Elamur Research⌕
☰
home / papers / 2026 / Seeing Words Differently: Visual Embeddings for Robust English-Arabic Machine Translation

Seeing Words Differently: Visual Embeddings for Robust English-Arabic Machine Translation

Paper · 2026 · Mahdi Alshaikh Saleh and Irfan Ahmad

Summary

The paper examines whether rendering source text as images helps English-to-Arabic neural MT cope with misspellings and out-of-vocabulary words, which tend to break subword-based systems. The authors train transformer encoder-decoder models on the IWSLT 2017 English-Arabic corpus (TED talk transcripts). A text baseline using BPE embeddings is compared with several visual-input variants. In these, English sentences are drawn as grayscale images, sliced into visual tokens, and projected by a CNN into embeddings. The variants are a sliding-window renderer (BASE1, BASE2), the same renderer applied after stripping English vowels (NON-VOWELS), and a character-trigram renderer with non-overlapping slices (TRI-GRAM). All systems are scored with BLEU on clean input and on artificially corrupted input. On clean text, the BPE baseline remains the strongest. As character noise increases, however, the visual systems lose quality far more slowly. The trigram variant offers the best trade-off between quality and training cost. The study frames visual input as a deliberate choice for noisy or user-generated text, accepting some loss of benchmark accuracy in exchange for stability.

Key points

  • Data: IWSLT 2017 English-Arabic, with 231,713 training pairs, 888 validation pairs and 8,583 test pairs.
  • Text baseline: joint BPE with 10,000 merges.
  • Visual models: NotoSans-Regular at 10pt. Sliding windows are 15 px wide with a 10 px stride, and each token gets a 512-dimensional CNN embedding.
  • Robustness: at 50% character corruption, visual models keep up to 2.4× the BLEU of BPE models. Noise was tested from 10% to 50%.
  • TRI-GRAM efficiency: it cuts input length from roughly 74 to roughly 32 visual tokens per sentence, which speeds up training.
  • Target side: BPE on the Arabic output proved important. BASE1 used word-level Arabic tokens and generalized worse than the BPE-decoding models.

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

 Institutions

🎓King Fahd University of Petroleum and MineralsDhahran · Saudi Arabia54 works · 4,666 citations

 Fields, topics and concepts

📚 Shelves 4The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksArts and Humanities · 17,450 worksLanguage and Linguistics · 17,296 works

🧭 Topics 3Research subjects and languages the work is about.Machine translation · 1,105 worksEnglish · 2,379 worksSemitic languages · 3,066 works

Cite

Cite as Mahdi Alshaikh Saleh and Irfan Ahmad. 2026. Seeing Words Differently: Visual Embeddings for Robust English-Arabic Machine Translation. Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.abjadnlp-1.9

@inproceedings{saleh2026seeing,
  title = {{Seeing Words Differently: Visual Embeddings for Robust English-Arabic Machine Translation}},
  author = {Mahdi Alshaikh Saleh and Irfan Ahmad},
  booktitle = {Proceedings of the 2nd Workshop on NLP for Languages Using Arabic Script},
  year = {2026},
  doi = {10.18653/v1/2026.abjadnlp-1.9},
  url = {https://doi.org/10.18653/v1/2026.abjadnlp-1.9}
}

References

7 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Maged Al-Shaibani and Irfan Ahmad. 2023. Consonant is all you need: a compact representation of English text for efficient NLP. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 11578–11588. Association for Computational Linguistics.
  2. Phillip Rust, Jonas F. Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Elliott. 2023. Language modelling with pixels. In International Conference on Learning Representations. find
  3. Elizabeth Salesky, David Etter, and Matt Post. 2021. Robust open-vocabulary translation from visual text representations. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7235–7252, Online and Punta Cana, Dominican Republic. Association for Computation… find
  4. Dorieh Alomari and Irfan Ahmad. 2024. Exploring character trigrams for robust arabic text classification: A comparative analysis in the face of vocabulary expansion and misspelled words. IEEE Access, 12:57103–57116. find
  5. Elizabeth Salesky, Neha Verma, Philipp Koehn, and Matt Post. 2023. Multilingual pixel representations for translation and effective cross-lingual transfer. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 13845–13861, Singapore. Association for Computa… find
  6. Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Neural machine translation of rare words with subword units. In Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 1715–1725.
  7. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. In Advances in Neural Information Processing Systems, volume 30.

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/saleh-2026-seeing-words-differently-visual-embeddings/