Pith. sign in

REVIEW 3 major objections 4 minor 44 references

ConECT Dataset: Overcoming Data Scarcity in Context-Aware E-Commerce MT

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding product images or category paths improves Czech-to-Polish e-commerce translation

desk verdict The ConECT dataset is a genuinely useful public resource for cs-pl e-commerce MT, but the context-improves-translation claim is not supported by the evidence as presented. read the letter →

arxiv 2506.04929 v2 pith:F7XBCM7L submitted 2025-06-05 cs.CL

Miko{l}aj Pokrywka , Wojciech Kusa , Mieszko Rutkowski , Miko{l}aj Koszowski This is my paper
classification cs.CL
keywords machinetranslationcontext-awaremultimodale-commerceCzech-Polishdatasetvision-languagemodelscategorycontext
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 introduces ConECT, a manually translated Czech-to-Polish dataset of 11,400 e-commerce sentences, each paired with the product's primary image and category path. It tests three ways to give a translation model context beyond the sentence itself: fine-tuning a vision-language model on real product images, prefixing text-to-text NMT models with the product's category path, and prefixing them with synthetic image descriptions. The paper's central claim is that the first two forms of context improve translation quality over text-only baselines, while the third degrades it. Per-sentence models routinely mistranslate words whose meaning depends on what the product is or looks like, and e-commerce platforms already hold the image and metadata signals these methods exploit.

What carries the argument

The load-bearing mechanism is the conditional context injection into two model families. For the vision-language route, the product image is fed alongside the source text into a fine-tuned 3-billion-parameter vision-language model, so the decoder can condition on visual features during generation. For the text-only route, the product's category path (a hierarchical string such as 'Sports » Bicycles » Tires') is prepended to the source sentence between special tokens, letting a standard Transformer encoder condition on the category hierarchy before decoding. Both are compared against their context-free controls—a blank image for the VLM and no prefix for the NMT—so the measured gain isolates the effect of the added context rather than the fine-tuning itself. The dataset, ConECT, is the enabling object: it provides the aligned image-metadata-text triples that make these comparisons possible for an under-resourced language pair.

What would settle it

Recompute the main comparisons on a test set consisting of product listings originally written in Czech (not translated from Polish); if the image- and category-prefixed models no longer beat their context-free controls, the paper's central claim fails to generalize beyond its own dataset. A cheaper check is to measure whether the image condition's gain concentrates on sentences whose ambiguous noun is visually resolvable; if the gain is uniform across unambiguous sentences, the apparent context benefit would instead reflect a domain shift.

Watch

Extended reading notes

Core claim

The central claim is that incorporating contextual information—specifically product images and category paths—improves Czech-to-Polish machine translation in the e-commerce domain. The authors construct ConECT, a test set of 11,400 Polish-to-Czech sentence pairs sampled from marketplace listings, with images and category paths attached, and manually translated into Czech. On this test set, a vision-language model fine-tuned with the true product image outperforms the same model given a blank image, and a text-to-text NMT model fine-tuned with a category-path prefix achieves higher COMET scores than the same model fine-tuned without the prefix. The authors interpret these results as evidence that both visual and structured metadata context help disambiguate translation, while noting that synthetic image-description prefixes hurt performance and that context does not help in every case.

Load-bearing premise

The load-bearing premise is that the ConECT test set measures natural Czech-to-Polish e-commerce translation, yet its Czech sentences were themselves translated from Polish text rather than written natively, so the reported gains may not transfer to naturally written Czech listings.

Editorial extensions

If this is right

  • E-commerce platforms with product images and category metadata can improve Czech-to-Polish MT by feeding those signals to the model without changing the underlying architecture.
  • Fine-tuning a VLM on real product images is a viable path to context-aware MT, with gains concentrated in product names and offer titles, while product descriptions show mixed metric movement.
  • Category-path prefixes are a cheap, effective way to add context to an existing text-to-text NMT model: the method needs no vision components and improves COMET on all test splits.
  • Synthetic image descriptions generated by a VLM are not a safe drop-in proxy for real images: prefixing them hurts translation quality, so the visual signal must come from the image itself or from better description methods.
  • The public release of ConECT gives the field a benchmark for context-aware MT on a low-resource language pair where none previously existed.

