Probing five architectures shows that character-level transformers represent complex Spanish L-shaped morphemes internally rather than just reproducing surface alternations. The encoding is localized and survives lemma-disjoint cross-validation even when stem variations are absent.
HOW THIS AFFECTS YOU
●
researcherYou can gain deeper insight into how transformer architectures represent linguistic irregularities beyond surface-level pattern matching.