Applying dependency parses and SRL Subject and Generic

Applying dependency parses and SRL: Subject and Generic Attribute Discovery Stephen Wu, Mayo Clinic SHARPn Summit 2012 June 11, 2012

Outline • Motivation and Role • Generic Attribute • Definition • Methods & Examples • Subject Attribute • Definition • Methods & Examples • Status & Future Work

Attribute Discovery • Clinical Element Models • Exclude generic • Family history • Methods: Dependency Parser and SRL

Methods summary • Types of rules • Noun phrase structure • Path to root • Path between pairs • Semantic arguments • Feature vector • Decision logic/ML

Generic: Attribute Definition (a) The patient was referred to the Lupus clinic. (b) We discussed increased risk of breast cancer Definition: “refers to mentions, which are generic, i. e. , not related to the instance of a disorder, sign/symptom, etc…” “… Mentioned as part of a general statement with no clear subject/experiencer. ” Values: in {true, false} default=false

Generic: Dependency parse rules Ex: Noun phrase structure Rule (a) The patient was referred to the Lupus clinic. sbj nmod was vc referred adv patient pmod to clinic nmod the • Find the headword of the NE • Modifies another noun (nmod)? Lupus generic=true

Generic: Dependency parse rules Ex: Path to root Rule (“Discussion” context) (b) We discussed increased risk of breast cancer sbj We discussed obj risk nmod increased nmod of pmod nmod cancer breast • Find NE headword • Path to top • “Discussion” word? generic=true discuss, ask, understand, understood, tell, told, mention, talk, speak, spoke, address

Subject: Attribute Definition (c) The patient’s son has schizophrenia. (d) Father died of MI in 50’s Definition: “The person the observation is on. This modifier refers to the entity experiencing the disorder. ” Values: in {Patient, Family_Member, Donor_Other, and Other} default=Patient

Subject: Semantic role labeling rules Ex: Semantic argument Rule (c) The patient’s son has schizophrenia. son has PRED ARG 0 schizophrenia patient the ARG 1 ‘s subject=family_member • Semantic argument (ARG 0, ARG 1) • Family term (Word. Net) father, dad, mother, mom, bro, sis, sib, cousin, aunt, uncle, grandm, grandp, grandf, wife, spouse, husband, child, offspring, progeny, son, daughter, nephew, niece, kin, family

Subject: Dependency parse rules Ex: Path to root Rule (family) (d) … father who died of MI in 50's father nmod who tmp died adv in of pmod 50 s MI • Find NE headword • Path to top • Family term? subject=family_member

Subject: Dependency parse rules Ex: Dependency paths Rule (d) Father died of MI in 50's sbj Father tmp died adv in of pmod 50 s MI subject=family_member • NE + “Family” pairs • Find dependency path • Once-removed?

Methods summary • Types of rules • Noun phrase structure • Path to root • Path between pairs • Semantic arguments • Feature vector • Decision logic/ML

Status and Future Work • c. TAKES v 2. 5 • “Assertion” module • Default • Future work (with data) • Evaulation & Error analysis • Improved rules • Features in machine learning

Task 4/6 team: Stephen Wu Cheryl Clark James Masanz Matt Coarr Ben Wellner https: //sites. google. com/site/stephentzeinnwu wu. stephen@mayo. edu THANK YOU. Special thanks to: Lee Becker Guergana Savova Pei Chen This work was supported in part by the SHARPn (Strategic Health IT Advanced Research Projects) Area 4: Secondary Use of EHR Data Cooperative Agreement from the HHS Office of the National Coordinator, Washington, DC. DHHS 90 TR 000201.