Elamur Research⌕
☰
home / papers / 2026 / PictoEduca: Building a Dataset for Spanish Text-to-Pictogram Generation

PictoEduca: Building a Dataset for Spanish Text-to-Pictogram Generation

Paper · 2026 · Alfonso Manuel Paredes Umeres and Marco Antonio Sobrevilla Cabezudo

Summary

Spanish has lacked a sizable parallel corpus for converting sentences into pictogram sequences, which hinders data-driven work in augmentative and alternative communication (AAC). The authors address this with PictoEduca, a resource built from official Peruvian primary-school textbooks and linked to the ARASAAC pictogram inventory.

Their pipeline has several stages. Sentences are extracted and manually checked, then filtered by readability, and GPT-4 rewrites them into short, child-friendly versions. ARAWORD performs automatic pictogram annotation for training, and expert psychologists label the validation and test sets.

They compare ARAWORD, a fine-tuned T5 Base, and a prompted LLaMA 3.1-8B-Instruct. Each system is tested on direct mapping and on a two-stage route that first produces textual concepts and then maps them to pictograms via CLIP-based selection. ARAWORD proves hard to beat. Neural systems gain from the intermediate concept layer, which reduces ambiguity.

The authors also study data selection. Ranking candidates by domain similarity plus a quality estimate produces cleaner silver data and better low-resource models, with smaller gains as data grows. The work provides a reusable benchmark and a construction recipe adaptable to other languages and domains.

Key points

  • Source material: 18 MINEDU textbooks (Communication; Science and Technology; grades 3–6; 2020–2023; regular and rural modalities).
  • Corpus size at each stage: 27,650 extracted sentences, then 21,962 after cleaning, then 17,626 rated "simple," then 16,319 after GPT-4 simplification.
  • Splits: 1,111 expert-annotated sentences and 835 sentences from ARASAAC materials were combined into 973 validation and 973 test items. The remaining 15,208 auto-annotated sentences form the training set. ARASAAC contains over 13,500 pictograms.
  • Data selection: FDA (domain representativeness) was combined with TQE (an X-CLIP–based quality proxy). Without manual correction, Similarity + Quality outperformed Similarity alone (Precision-1 0.4742 vs 0.3085; chrF++ 20.34 vs 19.03). Partially correcting 250 of 1,000 instances helped only the weaker strategy.
  • Limitations: the dataset covers only Peruvian educational text and has one reference per sentence. Human evaluation was small (50 sentences, 8 experts). Neural models struggle with pictogram IDs not seen during training.

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

 Fields, topics and concepts

📚 Shelves 7The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 worksArts and Humanities · 16,039 worksLanguage and Linguistics · 15,891 worksComputer Vision and Pattern Recognition · 2,869 worksHealth Professions · 69 worksOccupational Therapy · 9 works

Cite

Cite as Alfonso Manuel Paredes Umeres and Marco Antonio Sobrevilla Cabezudo. 2026. PictoEduca: Building a Dataset for Spanish Text-to-Pictogram Generation. Findings of the Association for Computational Linguistics: ACL 2026.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.findings-acl.1738

@inproceedings{umeres2026pictoeduca,
  title = {{PictoEduca: Building a Dataset for Spanish Text-to-Pictogram Generation}},
  author = {Alfonso Manuel Paredes Umeres and Marco Antonio Sobrevilla Cabezudo},
  booktitle = {Findings of the Association for Computational Linguistics: ACL 2026},
  year = {2026},
  doi = {10.18653/v1/2026.findings-acl.1738},
  url = {https://doi.org/10.18653/v1/2026.findings-acl.1738}
}

