Pith. sign in

REVIEW 4 major objections 4 minor 43 references

Evaluation of Multilingual Image Captioning: How far can we get with CLIP models?

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

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

Pith's one-line read The paper shows that a single fine-tuned multilingual CLIP model can evaluate image-caption quality across ten languages as reliably as English-only models do for English, using machine-translated benchmarks with quality filtering.

desk verdict A useful, reproducible multilingual captioning metric paper whose headline cross-lingual human-correlation numbers rest on machine-translated VICR with English ratings, a caveat the authors acknowledge but do not fully resolve. read the letter →

arxiv 2502.06600 v2 pith:FCUFKSPV submitted 2025-02-10 cs.CL cs.AI

Gonçalo Gomes , Chrysoula Zerva , Bruno Martins This is my paper · ORCID
classification cs.CLcs.AI
keywords multilingualimagecaptioningCLIPScorevision-languagemodelsmachinetranslationqualityestimationhumanjudgmentcorrelationmulticulturalbenchmarkscaptionevaluationmetrics
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 image-caption evaluation need not be English-centric: a single multilingual CLIP model, fine-tuned on multicultural image-text data and machine-translated captions carrying human ratings, can score caption quality in ten languages with strong correlation to human preferences. The authors build the missing multilingual test data by translating English benchmarks with a quality-aware machine-translation pipeline and by adapting natively multilingual datasets such as XVNLI and MaRVL into caption-scoring tasks. They report that the fine-tuned multilingual CLIPScore matches or beats English-only CLIPScore on English benchmarks while also generalizing across high-, medium-, and low-resource languages, and that it outperforms the unfine-tuned model even when no reference captions are used. If this holds, multilingual caption evaluation can be done with one model plus translated benchmarks, without collecting new human ratings for every language.

What carries the argument

The engine is CLIPScore, a reference-free metric defined as $\text{CLIPScore}(c,v)=w\cdot\max(\cos(c,v),0)$ with $w=2.5$, computed from the image and text embeddings of a CLIP model; the paper uses a large multilingual CLIP with a ViT-H/14 vision encoder and an XLM-RoBERTa text encoder. It fine-tunes this model with two losses: the standard CLIP contrastive loss on CrossModal-3600 images and captions, and a Pearson-correlation loss on machine-translated VICR pairs that pulls CLIPScore values toward human ratings. The translated training data is produced by the M2M100 translation model with beam search and filtered by COMETKiwi, selecting for each caption the translation with the highest quality estimate, which is what lets English human ratings stand in for all target languages.

What would settle it

Collect target-language human ratings for a sample of the translated VICR captions and compare them with the original English ratings: if the fine-tuned CLIPScore correlates with the target-language ratings no better than with the transferred English ratings, or if the correlation drops below monolingual English levels for any language, the translation-transfer assumption is falsified.

Watch

Extended reading notes

Core claim

The central claim is that a fine-tuned multilingual CLIPScore model carries human caption-quality judgments across languages. On machine-translated versions of the VICR dataset, it correlates strongly with the original English human ratings in English, French, German, Spanish, Chinese, Portuguese, Italian, Russian, Korean, and Dutch, with the fine-tuned reference-free model reaching an average Spearman $\rho$ around 83.5 and beating reference-based versions of the unfine-tuned model. The same model sets the best or near-best accuracy on the VALSE linguistic-phenomenon tests in most languages and generally outperforms an English-only model on the native multicultural XVNLI and MaRVL tasks, with the exception of very low-resource Tamil and Swahili cases. The authors also claim that multilingual models trained on the same amount of data but with less English-specific text perform just as well as English-only models on English benchmarks while being more versatile.

Load-bearing premise

The load-bearing premise is that English human quality ratings stay valid for machine-translated captions once a COMETKiwi quality filter accepts the translation, and the paper does not test this with target-language human ratings.

Editorial extensions

If this is right

  • Reference-free evaluation in ten languages becomes feasible with one fine-tuned model, removing the need for human-written references per language.
  • Machine-translated benchmarks with quality filtering can substitute for missing multilingual evaluation data, at least when the translation-transfer assumption holds.
  • Multilingual training with less English-specific data does not sacrifice English performance, so language coverage can be added without an English-quality tax.
  • The fine-tuned model's gains on VALSE quantifier, plurality, and counting foils show that a caption-scoring metric can be made more sensitive to linguistic phenomena, not just visual grounding.
  • Back-translating test captions to English in order to use an English-only model adds compute and latency without accuracy benefit, making direct multilingual scoring the cheaper route.

