Pith. sign in

REVIEW 3 major objections 5 minor 37 references

Investigating the Effect of Parallel Data in the Cross-Lingual Transfer for Vision-Language Encoders

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read For cross-lingual transfer of vision-language encoders, machine-translated task data is the best all-round parallel data, though authentic caption-like data wins in some languages and most languages benefit from multilingual pooling.

desk verdict Useful empirical study of parallel data types for cross-lingual VL encoder transfer, but the headline 'Task MT best' result is confounded by domain overlap with the downstream task. read the letter →

arxiv 2504.21681 v2 pith:AULNLWRP submitted 2025-04-30 cs.CL

classification cs.CL
keywords cross-lingualtransfervision-languagemodelsparalleldataselectionmultilingualtrainingvisualreasoningmachine-translatedencoderlow-resourcelanguages
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 investigates how to extend a fine-tuned English vision-language model to other languages by replacing its English text encoder with a multilingual one and adapting the new encoder on parallel data. It compares four kinds of parallel data — machine-translated task sentences, machine-translated captions, generic parallel text, and authentic caption-like sentences — and finds that machine-translated task data is best on average, while authentic caption-like data wins in several individual languages. It also compares bilingual transfer with transfer on a single encoder trained on many languages, finding that multilingual training helps most languages in very low-resource settings, with the benefit peaking around 20 languages. These results offer concrete guidance for building multilingual vision-language systems without large multilingual pre-training.

What carries the argument

The method replaces BridgeTower's English RoBERTa text encoder with XLM-R and trains XLM-R to reproduce RoBERTa's hidden states. The training signal is a subword-aligned imitation loss: Eflomal word alignments are projected to subword level, and the mean squared error between aligned hidden states is averaged with a mean-pooled state-matching loss over the last six layers that feed the cross-modal encoder. A learnable weighted mixture of layers, followed by a bottleneck projection with GELU and layer normalization, gives XLM-R the flexibility to match RoBERTa's representation space. This machinery is what makes the transfer possible, and the paper varies the parallel data and the number of languages while keeping it fixed.

What would settle it

Repeat the transfer experiments with machine-translated data from a different task (e.g., translated VQA or image-captioning data) and authentic parallel data from that same task. If the translated task data no longer beats authentic task data on average, the claim that machine-translated task data are the best parallel data type is not general.

Watch

Extended reading notes

Core claim

The paper's central empirical claim is that for cross-lingual transfer of a vision-language encoder via parallel data, the data's source matters more than whether it is machine-translated or authentic. On average across MARVL, M5-VGR, and XVNLI, machine-translated sentences from the same downstream task (NLVR2) give the highest accuracy, beating machine-translated COCO captions, generic OPUS-100 parallel sentences, and caption-like filtered OPUS-100 data. But the caption-like authentic data outperform machine-translated data for several languages, so translation quality alone does not determine the ranking. In addition, a single XLM-R encoder trained on parallel data from all five MARVL languages beats separate bilingual encoders when each language has only 5k pairs, and adding up to 20 languages continues to improve average performance before diminishing returns set in.

Load-bearing premise

The comparison of parallel data types assumes the type itself is the explanatory variable, but the datasets also differ in domain and filtering: Task MT comes from the same NLVR2 task used for fine-tuning, so its average advantage may be an in-domain effect rather than a property of machine-translated task data.

Editorial extensions

If this is right

  • When extending a vision-language encoder to many languages, machine-translated data from the target task is a strong default: it gives the highest average accuracy in the paper's three benchmarks.
  • Authentic caption-like parallel data, selected by a caption classifier, can replace machine-translated captions and outperform them on several languages, offering a cheaper and sometimes better alternative.
  • Pooling several languages into a single multilingual encoder helps most languages in the very low-resource case (5k parallel pairs per language), compared with separate bilingual encoders.
  • The multilingual benefit grows as languages are added but peaks around 20 languages; beyond that, average accuracy slightly declines.
  • The paper's approach matches CliCoTea's reported performance using a fraction of the parallel data (25k pairs), suggesting data selection quality matters as much as data quantity.