Reading between the lines

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

  • Because the Czech source sentences in the test set were themselves produced by translating Polish text, the reported gains may be optimistic for naturally written Czech listings; a test set built from original Czech copy would show whether the context advantage survives translationese-free input.
  • The category-path gains suggest a general recipe: any structured metadata with high mutual information about word sense—such as price range, brand, or product attributes—could be injected the same way and would be worth testing against the category path.
  • A falsifiable extension of the VLM result is to check whether the image helps most on sentences that contain an ambiguous noun whose sense the image disambiguates; if the gains are spread uniformly across unambiguous sentences, the mechanism would reflect domain adaptation more than disambiguation.
  • The negative result with image descriptions implies that the bottleneck is how the description is generated or injected, not the information content per se; a contrastive study comparing different captioning models and prefix formats could separate those factors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces ConECT, a Czech-to-Polish e-commerce translation dataset of about 11,400 product-name/offer-title/description sentence pairs with images and category-path metadata. The authors compare three context-aware approaches: fine-tuning a PaliGemma vision-language model with real product images versus black images, fine-tuning a Marian text-to-text baseline with category-path prefixes, and fine-tuning with generated image-description prefixes. They report that real images improve VLM translation, category-path prefixes yield a small COMET gain, and image-description prefixes hurt performance. The dataset is publicly released.

Significance. If the claims hold, the dataset and the VLM fine-tuning result would be a useful contribution for an under-resourced language pair and for e-commerce MT. The paper is an empirical study with a clear setup, standard automatic metrics, and an openly released resource; it does not rely on any formal derivation or circular fitting loop. However, the significance is currently limited by two load-bearing validity concerns: the test set is constructed by translating Polish reference sentences into Czech, and the category-path context is provided in the target language. These issues directly affect the central claim that contextual information improves translation.

major comments (3)
  1. [Section 3 and Section 4.1] The ConECT test set is created by extracting Polish sentences from allegro.pl and manually translating them into Czech, so every Czech source in the cs->pl test is a translation of the Polish reference. This turns the evaluation task into reconstructing the original Polish from its Czech translation rather than translating naturally written Czech e-commerce text. Section 4.1 compounds the issue: the image-conditioned training data uses 440,000 Polish product names back-translated into Czech, and the category-path fine-tuning data uses 7M+7M back-translated product names and descriptions. Consequently, both training and evaluation for the contextual models are specific to a synthetic Czech-from-Polish register, and the available images and Polish category paths may provide reconstruction cues that would not be present for genuinely Czech listings. The Limitations section acknowledges that text alone is often sufficient but never flags this reverse-translation construction. Please provide an evaluation on naturally written Czech e-commerce text, or otherwise demonstrate that the round-trip construction does not drive the reported context gains (e.g., through a human evaluation on native Czech listings).
  2. [Section 4.2 and Appendix A.3] The category-path context is provided in the target language, Polish, as explicitly stated in Appendix A.3: 'The subcategories were provided in the target language.' The example source sentence shows the Czech source prefixed with the Polish tokens 'Moda', 'Obuwie', 'Męskie', 'Sportowe'. This is therefore not a source-side or neutral contextual signal; it supplies target-language lexical information that can directly hint at the translation, potentially inflating the apparent benefit of 'category context'. To support the general claim that contextual information improves MT, the authors should either ablate with Czech (source-language) category paths or otherwise control for the overlap between the prefix vocabulary and the target reference. Without such a control, the reported COMET gain for category context cannot be attributed to contextual disambiguation rather than target-language leakage.
  3. [Section 5, Table 2] The reported advantage of the category-context model over its no-context fine-tuned control is small and inconsistent: on 'All sets', chrF is 78.56 for category context versus 78.87 without it, while COMET is 0.9362 versus 0.9354. No significance tests, confidence intervals, or bootstrap estimates are reported, and the differences are of the same order as typical COMET noise. The PaliGemma real-image versus black-image comparison likewise lacks significance testing. Given that the training and evaluation data share the same back-translation construction, these small automatic-metric gaps are not robust evidence for the central claim. Please report paired significance tests and confidence intervals for all comparisons, and ideally complement the automatic metrics with a human evaluation on the context-sensitive subsets.
minor comments (4)
  1. [Abstract and Section 3] The abstract states the dataset consists of 11,400 sentence pairs, Section 3 says 11,000 sentences were extracted, and Table 1 sums to 10,295 test plus 1,111 validation, i.e., 11,406. Please reconcile these numbers.
  2. [Table 2] The header 'Category paths experiements' contains a typo; it should read 'experiments'.
  3. [Section 4.2 and Appendix A.4] The image-description experiment uses a smaller training set than the category-context experiment and is fine-tuned only on synthetic descriptions. The paper should state this asymmetry explicitly in the results discussion, since it weakens the conclusion that image-description prefixes inherently degrade quality.
  4. [Section 4.2] The NLLB-600M baseline is listed without a citation; please add a reference to the NLLB model paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the context-gain claims rest on held-out test-set evaluations, not on fitted parameters or self-citations that reduce to the inputs.