Reading between the lines

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

  • Beyond the paper, the same translation-plus-quality-filter pipeline could be reused to extend caption-quality evaluation to dozens of languages with no new human annotation, simply by running M2M100 and COMETKiwi over existing English benchmarks.
  • Beyond the paper, the cross-language CLIPScore correlation heatmaps suggest the metric could double as a detector of translationese or mistranslation: a drop in cross-lingual score correlation for a language pair would flag captions whose quality was not preserved.
  • Beyond the paper, the qualitative finding that fine-tuned scores favor culturally specific captions (e.g., codfish for Portugal, cheongsam for China) points to a testable extension: measure whether the finetuning reduces cultural bias in caption ranking on a dedicated bias benchmark.
  • Beyond the paper, if target-language human ratings were collected, one could quantify how much COMETKiwi's threshold should vary per language, turning the translation-transfer assumption into a tunable, per-language calibration.
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 multilingual extension of CLIPScore for image captioning evaluation. The authors select a large multilingual CLIP model, fine-tune it on a combination of natively multilingual CrossModal-3600 captions and machine-translated VICR captions with transferred English human ratings, using a contrastive loss plus a Pearson-correlation loss against the transferred ratings. They evaluate the resulting model (MLF) in three settings: correlation with human ratings on machine-translated versions of four English benchmarks (headline result in Table 2 for VICR), robustness classification on machine-translated VALSE, and classification accuracy on natively multilingual XVNLI and MaRVL. They report that MLF matches or exceeds English-only models on English assessments and achieves high correlations with human judgments across ten languages, while also performing well on multilingual/multicultural classification tasks. The paper releases code and adapted datasets and provides extensive appendices on model selection, English-data comparisons, the machine-translation scheme, and per-language results.

Significance. If the headline claim is correct, the paper offers a practical recipe: a single fine-tuned multilingual CLIP model plus quality-filtered machine-translated benchmarks can evaluate captions in ten languages without language-specific human ratings. The experimental effort is substantial: multiple models, multiple correlation metrics with significance tests, held-out VICR splits, natively multilingual evaluation datasets, and public code/data. The paper is also honest in its Limitations section about the risks of machine-translated data. However, the central cross-lingual claim depends on the unverified transferability of English human ratings to M2M100/COMETKiwi-selected translations; this assumption is acknowledged in Appendix D and the Limitations section but not independently validated. The contribution is significant if that assumption is tested and confirmed; until then, the measured cross-lingual correlations may overstate performance on natural captions.

major comments (4)
  1. [Section 4.3.2, Table 2, and Appendix D] The headline cross-lingual correlations are computed against human ratings that were transferred from the original English VICR captions to M2M100 translations selected by COMETKiwi (Appendix D). No target-language human judgments are collected. The test set is therefore generated by exactly the same machine-translation and quality-estimation pipeline used to create the fine-tuning data, making the evaluation in-distribution with respect to both the MT system and the selection filter. This setup cannot establish that MLF 'strongly correlates with human preferences' for natively written or naturally occurring captions in German, French, Spanish, Chinese, Portuguese, Italian, Russian, Korean, or Dutch. The authors should either collect native-speaker ratings on a sample of the translated test items (or on natively written captions) and re-report the correlations, or explicitly restrict the claim to machine-translated caption evaluation.
  2. [Section 4.3.2, paragraph on heatmaps] The statement that the improved cross-lingual correlations observed after fine-tuning 'reflect the high quality of our multilingual training data, further validating our translation strategy' is circular. The fine-tuned model was trained on the same machine-translated VICR data whose translations are being validated, so an improvement on that test distribution is expected regardless of whether the English human ratings remain valid after translation. The translation strategy can be validated only with independent evidence, such as human judgments in the target languages or a separate translation-quality evaluation; the current paragraph does not provide such evidence.
  3. [Appendix D and Section 4.3.2] The best-of-25 COMETKiwi selection procedure changes the distribution of captions: the selected translations are likely more fluent, more uniform, and simpler than ordinary captions, and may contain translationese. Even if the English ratings were perfectly preserved, correlations on such filtered translations need not equal correlations on natural captions in the target language. The paper should analyze how the correlation varies with the COMETKiwi threshold (e.g., a continuous curve rather than only top/bottom quartiles) and should temper the inference from Table 2 to natural multilingual captions.
  4. [Limitations and Ethical Considerations; Abstract; Section 1] The Limitations section concedes that machine-translated data 'can still lead to different types of biases and to the reliance on artificially impoverished language' and defers manual curation to future work. This directly conflicts with the abstract's and Section 1's framing that the MT-based extension preserves the original benchmarks' quality and with the paper's claim that multilingual CLIPScore models 'maintain a high correlation with human judgements across different languages.' The authors should reconcile these statements by either adding the missing validation or reformulating the contribution and the headline claims to match the acknowledged limitation.
minor comments (4)
  1. [Appendix E, Tables 7 and 8 caption] The caption says 'respectfully' where it should say 'respectively'; this occurs in the sentence describing the tables.
  2. [Table 2 caption and Section 4.3.2] The reported 'std' values are bootstrap standard deviations of the correlation estimates across 1,000 stratified 80% subsets, not the standard deviation across models or languages; a footnote or clarified caption would prevent misinterpretation.
  3. [Table 1 caption] The columns named 'Size (B)' mix model parameter counts and training-data sizes in billions; the caption should clearly distinguish the two quantities, as the current wording is ambiguous.
  4. [Throughout, e.g., Section 4.1 and Table 3] The dataset name 'VALSE' is inconsistently rendered with a space ('V ALSE') in several places; these should be normalized.

