Elamur Research⌕
☰
home / papers / 2026 / FFT-UniBa at Cruciverb-IT: Special Length Tokens and CSP for Italian Crossword Solving

FFT-UniBa at Cruciverb-IT: Special Length Tokens and CSP for Italian Crossword Solving

Paper · 2026 · Andrea Porcelli, Filippo Di Gravina, Emanuele Fontana et al.

Summary

This paper describes the FFT-UniBa entry to Cruciverb-IT, an EVALITA 2026 shared task on Italian crosswords. The task has two parts: producing ranked answers for single clues of known length, and filling entire grids. The authors argue that subword tokenization hides character counts, so language models struggle to meet strict length constraints. For clue answering, they fine-tuned the Italian T5 variant it5-efficient-small-el32. They added length-specific start and end tokens ([SL=L], [EL=L]) to the input and the target, so the model learns during training where a word of a given length should stop. This differs from forcing length only at decoding time. For grid filling, they treated the puzzle as a CSP. Each slot's domain holds the model's top candidates, ranked by perplexity, plus a small dictionary fallback. A seeded depth-first backtracking search solves the grid, with budgets that scale with grid size. Both components clearly beat the task baselines. The work shows that explicit length conditioning helps compact seq2seq models handle the meta-linguistic side of word games.

Key points

  • Sub-task 1 results: up to 0.58 Acc@1, 0.75 Acc@10 and 0.63 MRR. The best configuration was trained for 22 epochs and decoded with 1000 beams.
  • Sub-task 2 results: on the test set, the system reached at most 0.93 accuracy at the character level, 0.85 at the word level and 0.34 for complete grids.
  • Solver budgets: the number of seeds, model candidates (K) and dictionary candidates (K′) all grow with grid size. K′ is kept much smaller than K so that the system remains neural rather than a dictionary lookup.
  • Candidate-count ablation:
    • Grid performance mostly levels off after about 200 candidates per clue.
    • Dictionary fallbacks still give a steady gain on the hardest crossings.
  • Clue-type analysis:
    • The authors labelled test clues with Gemini 3 Flash into four types: definition, wordplay, fill-in-the-blank and cryptic.
    • Adding dictionary training data gave only marginal gains (about +0.01 MRR) and hurt cryptic clues.
    • The authors conclude that reasoning, not lexical coverage, is the main bottleneck.

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 2The field of science this work belongs to (OpenAlex fields).Computer Science · 16,667 worksArtificial Intelligence · 16,194 works

Cite

Cite as Andrea Porcelli, Filippo Di Gravina, Emanuele Fontana, Mattia Curri and Francesco Damiano Di Gregorio. 2026. FFT-UniBa at Cruciverb-IT: Special Length Tokens and CSP for Italian Crossword Solving. Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026).

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{porcelli2026fft,
  title = {{FFT-UniBa at Cruciverb-IT: Special Length Tokens and CSP for Italian Crossword Solving}},
  author = {Andrea Porcelli, Filippo {Di Gravina}, Emanuele Fontana, Mattia Curri and Francesco Damiano {Di Gregorio}},
  booktitle = {Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026)},
  year = {2026},
  url = {https://aclanthology.org/2026.evalita-1.38/}
}

