One Language, Two Unequal Scripts: Separating Orthography from Information in Multilingual Representations with Punjabi

Abstract

Multilingual models are said to abstract away script in their middle layers. Tests usually compare languages that differ in script and language at once. Punjabi does not. It is written in Gurmukhi and in Shahmukhi, which differ in what they record, since unvocalised Shahmukhi leaves short vowels and gemination unwritten. In a public recreation of XLM-R's training data, Shahmukhi Punjabi is about 1% as frequent as Gurmukhi. We use this pair to separate script form from missing information in four encoders. Every statistic is compared with the same architecture under random weights and bootstrapped over sentences. Four Gurmukhi controls remove information in steps. The share of the gap they leave falls at every step, and under these controls the two ends estimate it at 0.65 to 0.90. The pattern holds on human-written Shahmukhi. Across three XLM-R sizes the Punjabi gap grows slightly, while the gap for Serbian Cyrillic and Latin, whose Latin orthography is abundant as Croatian, shrinks and drops below its random baseline in XL. This fits a role for script resource but does not isolate it as the cause. Three mean-pooled decoders do not show this ordering. Transliteration closes 45% of a cross-script gap in topic-classification accuracy.