Reading between the lines

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

  • The task-MT advantage may partly reflect domain match rather than translation: the machine-translated task data come from NLVR2, the same task used for fine-tuning; a test on a different task domain would separate these factors.
  • The caption-like selection method suggests that authentic parallel data filtered for task similarity could reduce dependence on machine translation services, which is a practical route for low-resource settings where MT engines are unavailable.
  • The language-synergy curve (help up to 20 languages, then decline) echoes the known trade-off in multilingual models between shared representation and per-language capacity; the optimal language count likely shifts with per-language data size and language distance.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper studies cross-lingual transfer for vision-language encoders by replacing a fine-tuned English text encoder with a multilingual encoder trained to mimic the English encoder's hidden states on parallel data. The authors extend CliCoTea with a subword-level alignment loss and a bottleneck projection, and they evaluate four parallel-data strategies (machine-translated task data, machine-translated captions, generic parallel data, and caption-like authentic parallel data) plus scaling in the number of languages. Experiments on MARVL, M5-VGR, and XVNLI lead to the claims that machine-translated task data is best on average, caption-like authentic data is better in some languages, and most languages benefit from multilingual training.

Significance. If the claims hold, the paper gives practical guidance for building multilingual VL systems: task-domain machine-translated data is a strong transfer signal, and pooling multiple languages helps in low-resource settings. The paper also demonstrates that a lightweight encoder-transfer method can match CliCoTea with less data, and it provides a useful comparison across several multilingual benchmarks. These contributions are of interest to practitioners in multilingual multimodal learning. However, the significance is conditional on ruling out the confounds described below and on establishing that the reported differences are not within run-to-run variance.

major comments (3)
  1. [§4.2, Table 1] The Task MT condition uses machine-translated sentences from NLVR2, while the VL model in the same pipeline is fine-tuned on English task data, including NLVR2 (Section 4.1, Figure 1). The alignment loss in Eq. (1) therefore encourages XLM-R to reproduce the representations of exactly the in-domain task text, whereas Caption MT (MSCOCO) and Caption-like (OPUS-100) come from other domains. This confounds 'translation method' with 'domain overlap with the fine-tuning task,' so the reported advantage of Task MT over the other conditions could be due to the latter rather than to any property of machine-translated task data. Please add a control (e.g., task MT from a task other than the fine-tuning task, or authentic task-domain data) or re-frame the headline conclusion.
  2. [Tables 1-4] All results are from single runs without error bars, confidence intervals, or significance tests. Many differences that carry the conclusions are small, for example Table 3's averages 66.3, 66.5, 66.7, 66.7, 66.3 across language counts, and Table 2's Biling. 5k average (65.9) versus Multi. 25k average (66.3). Without multiple seeds and a paired significance test across the five MARVL languages or across benchmarks, the claims 'task MT is best on average' and 'most languages benefit from multilingual training' are not established.
  3. [§4.3, Tables 2-3] The bilingual-versus-multilingual comparison and the language-scaling experiment both confound the number of languages with the total amount of training data. In Table 2, 'Biling. 5k' uses 5k pairs total for one language while 'Multi. 25k' uses 25k pairs total (5k per language for five languages); in Table 3, the language count grows from 5 to 40 while keeping 5k pairs per language, so the total data grows from 25k to 200k. The improvements attributed to multilingual training or to adding languages could therefore be driven by larger training data rather than by language diversity. Please include matched-total-data controls (or state the data budget explicitly per condition) before drawing the Section 6 conclusion.
minor comments (5)
  1. [Table 1 caption] The caption refers to 'M5B-VGR' while the rest of the paper uses 'M5-VGR'; please unify the notation.
  2. [§4.3] The phrase 'cross-signal transfer' appears to be a typo for 'cross-lingual transfer'; please correct it.
  3. [§4.2] The manuscript does not state the number of sentence pairs used for the Task MT and Caption MT conditions; please specify the per-language data budget for these conditions so that the comparison with Generic and Caption-like (5k pairs) is explicit.
  4. [Table 2] The label 'Multi. 25k' is unclear because the Section 4.3 text says 5k caption-like pairs per language; please clarify whether 25k is the total across five languages.
  5. [References] References [13] and [14] appear to refer to the same paper (Visually Grounded Reasoning across Languages and Cultures); please deduplicate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical comparisons on external benchmarks, and the Task MT in-domain confound is a validity concern, not a definitional reduction.