References

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

  1. C. Ciaccio, G. Sarti, A. Miaschi, F. Dell’Orletta, M. Nissim, Cruciverb-it @ evalita 2026: Overview of the crossword solving in italian task, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for Italian. Final Workshop (EVALITA 2026), CEUR.org, Bari, I… find
  2. F. Cutugno, A. Miaschi, A. P. Aprosio, G. Rambelli, L. Siciliani, M. A. Stranisci, Evalita 2026: Overview of the 9th evaluation campaign of natural language processing and speech tools for italian, in: Proceedings of the Ninth Evaluation Campaign of Natural Language Processing and Speech Tools for I… find
  3. G. Barlacchi, M. Nicosia, A. Moschitti, Sacry: Syntax-based automatic crossword puzzle resolution system, in: Proceedings of 53nd Annual Meeting of the Association for Computational Linguistics: System Demonstrations, Beijing, China, July. Association for Computational Linguistics, 2015. find
  4. C. Ciaccio, G. Sarti, A. Miaschi, F. Dell’Orletta, Crossword space: Latent manifold learning for italian crosswords and beyond, in: Proceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025), 2025, pp. 245–255. find
  5. G. Sarti, T. Caselli, A. Bisazza, M. Nissim, EurekaRebus - verbalized rebus solving with LLMs: A CALAMITA challenge, in: F. Dell’Orletta, A. Lenci, S. Montemagni, R. Sprugnoli (Eds.), Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), CEUR Workshop Proceedings,… find
  6. G. Sarti, T. Caselli, M. Nissim, A. Bisazza, Non verbis, sed rebus: Large language models are weak solvers of Italian rebuses, in: F. Dell’Orletta, A. Lenci, S. Montemagni, R. Sprugnoli (Eds.), Proceedings of the Tenth Italian Conference on Computational Linguistics (CLiC-it 2024), CEUR Workshop Pro… find
  7. E. Mensa, L. Zane, C. J. Scozzaro, M. Delsanto, T. Milani, D. P. Radicioni, Easy to complete, hard to choose: Investigating llm performance on the proverbit benchmark, in: Proceedings of the Eleventh Italian Conference on Computational Linguistics (CLiC-it 2025), 2025, pp. 722–734. find
  8. M. Ernandes, G. Angelini, M. Gori, et al., Webcrow: A web-based system for crossword solving, in: AAAI, 2005, pp. 1412–1417. find
  9. E. Wallace, N. Tomlin, A. Xu, K. Yang, E. Pathak, M. Ginsberg, D. Klein, Automated crossword solving, arXiv preprint arXiv:2205.09665 (2022). — [arXiv:2205.09665](https://arxiv.org/abs/2205.09665) find
  10. J. Rozner, C. Potts, K. Mahowald, Decrypting cryptic crosswords: Semantically complex wordplay puzzles as a target for nlp, Advances in Neural Information Processing Systems 34 (2021) 11409– 11421. find
  11. A. Sadallah, D. Kotova, E. Kochmar, What makes cryptic crosswords challenging for llms?, in: Proceedings of the 31st International Conference on Computational Linguistics, 2025, pp. 5102– 5114. find
  12. X. Zhang, J. Cao, C. You, Counting ability of large language models and impact of tokenization, 2024. URL: https://arxiv.org/abs/2410.19730. arXiv:2410.19730 . ↗
  13. S. Saha, S. Chakraborty, S. Saha, U. Garain, Language models are crossword solvers, in: Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), 2025, pp. 2074–2090. find
  14. N. Shazeer, M. Littman, G. Keim, Solving crossword puzzles as probabilistic constraint satisfaction, Proc. AAAI ’99 (1999). find
  15. S. Kulshreshtha, O. Kovaleva, N. Shivagunde, A. Rumshisky, Down and across: Introducing crossword-solving as a new NLP benchmark, in: S. Muresan, P. Nakov, A. Villavicencio (Eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Associ… find
  16. M. L. Ginsberg, Dr. fill: Crosswords and an implemented solver for singly weighted csps, Journal of Artificial Intelligence Research 42 (2011) 851–886. find
  17. G. Barlacchi, M. Nicosia, A. Moschitti, Learning to rank answer candidates for automatic resolution of crossword puzzles, in: Proceedings of the Eighteenth Conference on Computational Natural Language Learning, 2014, pp. 39–48. find
  18. A. Severyn, M. Nicosia, G. Barlacchi, A. Moschitti, Distributional neural networks for automatic resolution of crossword puzzles, in: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Vol… find
  19. G. Sarti, M. Nissim, IT5: Text-to-text pretraining for Italian language understanding and generation, in: N. Calzolari, M.-Y. Kan, V. Hoste, A. Lenci, S. Sakti, N. Xue (Eds.), Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC… find
  20. C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, P. J. Liu, Exploring the limits of transfer learning with a unified text-to-text transformer, Journal of machine learning research 21 (2020) 1–67.
  21. Y. Tay, M. Dehghani, J. Rao, W. Fedus, S. Abnar, H. W. Chung, S. Narang, D. Yogatama, A. Vaswani, D. Metzler, Scale efficiently: Insights from pre-training and fine-tuning transformers, arXiv preprint arXiv:2109.10686 (2021). find
  22. K. Zeinalipour, T. Iaquinta, A. Zanollo, G. Angelini, L. Rigutini, M. Maggini, M. Gori, Italian crossword generator: Enhancing education through interactive word puzzles (2023). find
  23. L. De Moura, N. Bjørner, Z3: An efficient smt solver, in: International conference on Tools and Algorithms for the Construction and Analysis of Systems, Springer, 2008, pp. 337–340. find
  24. L. Xue, A. Barua, N. Constant, R. Al-Rfou, S. Narang, M. Kale, A. Roberts, C. Raffel, Byt5: Towards a token-free future with pre-trained byte-to-byte models, Transactions of the Association for Computational Linguistics 10 (2022) 291–306. find
  25. Y. Tay, V. Q. Tran, S. Ruder, J. Gupta, H. W. Chung, D. Bahri, Z. Qin, S. Baumgartner, C. Yu, D. Metzler, Charformer: Fast character transformers via gradient-based subword tokenization, arXiv preprint arXiv:2106.12672 (2021).
  26. J. Wang, T. Gangavarapu, J. N. Yan, A. M. Rush, Mambabyte: Token-free selective state space model, arXiv preprint arXiv:2401.13660 (2024).

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/porcelli-2026-fft-uniba-cruciverb-it-special/