Symphonym: Universal Phonetic Embeddings for Cross-Script Toponym Matching via Teacher-Student Distillation

TL;DR

Symphonym tackles cross-script toponym matching by learning phonetic embeddings that are script-agnostic. It uses a Teacher network grounded in articulatory PanPhon features to shape a 128-dimensional embedding space and a lightweight Student network that learns to approximate these embeddings directly from raw characters, enabling inference without runtime phonetic processing. The training follows a three-phase curriculum—phonetic feature learning, teacher-student distillation, and hard negative discriminative fine-tuning—applied to 57.6 million toponyms from GeoNames, Wikidata, and Getty TGN. On the MEHDIE Hebrew-Arabic benchmark, Symphonym surpasses Levenshtein and Jaro-Winkler, achieving Recall@1 of 89.2% and demonstrating strong cross-script generalization that supports fuzzy phonetic search over the World Historical Gazetteer’s 67 million toponyms, with public code and models provided for reproducibility and deployment.

Abstract

Linking place names across languages and writing systems is a fundamental challenge in digital humanities and geographic information retrieval. Existing approaches rely on language-specific phonetic algorithms or transliteration rules that fail when names cross script boundaries -- no string metric can determine that "Moscow" when rendered in Cyrillic or Arabic refer to the same city. I present Symphonym, a neural embedding system that maps toponyms from 20 writing systems into a unified 128-dimensional phonetic space. A Teacher network trained on articulatory phonetic features (via Epitran and PanPhon) produces target embeddings, while a Student network learns to approximate these from raw characters. At inference, only the lightweight Student (1.7M parameters) is required, enabling deployment without runtime phonetic conversion. Training uses a three-phase curriculum on 57 million toponyms from GeoNames, Wikidata, and the Getty Thesaurus of Geographic Names. Phase 1 trains the Teacher on 467K phonetically-grounded triplets. Phase 2 aligns the Student to Teacher outputs across 23M samples, achieving 96.6% cosine similarity. Phase 3 fine-tunes on 3.3M hard negative triplets -- negatives sharing prefix and script with the anchor but referring to different places -- to sharpen discrimination. Evaluation on the MEHDIE Hebrew-Arabic benchmark achieves 89.2% Recall@1, outperforming Levenshtein (81.5%) and Jaro-Winkler (78.5%). The system is optimised for cross-script matching; same-script variants can be handled by complementary string methods. Symphonym will enable fuzzy phonetic reconciliation and search across the World Historical Gazetteer's 67 million toponyms. Code and models are publicly available.

Figures (2)

• Figure 1: Teacher-Student architecture. Left: The Teacher pathway converts toponyms to IPA via Epitran, extracts PanPhon articulatory features, and encodes to 128-d embeddings. Used only during training. Right: The Student pathway processes raw characters with script/language metadata. Deployed at inference time without phonetic resources. Bottom: Three training phases progressively transfer phonetic knowledge from Teacher to Student.
• Figure 2: Training curves for the three-phase curriculum. Left: Phase 1 Teacher training showing triplet loss convergence. Center: Phase 2 Student-Teacher alignment showing distillation loss and cosine similarity between Student and Teacher embeddings. Right: Phase 3 hard negative fine-tuning.