A Feature-Based Graph Beats Embeddings at Matching Idioms Across Eight Languages

arXiv·low signal

This interpretable framework annotates 160 idiomatic expressions across eight typologically diverse languages with binary conceptual features, then builds Jaccard-weighted graphs where community detection shows idioms cluster by conceptual schema rather than by language. Conceptual proximity identifies acceptable translation equivalents with substantial gains over embedding-based baselines across five language families, and ablations confirm schemas, roles, and valence each contribute non-redundantly.