References

19 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Vladimir Araujo, Maria Mihaela Trusca, Rodrigo Tufiño, and Marie-Francine Moens. 2024. Sequence-to-sequence Spanish pre-trained language models. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 14729–1… find
  2. Susana Bautista, Raquel Hervás, Agustín Hernández-Gil, Carlos Martínez-Díaz, Sergio Pascua, and Pablo Gervás. 2017. Aratraductor: text to pictogram translation using natural language processing techniques. In Proceedings of the XVIII International Conference on Human Computer Interaction, Interacció… find
  3. Ergun Biçici and Deniz Yuret. 2015. Optimizing instance selection for statistical machine translation with feature decay algorithms. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 23(2):339–350. find
  4. L. Cabello, E. Lleida, J. Simon, A. Miguel, and A. Ortega. 2018. Text-to-Pictogram Summarization for Augmentative and Alternative Communication. Procesamiento de Lenguaje Natural, 61:15–22. find
  5. Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, and 542 o… — The Llama 3 Herd of Models
  6. Piyali Karmakar and Manjira Sinha. 2024. Aiding non-verbal communication: A bidirectional language agnostic framework for automating text to AAC generation. In Proceedings of the 21st International Conference on Natural Language Processing (ICON), pages 324–331, AU-KBC Research Centre, Chennai, Indi… find
  7. Avaneesh Koushik, Jithu Morrison, P Mirunalini, and 1 others. 2024. A transformer based approach for text-to-picto generation. In Notebook for the ImageCLEF Lab at CLEF 2024, pages 1656–1661. find
  8. Cécile Macaire, Chloé Dion, Jordan Arrigo, Claire Lemaire, Emmanuelle Esperança-Rodier, Benjamin Lecouteux, and Didier Schwab. 2024. A multimodal French corpus of aligned speech, text, and pictogram sequences for speech-to-pictogram machine translation. In Proceedings of the 2024 Joint International… find
  9. Cécile Macaire, Diandra Fabre, Benjamin Lecouteux, and Didier Schwab. 2025. Overview of the 2025 ImageCLEFtoPicto Task -Investigating the Generation of Pictogram Sequences from Text and Speech Notebook for the ImageCLEF Lab at CLEF 2025. In Springer Lecture Notes in Computer Science (LNCS), Madrid,… find
  10. Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, and Haibin Ling. 2022. Expanding language-image pretrained models fornbsp;general video recognition. In Computer Vision – ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23–27, 2022, Proceed… find
  11. Magali Norré, Vincent Vandeghinste, Pierrette Bouillon, and Thomas François. 2021. Extending a text-to-pictograph system to French and to arasaac. In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1050–1059, Held Online. INCOMA Ltd. find
  12. Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311–318, Philadelphia, Pennsylvania, USA. Association for Computational…
  13. Jayr Pereira, Rodrigo Nogueira, Cleber Zanchettin, and Robson Fidalgo. 2024. Predictive authoring for brazilian portuguese augmentative and alternative communication. Natural Language Processing, 31(2):535–558. find
  14. Jayr Alencar Pereira, David Macêdo, Cleber Zanchettin, Adriano Lorena Inácio de Oliveira, and Robson do Nascimento Fidalgo. 2022. Pictobert: Transformers for next pictogram prediction. Expert Systems with Applications, 202:117231. find
  15. Maja Popović. 2017. chrF++: words helping character n-grams. In Proceedings of the Second Conference on Machine Translation, pages 612–618, Copenhagen, Denmark. Association for Computational Linguistics.
  16. Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of Machine Learning Research, 21(1).
  17. Nils Reimers and Iryna Gurevych. 2019. Sentence-bert: Sentence embeddings using siamese bert-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics.
  18. Didier Schwab, Pauline Trial, Céline Vaschalde, Loïc Vial, Emmanuelle Esperanca-Rodier, and Benjamin Lecouteux. 2020. Providing semantic knowledge to a set of pictograms for people with disabilities: a set of links between WordNet and arasaac: Arasaac-WN. In Proceedings of the Twelfth Language Resou… find
  19. B Rule-based algorithm for automatic text-to-pictogram generation. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/umeres-2026-pictoeduca-building-dataset-spanish-text/