Circularity Check

2 steps flagged · score 4.0 of 10

Cross-lingual human-correlation claims rest on English VICR ratings transferred through the same M2M100/COMETKiwi pipeline used to create the training data, and the paper explicitly validates that pipeline with in-distribution improvement; English and native XVNLI/MaRVL results remain independent.

  1. self definitional [Section 4.3.2 (Table 2) with Section 3 Eq. (2) and Appendix D]
    "we selected for each instance the translation that scored higher based on a large COMETKiwi model ... Low-quality translations ... will impact the caption and break our assumption that human ratings for the English data can be transferred across languages ... Table 2 displays the correlation between multilingual CLIPScore values and human ratings across the different languages."

    The non-English 'human ratings' in Table 2 are not native-language judgments; they are the original English VICR ratings attached to M2M100 translations selected by COMETKiwi. The model is fine-tuned with a Pearson correlation loss (Eq. 2) on exactly this kind of machine-translated VICR data, so the evaluation target is generated by the same pipeline that produced the training target. The reported cross-lingual human correlation is therefore, by construction, a correlation with transferred English preferences on MT output, not with independent human preferences in those languages. The paper's own Appendix D states that transferability is an assumption, and no native-language human ratings are collected to validate it.

  2. other [Section 4.3.2, final paragraph after Figure 1]
    "Although this improvement is expected, given that the finetuned model was trained on in-distribution data, it also reflects the high quality of our multilingual training data, further validating our translation strategy."

    This passage uses the finetuned model's improved cross-lingual correlations to validate the translation strategy, but immediately concedes that the improvement is expected because the model was trained on in-distribution data. Since the training data and the evaluation data come from the same M2M100/COMETKiwi pipeline, the observed consistency is a self-referential check: it confirms that the model fits the distribution created by that pipeline, not that the pipeline preserves human quality judgments. Without an external native-language ground truth, this step cannot validate the translation strategy.

full rationale

The paper is not circular in its English-centric core: Tables 1, 6, 7, and 8 evaluate CLIPScore variants against real English human ratings on held-out splits, which is standard supervised evaluation and is self-contained. The native multilingual benchmarks XVNLI and MaRVL (Tables 4 and 5) provide independent, externally sourced ground truth and support the secondary claim that multilingual models handle diverse semantic and cultural content. The circularity is localized to the cross-lingual human-correlation claims in Table 2 and to the validation of the translation scheme. There, the non-English ground truth is defined as English VICR ratings transferred through M2M100 and filtered by COMETKiwi; the fine-tuning objective (Pearson correlation with those transferred ratings) matches the evaluation objective on the same constructed distribution. The paper is transparent about this: Appendix D explicitly calls the transferability of English ratings an assumption, and the Limitations section concedes that machine-translated data 'can still lead to different types of biases and to the reliance on artificially impoverished language' and leaves manually curated benchmarks to future work. Section 4.3.2 further uses in-distribution improvement to 'validate' the translation strategy, which is a self-referential loop. These issues make the headline claim of strong correlation with human preferences in medium- and low-resource languages partially circular, but they do not undermine the English results or the independent native-data experiments. There is no load-bearing self-citation: COMETKiwi and M2M100 are external, publicly evaluated systems, and the authors' self-citations appear only as related work. Overall score 4 reflects partial circularity in the central cross-lingual evaluation claim with substantial independent content remaining.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new theoretical entities are introduced. The evaluation rests on standard CLIPScore plus assumptions about MT quality transfer and the validity of repurposed datasets.