full rationale

The paper is an empirical study with no derivation chain in which a prediction is constructed from its own inputs. The ConECT test set is created by extracting 11,000 Polish sentences from allegro.pl and manually translating them into Czech, which may introduce translationese or distribution mismatch, but the models are evaluated on unseen manually translated test items rather than on their training data. The category-path, image, and image-description experiments compare model variants under matched training setups, and the reported improvements are measured by automatic metrics; early stopping on the validation set is standard practice and does not amount to fitting the test labels. The back-translated synthetic training data is a legitimate data-augmentation strategy, and any concern that it inflates context gains is a validity or correctness issue, not circularity. There is no load-bearing self-citation, no imported uniqueness theorem, and no fitted parameter renamed as a prediction. The paper is self-contained as an empirical benchmark contribution, so the appropriate circularity score is 0.

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

The central empirical comparisons rest on standard modeling assumptions: automatic metrics reflect quality, back-translated Czech is acceptable training data, a black image is a valid no-context control, and the manually translated test set is suitable for cs->pl evaluation. The test-set translation direction is the most fragile assumption.

free parameters (5)
  • LoRA rank r = 8
    Chosen hyperparameter for PaliGemma fine-tuning; not fitted to test data.
  • LoRA alpha = 8
    Chosen hyperparameter; hand-picked.
  • VLM learning rate = 1e-4
    Chosen hyperparameter for PaliGemma fine-tuning.
  • NMT fine-tuning learning rate = 5e-6
    Chosen hyperparameter for category and image-description fine-tuning.
  • VLM fine-tuning epochs = 4
    Chosen number of epochs for PaliGemma fine-tuning.
assumptions (5)
  • domain assumption The test set is a valid sample of Czech e-commerce product text for cs->pl translation.
    The test set was created by translating Polish product text into Czech, so the Czech source may exhibit translationese; the paper does not discuss this.
  • domain assumption Back-translated Czech product names are valid training data for cs->pl MT.
    Synthetic data from MT back-translation may contain errors; no quality check is reported.
  • domain assumption chrF and COMET capture translation quality relevant to e-commerce context.
    These are standard automatic metrics; no human evaluation is performed.
  • domain assumption A black image is an appropriate control for testing visual context influence.
    The paper uses a black image as unrelated context; a different natural image might behave differently.
  • domain assumption The category path in the target language provides useful disambiguating context.
    Subcategories are included in Polish while the source is Czech; this is a deliberate design but its benefit is not theoretically motivated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ConECT Dataset: Overcoming Data Scarcity in Context-Aware E-Commerce MT." pith.science (2026). https://pith.science/paper/F7XBCM7L

@misc{pith2026250604929,
  author       = {Pith},
  title        = {Pith review of: ConECT Dataset: Overcoming Data Scarcity in Context-Aware E-Commerce MT},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F7XBCM7L}},
  note         = {Machine review of arXiv:2506.04929}
}
read the original abstract

