Familiar scripts help models translate unseen languages

Researchers used the GlyphMind model to demonstrate that familiar characters, rather than linguistic relatives, help artificial intelligence map new languages into a shared space of meaning.

The team tested GlyphMind, a vision transformer that processes text as pixels, on Lao, a Southeast Asian language and script the model had never encountered during training. The model successfully retrieved English translations for Lao seven times more often than the strongest simple baseline.

This suggests that the visual similarity of characters plays a major role in how models organize information. The study found that linguistic similarity alone does not bridge the gap for unseen languages. When Lao was rewritten in unfamiliar letters, the model failed to process it even though it had been trained on Thai, a closely related language.

Conversely, Santali, a language with no prior training data, was successfully mapped when written in Devanagari, a script familiar to the model from its Hindi training. These results suggest that transliterating rare languages into familiar scripts could allow multilingual models to process languages that currently lack available translations.