full rationale

The paper's central claims are empirical, not derivational. The alignment loss (Eq. 1-3) trains XLM-R to mimic RoBERTa hidden states on parallel data; it never optimizes for the reported MARVL, M5-VGR, or XVNLI accuracies, and no parameter is fitted to those benchmarks. The 'Task MT best on average' result compares machine-translated NLVR2 sentences against other data types, and although Task MT shares a domain with the NLVR2 fine-tuning task, that is a potential confound in experimental design rather than a circular reduction: the reported accuracies come from held-out multilingual benchmarks, not from the training data or from any quantity defined in terms of those benchmarks. The 'caption-like' condition is selected with a classifier trained on COCO versus OPUS-100, which again does not use the target evaluation labels. The multilingual-training result compares the number of training languages against external test accuracy, and the comparison with CliCoTea is based on independently reported numbers from another group. There are no load-bearing self-citations: the method builds on CliCoTea but the cited work is external, and the paper reproduces the method and evaluates it on new benchmarks. No step in the paper reduces by construction to its own inputs, so the appropriate finding is no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper makes no mathematical derivation; its central claims are empirical and depend on corpus choices, alignment tools, and benchmark validity. No new entities are postulated. The main free parameters are hand-chosen data budgets, training epochs, and filtering thresholds.

free parameters (4)
  • Parallel sentence budget per language = 5,000 pairs per language; 25,000 pairs in the bilingual 25k condition
    The choice of 5k pairs per language is a hand-chosen constant that affects the balance between data-type comparisons; different budgets could change rankings. See Sections 4.2 and 4.3.
  • English sample size = 200,000 sentences
    Uniformly sampled from IGLUE training sets; the size is selected by hand and affects the English-only baseline and alignment priors. See Section 4.1.
  • Training epochs = 50 epochs for multilingual, 10 epochs for bilingual
    Epoch counts are not matched between bilingual and multilingual conditions, so the comparison in Table 2 may partly reflect under-training of bilingual models. See Section 4.3.
  • Caption selection thresholds = Top 5,000 pairs; maximum 450 tokens; BERTScore filtering
    Manual thresholds for filtering caption-like data; different thresholds could change the caption-like condition's quality. See Appendix A.
assumptions (5)
  • domain assumption Eflomal word alignments with the grow-diagonal heuristic produce reliable subword alignments between RoBERTa and XLM-R tokenizers.
    The alignment loss in Eq. 1 depends on alignment quality; noisy alignments would add noise to the distillation signal. Invoked in Section 4.1.
  • domain assumption XLM-R Large can be adapted to mimic RoBERTa Large's hidden states using only 5k-25k parallel sentences per language.
    If capacity or data is insufficient, the observed transfer differences could reflect training instability rather than data type. This is assumed in Sections 4.1 and 4.3.
  • domain assumption Google Translate output is representative of machine-translated parallel data.
    Both Task MT and Caption MT are produced with Google Translate; if the translation system is unrepresentative, the comparison of MT versus authentic data may not generalize. Used in Section 4.2.
  • domain assumption The caption classifier trained on COCO and OPUS-100 selects genuinely caption-like in-domain parallel data.
    The Caption-like condition depends on this filter; if the classifier selects on spurious cues, the condition's label is misleading. See Appendix A.
  • domain assumption MARVL, M5-VGR, and XVNLI are valid measures of cross-lingual vision-language competence.
    All conclusions are evaluated on these benchmarks; the accuracy numbers are treated as comparable across languages and datasets, which assumes benchmark validity. Used in Section 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Investigating the Effect of Parallel Data in the Cross-Lingual Transfer for Vision-Language Encoders." pith.science (2026). https://pith.science/paper/AULNLWRP

@misc{pith2026250421681,
  author       = {Pith},
  title        = {Pith review of: Investigating the Effect of Parallel Data in the Cross-Lingual Transfer for Vision-Language Encoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AULNLWRP}},
  note         = {Machine review of arXiv:2504.21681}
}
read the original abstract

Most pre-trained Vision-Language (VL) models and training data for the downstream tasks are only available in English. Therefore, multilingual VL tasks are solved using cross-lingual transfer: fine-tune a multilingual pre-trained model or transfer the text encoder using parallel data. We study the alternative approach: transferring an already trained encoder using parallel data. We investigate the effect of parallel data: domain and the number of languages, which were out of focus in previous work. Our results show that even machine-translated task data are the best on average, caption-like authentic parallel data outperformed it in some languages. Further, we show that most languages benefit from multilingual training.