free parameters (2)
  • CLIPScore re-scaling constant w = 2.5
    Fixed in Eq. (3) and inherited from Hessel et al. (2021); not re-estimated in this paper, but it enters every CLIPScore value whose correlations are reported.
  • Contrastive temperature tau = not reported
    Used in Eq. (1) for finetuning; the paper does not state its value or whether it was tuned.
assumptions (4)
  • domain assumption English human quality ratings transfer to machine-translated captions that pass the COMETKiwi filter.
    This is the load-bearing premise of the MT-based benchmarks (Sections 4.1 and Appendix D); no target-language human judgments are collected to test it.
  • domain assumption COMETKiwi scores are a reliable proxy for translation quality in this setting.
    Used in Appendix D to choose one of 25 candidate translations per caption; no validation against human-rated translations for captioning.
  • domain assumption CLIP embedding cosine similarity is a valid measure of caption-image semantic correspondence.
    Foundation of CLIPScore (Eq. 3), inherited from Hessel et al. (2021).
  • domain assumption The repurposed XVNLI and MaRVL classification tasks measure caption evaluation quality.
    Defined in Section 4.3.3 and Appendix F; the paper itself notes SNLI-VE text-only bias and CLIP concept-association bias in MaRVL, so this assumption is only partially supported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluation of Multilingual Image Captioning: How far can we get with CLIP models?." pith.science (2026). https://pith.science/paper/FCUFKSPV

@misc{pith2026250206600,
  author       = {Pith},
  title        = {Pith review of: Evaluation of Multilingual Image Captioning: How far can we get with CLIP models?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FCUFKSPV}},
  note         = {Machine review of arXiv:2502.06600}
}
read the original abstract

The evaluation of image captions, looking at both linguistic fluency and semantic correspondence to visual contents, has witnessed a significant effort. Still, despite advancements such as the CLIPScore metric, multilingual captioning evaluation has remained relatively unexplored. This work presents several strategies, and extensive experiments, related to evaluating CLIPScore variants in multilingual settings. To address the lack of multilingual test data, we consider two different strategies: (1) using quality aware machine-translated datasets with human judgements, and (2) re-purposing multilingual datasets that target semantic inference and reasoning. Our results highlight the potential of finetuned multilingual models to generalize across languages and to handle complex linguistic challenges. Tests with machine-translated data show that multilingual CLIPScore models can maintain a high correlation with human judgements across different languages, and additional tests with natively multilingual and multicultural data further attest to the high-quality assessments.

Figures

Figures reproduced from arXiv: 2502.06600 by the authors.

Figure 1
Figure 1. Pearson correlation scores between different languages, for the original multilingual CLIPScore model [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. The three different XVNLI multilingual clas [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Multilingual CLIPScore values for image-caption pairs featuring concepts biased to particular languages. [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 33 canonical work pages

  1. [1]

    Somak Aditya, Yezhou Yang, Chitta Baral, Cornelia Fermuller, and Yiannis Aloimonos. 2015. From images to sentences through scene description graphs using commonsense reasoning and knowledge. arXiv preprint arXiv:1511.03292

    arXiv 2015

  2. [2]

    Saba Ahmadi and Aishwarya Agrawal. 2024. An examination of the robustness of reference-free image captioning evaluation metrics. In Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics

    work page 2024

  3. [3]

    Duarte M Alves, Jos \'e Pombal, Nuno M Guerreiro, Pedro H Martins, Jo \ a o Alves, Amin Farajian, Ben Peters, Ricardo Rei, Patrick Fernandes, Sweta Agrawal, et al. 2024. Tower: An open multilingual large language model for translation-related tasks. arXiv preprint arXiv:2402.17733

    arXiv 2024

  4. [4]

    Emanuele Bugliarello, Fangyu Liu, Jonas Pfeiffer, Siva Reddy, Desmond Elliott, Edoardo Maria Ponti, and Ivan Vuli \'c . 2022. IGLUE : A benchmark for transfer learning across modalities, tasks, and languages. In Proceedings of the International Conference on Machine Learning

    work page 2022

  5. [5]

    David Chan, Suzanne Petryk, Joseph Gonzalez, Trevor Darrell, and John Canny. 2023. CLAIR : Evaluating image captions with large language models. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    work page 2023

