Symphonym: Universal Phonetic Embeddings for Cross-Script Name Matching

What it is about

Symphonym maps place names from 36 writing systems into a 128-dimensional phonetic space, so that the same place written in different scripts lands close together. It is a Teacher-Student distillation: a Teacher (a BiLSTM over PanPhon articulatory features of IPA transcriptions from Epitran, Phonikud and CharsiuG2P) is trained with a triplet loss, and a character-level Student is trained to copy it, so that at inference time no IPA or language identification is needed. The training data are 73.5 million toponyms from GeoNames, Wikidata and the Getty TGN. On the MEHDIE medieval Hebrew–Arabic benchmark (137 queries in five test sets) the model reaches Recall@1 89.3% and MRR 92.8%, against 85.2% and 90.8% for the first generation, and well above the string baselines and a raw-feature ablation (41.1% R@1). The revision reports three corrections: a language tag error made Chinese names learn Japanese readings; the model had barely learned character order, with 70.5% of random anagrams passing the retrieval gate, which permutation negatives cut to 5.0%; and the better-ranking model separates true from false pairs slightly worse (AUC 0.9324 to 0.9270). String metrics still win within the Latin script.

Ideas

• I1. An order-sensitivity band test for any Mova matcher that compares two forms (for example pointed and unpointed lookup in the lexicon). Use five bands: a true variant, an adjacent transposition, a full permutation, a reversal, and an unrelated form. A good matcher keeps the variant and the transposition, and drops the permutations and the reversal. Score the bands, not single examples.
• I2. Keep the two kinds of matching apart. Within one alphabet, use plain edit distance on consonantal forms; across scripts, use a deterministic table, not a learned embedding. This is the paper's own split, and it fits tools/ocrfix, where the correction is already deterministic.
• I3. Ship weights, vocabularies and indexes as one versioned set with checksums, because a mismatched pair raises no error and gives plausible nonsense. Apply this to the segmenter tables (lex, pre) and the gparse model in tools/hbo, so a release fails loudly when one part is replaced alone.

For Mova

I1 is a test for the lexicon lookup in The lexicon and for the unpointed mode of tools/hbo. I2 is a rule for tools/ocrfix and for the orthography notes in 2. Script and spelling. I3 is a release check for tools/hbo and 13. Biblical Hebrew in Mova: the tools. Phonikud, the Hebrew G2P used in the paper, is a neural backend and is not used here.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. arXiv 2601.06932

@misc{gadd2026symphonym,
  title = {{Symphonym: Universal Phonetic Embeddings for Cross-Script Name Matching}},
  author = {Gadd, Stephen},
  year = {2026},
  eprint = {2601.06932},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.06932}
}

References

21 references, 2 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Mikel Artetxe, Gorka Labaka, and Eneko Agirre. A robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. In Iryna Gurevych and Yusuke Miyao, editors, Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 789–798, Melbourne, Australia, July 2018. Association for Computational Linguistics. doi: 10.18653/v1/P18-1073. URL https://aclanthology.org/P18-1073/. ↗
2. Alexis Conneau, Guillaume Lample, Marc’Aurelio Ranzato, Ludovic Denoyer, and Hervé Jégou. Word translation without parallel data. In Proceedings of the International Conference on Learning Representations, 2018. find
3. Stephen Gadd. Symphonym: Universal phonetic embeddings for cross-script name matching, 2026a. URL https://arxiv.org/abs/2601.06932. ↗ our card
4. Stephen Gadd. Symphonym v8 — universal phonetic embeddings for cross-script toponym matching, 2026b. URL https://huggingface.co/docuracy/symphonym-v8. ↗
5. Stephen Gadd. Symphonym v8 — universal phonetic embeddings for cross-script toponym matching, 2026c. URL https://doi.org/10.5281/zenodo.22767194. ↗
6. T. N. Gadd. PHONIX: The algorithm. Program: Automated Library and Information Systems, 24(4):363–366, 1990. doi: 10.1108/eb047069. ↗
7. Linda L. Hill. Georeferencing: The geographic associations of information. Digital Libraries and Electronic Publishing, 2006. doi: 10.7551/mitpress/3260.001.0001. ↗
8. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. doi: 10.48550/arXiv.1503.02531. URL https://arxiv.org/abs/1503.02531. ↗
9. Yakov Kolani, Maxim Melichov, Cobi Calev, and Morris Alper. Phonikud: Hebrew grapheme-to-phoneme conversion for real-time text-to-speech. 2025. doi: 10.48550/arXiv.2506.12311. URL https://arxiv.org/abs/2506.12311. ↗ our card
10. Yu. A. Malkov and D. A. Yashunin. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(4):824–836, 2020. doi: 10.1109/TPAMI.2018.2889473. ↗
11. Leland McInnes, John Healy, and Steve Astels. hdbscan: Hierarchical density based clustering. Journal of Open Source Software, 2(11):205, 2017. doi: 10.21105/joss.00205. ↗
12. David R. Mortensen, Patrick Littell, Akash Bharadwaj, Kartik Goyal, Chris Dyer, and Lori Levin. Panphon: A resource for mapping IPA segments to articulatory feature vectors. In Proceedings of the 26th International Conference on Computational Linguistics (COLING), pages 3475–3484, 2016. URL https://aclanthology.org/C16-1328/. ↗
13. David R. Mortensen, Siddharth Dalmia, and Patrick Littell. Epitran: Precision G2P for many languages. In Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC), pages 2634–2639, 2018. URL https://aclanthology.org/L18-1140/. ↗
14. Lawrence Philips. Hanging on the metaphone. Computer Language, 7(12):39–43, 1990. find
15. Qinjun Qiu, Haiyan Li, Xinxin Hu, Miao Tian, Kai Ma, and Yunqiang Zhu. A deep neural network model for chinese toponym matching with geographic pre-training model. International Journal of Digital Earth, 17(1):2353111, 2024. doi: 10.1080/17538947.2024.2353111. Proposal of the GeoBERT model. ↗
16. Taraka Rama. Siamese convolutional networks for cognate identification. In Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1018–1027, 2016. URL https://aclanthology.org/C16-1097/. ↗
17. Gabriel Recchia and Max Louwerse. A comparison of string similarity measures for toponym matching. In Vagueness in Communication, pages 54–67. Springer, 2013. doi: 10.1007/978-3-642-18446-8_4. ↗
18. Robert C. Russell. Index. U.S. Patent 1,261,167, 1918. URL https://patents.google.com/patent/US1261167A/. Patented Apr. 2, 1918. ↗
19. Tomer Sagi, Moran Zaga, Sinai Rusinek, Marcell Richard Fekete, Johannes Bjerva, and Katja Hose. Utilizing phonetic similarity for cross-source and cross-language toponym matching: a benchmark and prototype. Language Resources and Evaluation, 59(3):2427–2451, 2025. doi: 10.1007/s10579-025-09812-9. ↗
20. Rodrigo Santos, Patricia Murrieta-Flores, Pablo Calafiore, and Bruno Martins. Learning to combine multiple string similarity metrics for effective toponym matching. International Journal of Digital Earth, 11(9):949–965, 2018. doi: 10.1080/17538947.2017.1371253. ↗
21. Jian Zhu, Cong Zhang, and David Jurgens. Byt5 model for massively multilingual grapheme-to-phoneme conversion. 2022. doi: 10.48550/ARXIV.2204.03067. URL https://arxiv.org/abs/2204.03067. ↗

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/langstream-2026-gadd-symphonym-universal-phonetic-embeddings-cross/