Elamur Research⌕
☰
home / papers / 2025 / Assessing Dutch Syllabification Algorithms and Improving Accuracy by Combining Phonetic and Orthographic Information through Deep Learning

Assessing Dutch Syllabification Algorithms and Improving Accuracy by Combining Phonetic and Orthographic Information through Deep Learning

Paper · 2025 · Gus Lathouwers, Wieke Harmsen, Catia Cucchiarini et al.

Summary

Automatically dividing Dutch words into syllables is hard. Competing principles (sonority, maximum onset, priority for prefixes and suffixes) interact with many exceptions, and earlier Dutch systems were evaluated on different data with incompatible metrics. The authors run four syllabifiers under one shared protocol: the knowledge-based Brandt Corstius system, Liang's pattern method, the Trogkanis-Elkan CRF, and a new BiLSTM-based deep-learning model. They test on three sets: CELEX dictionary words, loanwords, and pseudowords. They then build a second network that reads both a word's spelling and its phonetic transcription. Data-driven methods win in almost every condition, though the hand-coded rules do best on pseudowords. The spelling-only network reaches 99.55% word accuracy. Adding pronunciation raises this to 99.65%, a 28.6% error reduction relative to the previous best CRF result (99.51%). The gains come from words whose spelling is ambiguous but whose pronunciation is clear. This suggests phonetic signals could help other orthographic tasks, and the models could be ported to languages such as English or German.

Key points

  • Four algorithms (Brandt Corstius, Liang, Trogkanis-Elkan CRF, new deep-learning model) are compared on three datasets: CELEX dictionary words, loanwords, and pseudowords.
  • Orthography-only deep model: 99.55% word accuracy, an 8.2% error reduction over Trogkanis and Elkan (2010). The combined orthographic-plus-phonetic model reaches 99.65%, a 28.6% error reduction.
  • Brandt Corstius scored lowest on dictionary words and loanwords but highest on pseudowords. Rankings therefore depend on the test set, and hyphen accuracy can mislead compared with word accuracy.
  • The deep model and the CRF share only about 30% of their word-level errors. In exploratory runs a BiLSTM outperformed transformer variants, which matches earlier findings on Romanian.
  • Limitations:
    • The pseudoword set is small.
    • Phonetic input was unavailable for loanwords and pseudowords, so the combined model was not tested on them.
    • Bouma (2003) and Bartlett et al. (2008) were excluded because their code or documentation was missing.

Interesting ideas

  • I1. Evaluate on several deliberately different test sets (regular dictionary entries, borrowed words, invented words), because a system that tops one set can rank last on another.
  • I2. Feed a second, parallel representation of the same item (pronunciation alongside spelling) so the model can resolve cases where one view is ambiguous, such as long versus short vowels.
  • I3. Report whole-unit accuracy (every boundary in a word correct), not only per-boundary scores, since partial metrics can look good while many items are still wrong.
  • I4. Measure how much two systems' errors overlap; low overlap (here about 30%) shows they fail differently and are promising to combine.
  • I5. For short, locally determined sequence decisions on modest data, a BiLSTM can beat a transformer, so architecture choice should follow the task's locality and data size.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

  • I1 (learn): Add held-out suites of loanword-heavy and nonce-word text to tagger and parser evaluation, so that memorization of familiar vocabulary is separated from real generalization.
  • I2 (tokens): Morphology and lemma guessing for unknown words could use a pronunciation or phoneme channel to break ties, e.g. stress-dependent noun/verb pairs.
  • I3 (tree): Alongside LAS, track the share of sentences with a fully correct tree, which better reflects whether downstream frames will be right.
  • I4 (learn): Compute error overlap between the perceptron parser and gparse to decide where agreement-based silver data is trustworthy and where witness parsers truly complement each other.
  • I5 (tokens): Keep a lightweight character-level BiLSTM as a candidate for tokenizer and morphology boundary decisions instead of defaulting to a transformer.

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

🎓Radboud University NijmegenNijmegen · Netherlands20 works · 1,688 citations

 Fields, topics and concepts

📚 Shelves 6The 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 worksPsychology · 1,552 worksDevelopmental and Educational Psychology · 499 works

Cite

Cite as Gus Lathouwers, Wieke Harmsen, Catia Cucchiarini and Helmer Strik. 2025. Assessing Dutch Syllabification Algorithms and Improving Accuracy by Combining Phonetic and Orthographic Information through Deep Learning. arXiv:2605.28834 · Computational Linguistics in the Netherlands Journal, Vol. 14 (2025), pp. 365 to 383.