Figures

Figures reproduced from arXiv: 2504.21681 by the authors.

Figure 1
Figure 1. Overall scheme of the proposed approach. The VL model is depicted as a box containing both the image and the text encoder, combined with the violet cross-modal encoder. The VL model is modified in two phases: (1) Fine-tune the entire model using task-specific data and (2) Encoder Transfer, which transfers the English text encoder capabilities into the multilingual one, using English task-specific data and bilingual … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 15 canonical work pages

  1. [1]

    IGLUE: A Benchmark for Transfer Learning across Modalities, Tasks, and Languages

    Bugliarello, E., Liu, F., Pfeiffer, J., Reddy, S., Elliott, D., Ponti, E.M., Vulic, I.: IGLUE: A Benchmark for Transfer Learning across Modalities, Tasks, and Languages. CoRR abs/2201.11732 (2022), https://arxiv.org/abs/2201.11732

  2. [2]

    In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=r1xCMyBtPS

    Cao, S., Kitaev, N., Klein, D.: Multilingual alignment of contextual word rep- resentations. In: International Conference on Learning Representations (2020), https://openreview.net/forum?id=r1xCMyBtPS

  3. [3]

    In: Proceedings of the Thirteenth Language Resources and Evaluation Conference

    Carlsson, F., Eisen, P., Rekathati, F., Sahlgren, M.: Cross-lingual and multilingual CLIP. In: Proceedings of the Thirteenth Language Resources and Evaluation Conference. pp. 6848–6854. European Language Resources Association, Marseille, France (Jun 2022), https://aclanthology.org/2022.lrec-1.739/

  4. [4]

    In: Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX

    Chen, Y., Li, L., Yu, L., Kholy, A.E., Ahmed, F., Gan, Z., Cheng, Y., Liu, J.: UNITER: UNiversal Image-TExt Representation Learning. In: Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXX. Lecture Notes in Computer Science, vol. 12375, pp. 104–

  5. [5]

    In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics

    Conneau, A., Khandelwal, K., Goyal, N., Chaudhary, V., Wenzek, G., Guzmán, F., Grave, E., Ott, M., Zettlemoyer, L., Stoyanov, V.: Unsupervised cross-lingual representation learning at scale. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 8440–8451. ACL, Online (Jul 2020). https://doi.org/10.18653/v1/2020.a...

  6. [6]

    In: Proceedings of the 2019 Con- ferenceoftheNorthAmericanChapteroftheAssociationforComputationalLinguis- tics: Human Language Technologies, Volume 1 (Long and Short Papers)

    Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: Pre-training of deep bidi- rectional transformers for language understanding. In: Proceedings of the 2019 Con- ferenceoftheNorthAmericanChapteroftheAssociationforComputationalLinguis- tics: Human Language Technologies, Volume 1 (Long and Short Papers). pp. 4171–

  7. [7]

    CoRR abs/2101.08231 (2021), https://arxiv.org/abs/2101.08231

    Dou, Z., Neubig, G.: Word Alignment by Fine-tuning Embeddings on Parallel Corpora. CoRR abs/2101.08231 (2021), https://arxiv.org/abs/2101.08231

  8. [8]

    In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Feng, F., Yang, Y., Cer, D., Arivazhagan, N., Wang, W.: Language-agnostic BERT sentence embedding. In: Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 878–891. ACL, Dublin, Ireland (May 2022). https://doi.org/10.18653/v1/2022.acl-long.62

Show all 37 references
  1. [9]

    In: Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR)

    Geigle, G., Jain, A., Timofte, R., Glavaš, G.: mBLIP: Efficient bootstrapping of multilingual vision-LLMs. In: Proceedings of the 3rd Workshop on Advances in Language and Vision Research (ALVR). pp. 7–25. ACL, Bangkok, Thailand (Aug 2024). https://doi.org/10.18653/v1/2024.alvr...

