Companion data and analysis scripts for 'Robust but not portable: a tiered invariance audit of cross-lingual embedding comparisons'
收藏资源简介:
Companion data and analysis scripts for the paper 'Robust but not portable: a tiered invariance audit of cross-lingual embedding comparisons'. The deposit contains the R3 dual-round blind human coding of 30 news documents on three framework poles (conflict/cooperation, human-security, economic/political) under independent English-only and Chinese-only reading rounds, the deterministic invariance-test battery (distributional / covariance / mean-structure tiers under MiniLM and LaBSE), and the machine-readable result JSONs at the full n=7148 six-language-aligned analysis pool (the entire aligned population, not a sample). Raw News-Commentary v16 text is not redistributed (third-party copyright); the scripts regenerate all embedding-derived outputs from the source corpus.