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.48550/arxiv.2605.28834 · arXiv 2605.28834

@misc{lathouwers2025assessing,
  title = {{Assessing Dutch Syllabification Algorithms and Improving Accuracy by Combining Phonetic and Orthographic Information through Deep Learning}},
  author = {Gus Lathouwers, Wieke Harmsen, Catia Cucchiarini and Helmer Strik},
  year = {2025},
  doi = {10.48550/arxiv.2605.28834},
  eprint = {2605.28834},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2605.28834}
}

References

34 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

  1. Adsett, C. R. and Y. Marchand (2009), A comparison of data-driven automatic syllabification methods, in Karlgren, J., J. Tarhio, and H. Hyyrö, editors, String Processing and Information Retrieval, SPIRE 2009, Vol. 5721 of Lecture Notes in Computer Science, Springer, Berlin, Heidelberg, pp. 357–368. find
  2. Ahmed, S. F., M. S. B. Alam, M. Hassan, M. R. Rozbu, T. Ishtiak, N. Rafa, M. Mofijur, A. B. M. S. Ali, and A. H. Gandomi (2023), Deep learning modelling techniques: current progress, applications, advantages, and challenges, Artificial Intelligence Review 56 (11), pp. 13521– 13617. find
  3. Alfiansyah, F. and Suyanto (2018), Partial greedy algorithm to extract a minimum phonetically-and-prosodically rich sentence set, International Journal of Advanced Computer Science and Applications. ↗
  4. Ayyaswami, V., D. Padmanabhan, M. Patel, A. V. Prabhu, D. R. Hansberry, N. Agarwal, and J. W. Magnani (2019), A readability analysis of online cardiovascular disease-related health education materials, HLRP Health Literacy Research and Practice. ↗
  5. Bartlett, S., G. Kondrak, and C. Cherry (2008), Automatic syllabification with structured svms for letter-to-phoneme conversion, Meeting of the Association for Computational Linguistics, pp. 568–576. https://webdocs.cs.ualberta.ca/ kondrak/papers/acl08Syl.pdf. ↗
  6. Boot, M.N.M. (1984), Taal, tekst, computer, Servire, Katwijk. find
  7. Boulesteix, Anne-Laure, S. Lauer, and Manuel (2013), A plea for neutral comparison studies in computational sciences, PLOS ONE 8 (4), pp. e61562–e61562. find
  8. Bouma, G. (2003), Finite state methods for hyphenation, Natural Language Engineering 9 (1), pp. 5–20. find
  9. Brandt Corstius, H. (1970), Exercises in computational linguistics, Doctoral dissertation, University of Amsterdam. find
  10. Corlatescu, D. G., S. Ruseti, and M. Dascalu (2022), Romanian syllabification using deep neural networks, in Mealha, O., M. Dascalu, and T. Di Mascio, editors, Proceedings of the Ludic, Co-design and Tools Supporting Smart Learning Ecosystems and Smart Education, Vol. 249 of Smart Innovation, Systems and Technologies, Springer, pp. 89–100. find
  11. Daelemans, Walter (1989), Automatic hyphenation: Linguistics versus engineering, in Heyvaert, F.J. and F. Steurs, editors, Worlds behind words, Leuven University Press, Leuven, pp. 347– 364. find
  12. Daelemans, Walter and Antal van den Bosch (1992), A neural network for hyphenation, Proceedings of the 1992 International Conference on Artificial Neural Networks (ICANN–92), Brighton, United Kingdom, pp. 1647–1650. find
  13. Dascalu, Mihai A, Gabriel S Gutu, Stefan S Ruseti, Ionut S Cristian Paraschiv, Philippe Dessus, Danielle A Mcnamara, Scott A Crossley, and Stefan A Trausan-Matu (2017), ReaderBench Learns Dutch: Building a comprehensive automated essay scoring system for Dutch language, Springer International Publishing AG. find
  14. Friedrich, S. and T. Friede (2023), On the role of benchmarking data sets and simulations in method comparison studies, Biometrical Journal. ↗
  15. Gale, R., A. Salem, G. Fergadiotis, and S. Bedrick (2023), Mixed orthographic/phonemic language modeling: Beyond orthographically restricted transformers (bort), ACLWeb; Association for Computational Linguistics. find