  2. [10]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vol- ume 2: Short Papers)

    Karoui, Y., Lebret, R., Foroutan Eghlidi, N., Aberer, K.: Stop pre-training: Adapt visual-language models to unseen languages. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Vol- ume 2: Short Papers). pp. 366–375. ACL, Toronto, Can...

  3. [11]

    CoRR abs/2208.02131 (2022)

    Kwon, G., Cai, Z., Ravichandran, A., Bas, E., Bhotika, R., Soatto, S.: Masked Vision and Language Modeling for Multi-modal Representation Learn- ing. CoRR abs/2208.02131 (2022). https://doi.org/10.48550/ARXIV.2208.02131, https://doi.org/10.48550/arXiv.2208.02131

  4. [12]

    In: Advances in Neural Information Processing Systems

    Li, J., Selvaraju, R., Gotmare, A., Joty, S., Xiong, C., Hoi, S.C.H.: Align before fuse: Vision and language representation learning with momentum distillation. In: Advances in Neural Information Processing Systems. vol. 34, pp. 9694–9705. Curran Associates, Inc. (2021)

  5. [13]

    In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing

    Liu, F., Bugliarello, E., Ponti, E.M., Reddy, S., Collier, N., Elliott, D.: Visu- ally grounded reasoning across languages and cultures. In: Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. pp. 10467–10485. ACL, Online and Punta Cana, Dom...

  6. [14]

    CoRRabs/2109.13238 (2021), https://arxiv.org/abs/2109.13238

    Liu, F., Bugliarello, E., Ponti, E.M., Reddy, S., Collier, N., Elliott, D.: Visually Grounded Reasoning across Languages and Cultures. CoRRabs/2109.13238 (2021), https://arxiv.org/abs/2109.13238

  7. [15]

    CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692

    Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., Stoyanov, V.: RoBERTa: A Robustly Optimized BERT Pretraining Approach. CoRR abs/1907.11692 (2019), http://arxiv.org/abs/1907.11692

  8. [16]

    CoRR abs/2006.02635 (2020), https://arxiv.org/abs/2006.02635

    Ni, M., Huang, H., Su, L., Cui, E., Bharti, T., Wang, L., Gao, J., Zhang, D., Duan, N.:M3P:LearningUniversalRepresentationsviaMultitaskMultilingualMultimodal Pre-training. CoRR abs/2006.02635 (2020), https://arxiv.org/abs/2006.02635

  9. [17]

    The Prague Bulletin of Mathematical Linguistics106(1), 125 (2016)

    Östling, R., Tiedemann, J.: Efficient word alignment with markov chain monte carlo. The Prague Bulletin of Mathematical Linguistics106(1), 125 (2016)

  10. [18]

    In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies

    Pfeiffer, J., Goyal, N., Lin, X., Li, X., Cross, J., Riedel, S., Artetxe, M.: Lifting the curse of multilinguality by pre-training modular transformers. In: Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human...

  11. [19]

    Pires, T., Schlinger, E., Garrette, D.: How multilingual is multilingual BERT? In: Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 4996–5001. ACL, Florence, Italy (Jul 2019). https://doi.org/10.18653/v1/P19-1493

  12. [20]

    In: Proceedings of the 38th International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sas- try, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Proceedings of the 38th International Conference on M...

  13. [21]

    In: International conference on machine learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International conference on machine learning. pp. 8748–8763. PMLR (2021)

  14. [22]

    In: Proceedings of the 2020 Conference on Empirical Investigating Cross-Lingual Transfer for VL Encoders 9 Methods in Natural Language Processing (EMNLP)

    Reimers, N., Gurevych, I.: Making monolingual sentence embeddings multilingual using knowledge distillation. In: Proceedings of the 2020 Conference on Empirical Investigating Cross-Lingual Transfer for VL Encoders 9 Methods in Natural Language Processing (EMNLP). pp. 4512–4525...

  15. [23]

    In: Findings of the Association for Computational Linguis- tics: EMNLP 2024

    Schneider, F., Sitaram, S.: M5 – a diverse benchmark to assess the perfor- mance of large multimodal models across multilingual and multicultural vision- language tasks. In: Findings of the Association for Computational Linguis- tics: EMNLP 2024. pp. 4309–4345. ACL, Miami, Flo...

