: Bart: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension. arXiv preprint arXiv:1910.13461 (2019)

    2019 arXiv

  25. [33]

    In: Text Summarization Branches Out, pp

    Lin, C.-Y.: Rouge: A package for automatic evaluation of summaries. In: Text Summarization Branches Out, pp. 74–81 (2004) 16

    2004

  26. [34]

    In: Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pp

    Papineni, K., Roukos, S., Ward, T., Zhu, W.-J.: 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)

    2002

  27. [35]

    arXiv preprint arXiv:1904.09675 (2019) 17

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluat- ing text generation with bert. arXiv preprint arXiv:1904.09675 (2019) 17

    2019 arXiv

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.