Universal or Language-Family-Specific Script Unification for Cross-Lingual Transfer? A Case Study on Turkic Languages

Multilingual models often struggle to transfer knowledge between closely related languages when they are written in different scripts, as the lack of surface-level character overlap prevents the model from recognizing shared linguistic structures. This paper investigates whether it is better to use a universal romanizer (like uroman) or a family-specific script (like the Common Turkic Script, CTS) to unify these writing systems. The author compares these two approaches across 11 Turkic languages, training matched fastText models on transliterated Wikipedia corpora and evaluating them on named entity recognition (NER) and part-of-speech (POS) tagging.

The study yields three primary results. First, there is no significant difference between universal romanization and family-specific unification for NER tasks, though both methods substantially outperform official monolingual fastText baselines. Second, neither approach dominates POS tagging; uroman performs better for some languages (e.g., Azerbaijani and Tatar), while CTS is more effective for others (e.g., Uyghur). Third, these performance differences are linked to the character n-gram coverage induced by each representation. While universal romanization maximizes broad character sharing, family-specific unification preserves linguistically relevant contrasts. The study also notes that while more complex models like CANINE-c achieve higher overall POS averages, the simpler fastText-based systems remain highly competitive.

This research demonstrates that script unification is a powerful tool for improving cross-lingual transfer in related, lower-resource languages. It challenges the assumption that a single "universal" approach to script normalization is always optimal. Instead, the results suggest that the choice of unification strategy should be guided by the specific linguistic properties of the target family and the nature of the downstream task. By showing that simple, unified representations can outperform monolingual baselines, the paper provides a practical, computationally efficient pathway for enhancing multilingual NLP in resource-constrained settings.