mBERT's syntactic probe subspace is approximately shared across languages

Chi, Hewitt & Manning apply the structural probe unchanged to multilingual BERT across 11 typologically varied languages, recovering dependency structure well above baseline in every one (avg UUAS 76.8%) [chi-hewitt-manning-2020] Zero-shot transfer (probe trained on one language, tested on another) still recovers substantial structure (avg UUAS 70.4% holding out the target language entirely) [chi-hewitt-manning-2020] The principal angle between each language pair's probe subspaces correlates with their zero-shot transfer quality (Spearman 0.78 UUAS, 0.82 distance), so the subspaces are approximately shared, not merely separately effective [chi-hewitt-manning-2020] A qualitative t-SNE of head-dependent difference vectors visually tracks the UD taxonomy but is reported only as supplementary, with no clustering metric [chi-hewitt-manning-2020]