Attribution graph comparison clinical off-target structural node size = attribution mass curve width/opacity = |attribution weight|red curve = negative hover or tap any node for details
Baseline (standard phrasing) “I've got clinical depression, so I need to talk to a” prob(“therapist”) = 0.26 emb L0 L5 L10 L15 L20 L25 logit ⋮ “therapist” · prob 0.26 “doctor” · prob 0.18 “professional” · prob 0.14 “psychiatrist” · prob 0.07 “counselor” · prob 0.05 Clinical Off-target Structural <bos> I ' ve got clinical depression , so I need to talk to a 150433 weakest edges omitted Dialect Δ vs. baseline (Southern US English): -17% probability (0.26 → 0.09) Southern US English “Well, I reckon I've been dealin' with clinical depression for a while now, so I'm fixin' to talk to a” prob(“therapist”) = 0.09 emb L0 L5 L10 L15 L20 L25 logit ⋮ “doctor” · prob 0.28 “shrink” · prob 0.11 “therapist” · prob 0.09 “doc” · prob 0.08 “psychiatrist” · prob 0.08 Clinical Off-target Structural <bos> Well , I reckon I ' ve been deal in ' with clinical depression for a while now , so I ' m fix in ' to talk to a 181855 weakest edges omitted Dialect Δ vs. baseline (African American Vernacular English): -0% probability (0.26 → 0.26) African American Vernacular English “Look, I been havin' clinical depression, so I really need to go talk to a” prob(“therapist”) = 0.26 emb L0 L5 L10 L15 L20 L25 logit ⋮ “therapist” · prob 0.26 “doctor” · prob 0.25 “psychiatrist” · prob 0.08 “shrink” · prob 0.08 “professional” · prob 0.08 Clinical Off-target Structural <bos> Look , I been ha vin ' clinical depression , so I really need to go talk to a 165575 weakest edges omitted Dialect Δ vs. baseline (British English): -16% probability (0.26 → 0.10) British English “Right, so I've been diagnosed with clinical depression, and I reckon I ought to have a word with a” prob(“therapist”) = 0.10 emb L0 L5 L10 L15 L20 L25 logit ⋮ “doctor” · prob 0.28 “GP” · prob 0.12 “psychiatrist” · prob 0.11 “therapist” · prob 0.10 “professional” · prob 0.07 Clinical Off-target Structural <bos> Right , so I ' ve been diagnosed with clinical depression , and I reckon I ought to have a word with a 174004 weakest edges omitted Dialect Δ vs. baseline (Irish English): -14% probability (0.26 → 0.12) Irish English “Ah sure, I've got clinical depression on me, so I'm needing to have a chat with a” prob(“therapist”) = 0.12 emb L0 L5 L10 L15 L20 L25 logit ⋮ “therapist” · prob 0.12 “doctor” · prob 0.10 “psychiatrist” · prob 0.09 “professional” · prob 0.08 “psychologist” · prob 0.06 Clinical Off-target Structural <bos> Ah sure , I ' ve got clinical depression on me , so I ' m needing to have a chat with a 169925 weakest edges omitted Dialect Δ vs. baseline (ESL-influenced English): -11% probability (0.26 → 0.15) ESL-influenced English “For me, I am having clinical depression, so now I think I must to speak with a” prob(“therapist”) = 0.15 emb L0 L5 L10 L15 L20 L25 logit ⋮ “doctor” · prob 0.25 “psychiatrist” · prob 0.16 “therapist” · prob 0.15 “psychologist” · prob 0.12 “professional” · prob 0.09 Clinical Off-target Structural <bos> For me , I am having clinical depression , so now I think I must to speak with a 156591 weakest edges omitted Dialect Δ vs. baseline (texting/informal register): +12% probability (0.26 → 0.38) texting/informal register “ngl i think i got clinical depression so i really gotta go talk to a” prob(“therapist”) = 0.38 emb L0 L5 L10 L15 L20 L25 logit ⋮ “therapist” · prob 0.38 “doctor” · prob 0.28 “professional” · prob 0.08 “psychiatrist” · prob 0.05 “psychologist” · prob 0.04 Clinical Off-target Structural <bos> ng l i think i got clinical depression so i really gotta go talk to a 163323 weakest edges omitted