Attribution graph comparison clinical off-target structural red curve = negative
A · Medical lexicon + standard morphosyntax (prestige form) “Every time I climb the stairs I feel tachycardia, so this afternoon I am heading to the” prob(“cardio”) = 0.07 emb L0 L5 L10 L15 L20 L25 logit ⋮ “doctor” · prob 0.29 “hospital” · prob 0.12 “gym” · prob 0.07 “cardio” · prob 0.07 “emergency” · prob 0.07 <bos> Every time I climb the stairs I feel tachycardia, so this afternoon I am heading to the 416546 weakest edges omitted Variety shift Δ (A → B): -3% probability (0.07 → 0.04) shared circuit dimmed: 1077 features in both · 294 only in A · 311 only in B B · Medical lexicon + nonstandard morphosyntax (variety shift only) “Every time I be climbing them stairs I feel tachycardia, so this afternoon I be heading to the” prob(“cardio”) = 0.04 emb L0 L5 L10 L15 L20 L25 logit ⋮ “gym” · prob 0.18 “doctor” · prob 0.18 “hospital” · prob 0.14 “pharmacy” · prob 0.04 “cardio” · prob 0.04 <bos> Every time I be climbing them stairs I feel tachycardia, so this afternoon I be heading to the 407471 weakest edges omitted