  16. [24]

    CoRR abs/1811.00491 (2018), http://arxiv.org/abs/1811.00491

    Suhr, A., Zhou, S., Zhang, I., Bai, H., Artzi, Y.: A Corpus for Reasoning About Natural Language Grounded in Photographs. CoRR abs/1811.00491 (2018), http://arxiv.org/abs/1811.00491

  17. [25]

    In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)

    Tan, H., Bansal, M.: LXMERT: Learning cross-modality encoder representations from transformers. In: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). ...

  18. [26]

    In: Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12)

    Tiedemann, J.: Parallel data, tools and interfaces in OPUS. In: Proceedings of the Eighth International Conference on Language Resources and Evaluation (LREC’12). pp.2214–2218.EuropeanLanguageResourcesAssociation(ELRA),Istanbul,Turkey (May 2012), https://aclanthology.org/L12-1246/

  19. [27]

    In: Proceed- ings of the 2022 Conference on Empirical Methods in Natural Language Pro- cessing

    Vu, T., Barua, A., Lester, B., Cer, D., Iyyer, M., Constant, N.: Overcom- ing catastrophic forgetting in zero-shot cross-lingual generation. In: Proceed- ings of the 2022 Conference on Empirical Methods in Natural Language Pro- cessing. pp. 9279–9300. ACL, Abu Dhabi, United Ar...

  20. [28]

    Wu, S., Dredze, M.: Do explicit alignments robustly improve multilingual en- coders? In: Proceedings of the 2020 Conference on Empirical Methods in Nat- ural Language Processing (EMNLP). pp. 4471–4482. ACL, Online (Nov 2020). https://doi.org/10.18653/v1/2020.emnlp-main.362

  21. [29]

    In: Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI

    Xu, X., Wu, C., Rosenman, S., Lal, V., Che, W., Duan, N.: BridgeTower: Building Bridges between Encoders in Vision-language Representation Learn- ing. In: Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI

  22. [30]

    Zeng, Y., Zhang, X., Li, H., Wang, J., Zhang, J., Zhou, W.: X$^2$- VLM: All-In-One Pre-trained Model For Vision-Language Tasks (Jul 2023), http://arxiv.org/abs/2211.12402, arXiv:2211.12402 [cs]

  23. [31]

    In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)

    Zeng, Y., Zhou, W., Luo, A., Cheng, Z., Zhang, X.: Cross-view language mod- eling: Towards unified cross-lingual cross-modal pre-training. In: Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). pp. 5731–5746. ACL, T...

  24. [32]

    In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics

    Zhang, B., Williams, P., Titov, I., Sennrich, R.: Improving massively multilingual neural machine translation and zero-shot translation. In: Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics. pp. 1628–1639. ACL, Online (Jul 2020). https://...

  25. [33]

    Zhang, T., Kishore, V., Wu, F., Weinberger, K.Q., Artzi, Y.: Bertscore: Evaluating text generation with bert (2020), https://arxiv.org/abs/1904.09675

  26. [34]

    CoRRabs/2104.00332 (2021), https://arxiv.org/abs/2104.00332 10 A

    Zhou, M., Zhou, L., Wang, S., Cheng, Y., Li, L., Yu, Z., Liu, J.: UC2: Universal Cross- lingual Cross-modal Vision-and-language Pre-training. CoRRabs/2104.00332 (2021), https://arxiv.org/abs/2104.00332 10 A. Manea and J. Libovický A Caption Classifier To train the classifier, ...

  27. [120]

    https://doi.org/10.1007/978-3-030-58577-8\_7

    Springer (2020). https://doi.org/10.1007/978-3-030-58577-8\_7

  28. [2023]

    10637–10647

    pp. 10637–10647. AAAI Press, Washington, DC, USA (Feb 2023), https://ojs.aaai.org/index.php/AAAI/article/view/26263

  29. [4186]

    https://doi.org/10.18653/v1/N19- 1423

    ACL, Minneapolis, Minnesota (Jun 2019). https://doi.org/10.18653/v1/N19- 1423

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.