Neural Machine Translation (NMT) has improved translation by using Transformer-based models, but it still struggles with word ambiguity and context. This problem is especially important in domain-specific applications, which often have problems with unclear sentences or poor data quality. Our research explores how adding information to models can improve translations in the context of e-commerce data. To this end we create ConECT -- a new Czech-to-Polish e-commerce product translation dataset coupled with images and product metadata consisting of 11,400 sentence pairs. We then investigate and compare different methods that are applicable to context-aware translation. We test a vision-language model (VLM), finding that visual context aids translation quality. Additionally, we explore the incorporation of contextual information into text-to-text models, such as the product's category path or image descriptions. The results of our study demonstrate that the incorporation of contextual information leads to an improvement in the quality of machine translation. We make the new dataset publicly available.

Figures

Figures reproduced from arXiv: 2506.04929 by the authors.

Figure 1
Figure 1. We evaluate three methods for contextualisa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

44 extracted references · 16 canonical work pages

  1. [1]

    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 INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. GPT-4 technical report . arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. 2015. https://arxiv.org/abs/1409.0473 Neural machine translation by jointly learning to align and translate . In International Conference on Learning Representations (ICLR)

    arXiv 2015

  5. [5]

    Lo \" c Barrault, Fethi Bougares, Lucia Specia, Chiraag Lala, Desmond Elliott, and Stella Frank. 2018. https://doi.org/10.18653/v1/W18-6402 Findings of the third shared task on multimodal machine translation . In Proceedings of the Third Conference on Machine Translation: Shared Task Papers, pages 304--323, Belgium, Brussels. Association for Computational...

    doi:10.18653/v1/w18-6402 2018

  6. [6]

    Lucas Beyer, Andreas Steiner, André Susano Pinto, Alexander Kolesnikov, Xiao Wang, Daniel Salz, Maxim Neumann, Ibrahim Alabdulmohsin, Michael Tschannen, Emanuele Bugliarello, Thomas Unterthiner, Daniel Keysers, Skanda Koppula, Fangyu Liu, Adam Grycner, Alexey Gritsenko, Neil Houlsby, Manoj Kumar, Keran Rong, Julian Eisenschlos, Rishabh Kabra, Matthias Bau...

    arXiv 2024

  7. [7]

    Ozan Caglayan, Lo \" c Barrault, and Fethi Bougares. 2016. Multimodal attention for neural machine translation. arXiv preprint arXiv:1609.03976

    arXiv 2016

  8. [8]

    Iacer Calixto and Qun Liu. 2017. https://doi.org/10.18653/v1/D17-1105 Incorporating global visual features into attention-based neural machine translation. In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 992--1003, Copenhagen, Denmark. Association for Computational Linguistics

    work page doi:10.18653/v1/d17-1105 2017

Show all 44 references
  1. [9]

    Iacer Calixto, Qun Liu, and Nick Campbell. 2017 a . https://doi.org/10.18653/v1/P17-1175 Doubly-attentive decoder for multi-modal neural machine translation . In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pa...

    2017 doi

  2. [10]

    Iacer Calixto, Daniel Stein, Evgeny Matusov, Sheila Castilho, and Andy Way. 2017 b . https://doi.org/10.18653/v1/W17-2004 Human evaluation of multi-modal neural machine translation: A case-study on E -commerce listing titles . In Proceedings of the Sixth Workshop on Vision and...

    2017 doi

  3. [11]

    Iacer Calixto, Daniel Stein, Evgeny Matusov, Pintu Lohar, Sheila Castilho, and Andy Way. 2017 c . https://aclanthology.org/E17-2101 Using images to improve machine-translating E -commerce product listings. In Proceedings of the 15th Conference of the E uropean Chapter of the A...

    2017

  4. [12]

    Shizhe Chen, Qin Jin, and Jianlong Fu. 2019. From words to sentences: A progressive learning approach for zero-resource machine translation with visual pivots. arXiv preprint arXiv:1906.00872

    2019 arXiv

  5. [13]

    Desmond Elliott. 2018. https://doi.org/10.18653/v1/D18-1329 Adversarial evaluation of multimodal machine translation . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, pages 2974--2978, Brussels, Belgium. Association for Computational ...