  16. Gavrishchaka, Valeriy, Zhenyi Yang, Rebecca Miao, and Olga Senyukova (2018), Advantages of hybrid deep learning frameworks in applications with limited data, International Journal of Machine Learning and Computing 8 (6), pp. 549–558. find
  17. Gove, Philip Babcock (1993), Webster’s Third New International Dictionary of the English Language, Unabridged, Merriam-Webster Inc. find
  18. Hauer, B. and G. Kondrak (2013), Automatic generation of english respellings, North American Chapter of the Association for Computational Linguistics, pp. 634–643. https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.353.9670&rep=rep1&type=pdf. ↗
  19. Krantz, J., M. Dulin, and P. De Palma (2019), Language-agnostic syllabification with neural sequence labeling, 2021 20th IEEE International Conference on Machine Learning and Applications (ICMLA), Vol. 3, pp. 804–810. https://doi.org/10.1109/icmla.2019.00141. ↗
  20. Liang, F. M. (1983), Word hy-phen-a-tion by com-put-er, Doctoral dissertation, Stanford University. find
  21. Man, A. and C. Van Ballegooie (2021), Assessment of the readability of web-based patient education material from major canadian pediatric associations: cross-sectional study, JMIR Pediatrics and Parenting. ↗
  22. Marco, G., J. De La Rosa, J. Gonzalo, S. Ros, and E. Gonzalez-Blanco (2021), Automated metric analysis of spanish poetry: Two complementary approaches, IEEE Access 9, pp. 51734– 51746. find
  23. Nießl, C., M. Herrmann, C. Wiedemann, G. Casalicchio, and A. Boulesteix (2021), Over-optimism in benchmark studies and the multiplicity of design and analysis options when interpreting their results, Wiley Interdisciplinary Reviews Data Mining and Knowledge Discovery. ↗
  24. Phang, Jason, Angelica Chen, William Huang, and Samuel R. Bowman (2021), Adversarially constructed evaluation sets are more challenging, but may not be fair, arXiv. arXiv:2111.08181 [cs.CL]. https://doi.org/10.48550/arXiv.2111.08181. ↗
  25. Pradhan, A., A. S. S, A. Prakash, K. Veezhinathan, and H. Murthy (2013), A syllable based statistical text to speech system, European Signal Processing Conference, pp. 1–5. https://dblp.uni-trier.de/db/conf/eusipco/eusipco2013.html#PradhanSPVM13. ↗
  26. Sarker, I. H. (2021), Deep learning: A comprehensive overview on techniques, taxonomy, applications and research directions, SN Computer Science. https://doi.org/10.1007/s42979-021-00815-1. ↗
  27. Sarma, Parismita and S. K. Sarma (2020), Syllable based approach for text to speech synthesis of assamese language: A review, Journal of Physics: Conference Series. find
  28. Schiller, N. O., A. S. Meyer, R. H. Baayen, and W. J. M. Levelt (1996), A comparison of lexeme and speech syllables in dutch, Journal of Quantitative Linguistics 3 (1), pp. 8–28. find
  29. Strobl, C. and F. Leisch (2022), Against the “one method fits all data sets” philosophy for comparison studies in methodological research, Biometrical Journal. find
  30. Tits, Noé (2023), Must&p-srl: Multi-lingual and unified syllabification in text and phonetic domains for speech representation learning, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: Industry Track, Association for Computational Linguistics, Singapore, pp. 74–82. find
  31. Trogkanis, N. and C. Elkan (2010), Conditional random fields for word hyphenation, Meeting of the Association for Computational Linguistics, pp. 366–374. find
  32. Tutelaers, P. (1993), Herziene https://www.ntg.nl/maps/11/35.pdf. ↗
  33. Van Der Sijs, Nicoline (2005), Van Dale Groot Leenwoordenboek: De invloed van andere talen op het Nederlands, Van Dale Lexicografie. https://www.dbnl.org/tekst/sijs002groo01 01/. ↗
  34. Yousefi, M. R., J. Hua, C. Sima, and E. R. Dougherty (2009), Reporting bias when using real data sets to analyze classification performance, Bioinformatics 26 (1), pp. 68–76. afbreekpatronen voor het nederlands. find

Updated

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/lathouwers-2026-assessing-dutch-syllabification-algorithms-improving/