  6. [6]

    Tianyu Cui, Jinbin Bai, Guohua Wang, Qingguo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Ye Shi. 2025. Evaluating image caption via cycle-consistent text-to-image generation. arXiv preprint arXiv:2501.03567

    work page Pith review arXiv 2025

  7. [7]

    Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Michael Auli, and Armand Joulin. 2021. Beyond english-centric multilingual machine translation. Journal of Machine Learning Research

    work page 2021

  8. [8]

    Gregor Geigle, Abhay Jain, Radu Timofte, and Goran Glava s . 2023. m BLIP : Efficient bootstrapping of multilingual vision- LLM s. arXiv preprint arXiv:2307.06930

    arXiv 2023

Show all 43 references
  1. [9]

    Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPS core: A reference-free evaluation metric for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    2021

  2. [10]

    Micah Hodosh, Peter Young, and Julia Hockenmaier. 2013. Framing image description as a ranking task: Data, models and evaluation metrics. Journal of Artificial Intelligence Research

    2013

  3. [11]

    Anwen Hu, Shizhe Chen, Liang Zhang, and Qin Jin. 2023. InfoMetIC : An informative metric for reference-free image caption evaluation. In Proceedings of the Annual Meeting of the Association for Computational Linguistics

    2023

  4. [12]

    Maxime Kayser, Oana-Maria Camburu, Leonard Salewski, Cornelius Emde, Virginie Do, Zeynep Akata, and Thomas Lukasiewicz. 2021. E-VIL : A dataset and benchmark for natural language explanations in vision-language tasks. In Proceedings of the IEEE/CVF International Conference on ...

    2021

  5. [13]

    Jin-Hwa Kim, Yunji Kim, Jiyoung Lee, Kang Min Yoo, and Sang-Woo Lee. 2022. Mutual information divergence: A unified metric for multimodal generative models. In Proceedings of the Annual Meeting on Neural Information Processing Systems

    2022

  6. [14]

    Yongil Kim, Yerin Hwang, Hyeongu Yun, Seunghyun Yoon, Trung Bui, and Kyomin Jung. 2023. PR-MCS: perturbation robust metric for multilingual image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    2023

  7. [15]

    Yebin Lee, Imseong Park, and Myungjoo Kang. 2024. FLEUR : An explainable reference-free evaluation metric for image captioning using a large multimodal model. arXiv preprint arXiv:2406.06004

    2024 arXiv

  8. [16]

    Fangyu Liu, Emanuele Bugliarello, Edoardo Maria Ponti, Siva Reddy, Nigel Collier, and Desmond Elliott. 2021. Visually grounded reasoning across languages and cultures. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    2021

  9. [17]

    Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, and Luke Zettlemoyer. 2020. Multilingual denoising pre-training for neural machine translation. Transactions of the Association for Computational Linguistics

    2020

  10. [18]

    Jiasen Lu, Vedanuj Goswami, Marcus Rohrbach, Devi Parikh, and Stefan Lee. 2020. 12-in-1: Multi-task vision and language representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2020

  11. [19]

    Lothar D Narins, Andrew Scott, Aakash Gautam, Anagha Kulkarni, Mar Castanon, Benjamin Kao, Shasta Ihorn, Yue-Ting Siu, James M Mason, Alexander Blum, et al. 2024. Validated image caption rating dataset. In Proceedings of the Annual Meeting on Neural Information Processing Systems

    2024

  12. [20]

    Letitia Parcalabescu, Michele Cafagna, Lilitta Muradjan, Anette Frank, Iacer Calixto, and Albert Gatt. 2022. VALSE : A task-independent benchmark for vision and language models centered on linguistic phenomena. In Proceedings of the Annual Meeting of the Association for Comput...

    2022

  13. [21]

    Haoyi Qiu, Zi-Yi Dou, Tianlu Wang, Asli Celikyilmaz, and Nanyun Peng. 2023. Gender biases in automatic evaluation metrics for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    2023

  14. [22]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In Proceedings of the International Conferenc...