    2018 doi

  6. [14]

    Desmond Elliott, Stella Frank, Lo \" c Barrault, Fethi Bougares, and Lucia Specia. 2017. https://doi.org/10.18653/v1/W17-4718 Findings of the second shared task on multimodal machine translation and multilingual image description . In Proceedings of the Second Conference on Ma...

    2017 doi

  7. [15]

    Desmond Elliott, Stella Frank, Khalil Sima ' an, and Lucia Specia. 2016. https://doi.org/10.18653/v1/W16-3210 M ulti30 K : Multilingual E nglish- G erman image descriptions . In Proceedings of the 5th Workshop on Vision and Language, pages 70--74, Berlin, Germany. Association ...

    2016 doi

  8. [16]

    Fangxiaoyu Feng, Yinfei Yang, Daniel Cer, Naveen Arivazhagan, and Wei Wang. 2022. https://arxiv.org/abs/2007.01852 Language-agnostic bert sentence embedding . Preprint, arXiv:2007.01852

    2022 arXiv

  9. [17]

    Matthieu Futeral, Cordelia Schmid, Ivan Laptev, Beno \^ t Sagot, and Rachel Bawden. 2023. https://doi.org/10.18653/v1/2023.acl-long.295 Tackling ambiguity with images: Improved multimodal machine translation and contrastive evaluation . In Proceedings of the 61st Annual Meetin...

    2023 doi

  10. [18]

    Matthieu Futeral, Cordelia Schmid, Beno \^ t Sagot, and Rachel Bawden. 2024. Towards zero-shot multimodal machine translation. arXiv preprint arXiv:2407.13579

    2024 arXiv

  11. [19]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. https://arxiv.org/abs/2106.09685 Lora: Low-rank adaptation of large language models . Preprint, arXiv:2106.09685

    2021 arXiv

  12. [20]

    Marcin Junczys-Dowmunt, Roman Grundkiewicz, Tomasz Dwojak, Hieu Hoang, Kenneth Heafield, Tom Neckermann, Frank Seide, Ulrich Germann, Alham Fikri Aji, Nikolay Bogoychev, Andr\' e F. T. Martins, and Alexandra Birch. 2018. http://www.aclweb.org/anthology/P18-4020 Marian: Fast ne...

    2018

  13. [21]

    Taku Kudo and John Richardson. 2018. https://doi.org/10.18653/v1/D18-2012 S entence P iece: A simple and language independent subword tokenizer and detokenizer for neural text processing . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processin...

    2018 doi

  14. [22]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730--19742. PMLR

    2023

  15. [23]

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2024. Visual instruction tuning. Advances in neural information processing systems, 36

    2024

  16. [24]

    Pengbo Liu, Hailong Cao, and Tiejun Zhao. 2021. Gumbel-attention for multi-modal machine translation. arXiv preprint arXiv:2103.08862

    2021 arXiv

  17. [25]

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 3...

    2022

  18. [26]

    Maja Popovi \'c . 2015. https://doi.org/10.18653/v1/W15-3049 chr F : character n-gram F -score for automatic MT evaluation . In Proceedings of the Tenth Workshop on Statistical Machine Translation, pages 392--395, Lisbon, Portugal. Association for Computational Linguistics

    2015 doi

  19. [27]

    Matt Post. 2018. https://doi.org/10.18653/v1/W18-6319 A call for clarity in reporting BLEU scores . In Proceedings of the Third Conference on Machine Translation: Research Papers, pages 186--191, Brussels, Belgium. Association for Computational Linguistics

    2018 doi

  20. [28]

    Ricardo Rei, Jos \'e G. C. de Souza, Duarte Alves, Chrysoula Zerva, Ana C Farinha, Taisiya Glushkova, Alon Lavie, Luisa Coheur, and Andr \'e F. T. Martins. 2022. https://aclanthology.org/2022.wmt-1.52 COMET -22: Unbabel- IST 2022 submission for the metrics shared task . In Pro...

    2022

  21. [29]

    Annette Rios Gonzales, Laura Mascarell, and Rico Sennrich. 2017. https://doi.org/10.18653/v1/W17-4702 Improving word sense disambiguation in neural machine translation with sense embeddings . In Proceedings of the Second Conference on Machine Translation, pages 11--19, Copenha...

    2017 doi

  22. [30]

    Huangjun Shen, Liangying Shao, Wenbo Li, Zhibin Lan, Zhanyu Liu, and Jinsong Su. 2024. A survey on multi-modal machine translation: Tasks, methods and challenges. arXiv preprint arXiv:2405.12669

    2024 arXiv

  23. [31]

    Gunnar A Sigurdsson, Jean-Baptiste Alayrac, Aida Nematzadeh, Lucas Smaira, Mateusz Malinowski, Joao Carreira, Phil Blunsom, and Andrew Zisserman. 2020. Visual grounding in video for unsupervised word translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and...

    2020

  24. [32]

    Yuqing Song, Shizhe Chen, Qin Jin, Wei Luo, Jun Xie, and Fei Huang. 2021. Product-oriented machine translation with cross-modal cross-lingual pre-training. In Proceedings of the 29th ACM International Conference on Multimedia, pages 2843--2852

    2021

  25. [33]

    Lucia Specia, Stella Frank, Khalil Sima ' an, and Desmond Elliott. 2016. https://doi.org/10.18653/v1/W16-2346 A shared task on multimodal machine translation and crosslingual image description . In Proceedings of the First Conference on Machine Translation: Volume 2, Shared Ta...

    2016 doi

  26. [34]

    Yuanhang Su, Kai Fan, Nguyen Bach, C-C Jay Kuo, and Fei Huang. 2019. Unsupervised multi-modal neural machine translation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10482--10491

    2019

  27. [35]

    J \"o rg Tiedemann and Lars Nygaard. 2004. http://www.lrec-conf.org/proceedings/lrec2004/pdf/320.pdf The OPUS corpus - parallel and free: http://logos.uio.no/opus . In Proceedings of the Fourth International Conference on Language Resources and Evaluation ( LREC ' 04) , Lisbon...

    2004

  28. [36]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. LLaMA: Open and efficient foundation language models . arXiv preprint arXiv:2302.13971

    2023 arXiv

  29. [37]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017 a . https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in...

    2017

  30. [38]

    Gomez, ukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, ukasz Kaiser, and Illia Polosukhin. 2017 b . Attention is all you need. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, page 6000–601...

    2017

  31. [39]

    Dexin Wang and Deyi Xiong. 2021. Efficient object-level visual context modeling for multimodal machine translation: Masking irrelevant objects helps grounding. In Proceedings of the AAAI conference on artificial intelligence, volume 35, pages 2720--2728

    2021

  32. [40]

    Haoran Xu, Young Jin Kim, Amr Sharaf, and Hany Hassan Awadalla. 2023. A paradigm shift in machine translation: Boosting translation performance of large language models. arXiv preprint arXiv:2309.11674

    2023 arXiv

  33. [41]

    Pengcheng Yang, Boxing Chen, Pei Zhang, and Xu Sun. 2020. Visual agreement regularized training for multi-modal machine translation. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages 9418--9425

    2020

  34. [42]

    Shaowei Yao and Xiaojun Wan. 2020. https://doi.org/10.18653/v1/2020.acl-main.400 Multimodal transformer for multimodal machine translation . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4346--4350, Online. Association for Co...

    2020 doi

  35. [43]

    Yongjing Yin, Fandong Meng, Jinsong Su, Chulun Zhou, Zhengyuan Yang, Jie Zhou, and Jiebo Luo. 2020. https://doi.org/10.18653/v1/2020.acl-main.273 A novel graph-based multi-modal fusion encoder for neural machine translation . In Proceedings of the 58th Annual Meeting of the As...

    2020 doi

  36. [44]

    Vilém Zouhar, Pinzhen Chen, Tsz Kin Lam, Nikita Moghe, and Barry Haddow. 2024. https://arxiv.org/abs/2408.15366 Pitfalls and outlooks in using comet . Preprint, arXiv:2408.15366

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.