    2021

  15. [23]

    Rita Ramos, Emanuele Bugliarello, Bruno Martins, and Desmond Elliott. 2024. PAELLA : Parameter-efficient lightweight language-agnostic captioning model. In Findings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics

    2024

  16. [24]

    Rita Ramos, Desmond Elliott, and Bruno Martins. 2023 a . Retrieval-augmented image captioning. In Proceedings of the Conference of the European Chapter of the Association for Computational Linguistics

    2023

  17. [25]

    Rita Ramos, Bruno Martins, and Desmond Elliott. 2023 b . LMCap : Few-shot multilingual image captioning by retrieval augmented language model prompting. In Findings of the Association for Computational Linguistics

    2023

  18. [26]

    Rita Ramos, Bruno Martins, Desmond Elliott, and Yova Kementchedjhieva. 2023 c . SmallCap : Lightweight image captioning prompted with retrieval augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  19. [27]

    Ricardo Rei, Nuno M Guerreiro, Daan van Stigt, Marcos Treviso, Lu \' sa Coheur, Jos \'e GC de Souza, Andr \'e FT Martins, et al. 2023. Scaling up COMETKiwi : Unbabel-IST 2023 submission for the quality estimation shared task. In Proceedings of the Conference on Machine Translation

    2023

  20. [28]

    Ricardo Rei, Marcos Treviso, Nuno M Guerreiro, Chrysoula Zerva, Ana C Farinha, Christine Maroti, Jos \'e GC de Souza, Taisiya Glushkova, Duarte Alves, Lu \' sa Coheur, et al. 2022. COMETKiwi : IST -unbabel 2022 submission for the quality estimation shared task. In Proceedings ...

    2022

  21. [29]

    Sara Sarto, Manuele Barraco, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. 2023. Positive-augmented contrastive learning for image and video captioning evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2023

  22. [30]

    Sara Sarto, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. 2024 a . BRIDGE : Bridging gaps in image captioning evaluation with stronger visual cues. In Proceedings of the European Conference on Computer Vision

    2024

  23. [31]

    Sara Sarto, Nicholas Moratelli, Marcella Cornia, Lorenzo Baraldi, and Rita Cucchiara. 2024 b . Positive-augmented contrastive learning for vision-and-language evaluation and training. arXiv preprint arXiv:2410.07336

    2024 arXiv

  24. [32]

    Haoyu Song, Li Dong, Weinan Zhang, Ting Liu, and Furu Wei. 2022. CLIP models are few-shot learners: Empirical studies on VQA and visual entailment. In Proceedings of the Annual Meeting of the Association for Computational Linguistics

    2022

  25. [33]

    Alane Suhr, Stephanie Zhou, Ally Zhang, Iris Zhang, Huajun Bai, and Yoav Artzi. 2019. A corpus for reasoning about natural language grounded in photographs. In Proceedings of the Annual Meeting of the Association for Computational Linguistics

    2019

  26. [34]

    Ashish V Thapliyal, Jordi Pont Tuset, Xi Chen, and Radu Soricut. 2022. CrossModal-3600 : A massively multilingual multimodal evaluation dataset. In Proceedings of the Conference on Empirical Methods in Natural Language Processing

    2022

  27. [35]

    Tony Cheng Tong, Sirui He, Zhiwen Shao, and Dit-Yan Yeung. 2024. G- VE val: A versatile metric for evaluating image and video captions using GPT -4o. arXiv preprint arXiv:2412.13647

    2024 arXiv

  28. [36]

    Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. 2015. CIDEr : Consensus-based image description evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2015

  29. [37]

    Yuiga Wada, Kanta Kaneda, Daichi Saito, and Komei Sugiura. 2024. Polos: Multimodal metric learning from human feedback for image captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    2024

  30. [38]

    Yuiga Wada, Kanta Kaneda, and Komei Sugiura. 2023. Ja SPICE : Automatic evaluation metric using predicate-argument structures for image captioning models. In Proceedings of the Conference on Computational Natural Language Learning

    2023

  31. [39]

    Ning Xie, Farley Lai, Derek Doran, and Asim Kadav. 2019. Visual entailment: A novel task for fine-grained image understanding. arXiv preprint arXiv:1901.06706

    2019 arXiv

  32. [40]

    Yutaro Yamada, Yingtian Tang, and Ilker Yildirim. 2022. When are lemons purple? the concept association bias of CLIP . arXiv preprint arXiv:2212.12043

    2022 arXiv

  33. [41]

    Bang Yang, Fenglin Liu, Xian Wu, Yaowei Wang, Xu Sun, and Yuexian Zou. 2023. MultiCapCLIP : Auto-encoding prompts for zero-shot multilingual visual captioning. In Proceedings of the Annual Meeting of the Association for Computational Linguistics

    2023

  34. [42]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  35. [43]

    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 8, 2026 · model on record in the stance chip above.