Rationalizing Medical Relation Prediction from Corpus-level - PowerPoint PPT Presentation

ACL 2020 Rationalizing Medical Relation Prediction from Corpus-level Statistics Zhen Wang The Ohio State University In collaboration with Jennifer Lee (NCH), Simon Lin (NCH), Huan Sun (OSU)

One-Minute Summary What is this paper about? Task : Predicting relations between two given terms from a text corpus Goal : Make accurate prediction & Provide justifications for it (Rationalization) 1. Draw inspirations from human Ca��e��e M���a��e ma��treat memory recall and recognition 428 2341 521 Pa���Re��e� 2. Recall global associations (blue) 38� ma��ca�se A��i�i� Headache from corpus-level statistics and recognize meaningful relational 1356 Fe�e� 122 7�4 1�8 ma��treat assumptions (red) between them 3. Justify the prediction ( Aspirin may C�-�cc���e�ce�Li�k� 122 A���cia�i���Recall A������i���Rec�g�i�i�� treat headache ) by highlighting (�i�h�c�����) important associations and assumptions 1-Min Summary 1

Black-Box Relation Classifier Medical Records Aspirin Headache Input Pair Black-Box Classifier Class Distribution Overview 2

Black-Box Relation Classifier Medical Records Aspirin Headache Input Pair Black-Box Classifier Class Distribution High Risk! Medicine Finance Judiciary Overview 3

Can We Open the Black-Box? Medical Records Aspirin Headache Input Pair Class Distribution Overview 4

Can We Open the Black-Box? Medical Records Aspirin Headache Input Pair Class Distribution To solve the problem, we draw inspirations from human memory theories – Recall and Recognition! Overview 5

Memory Recall and Recognition Recall Retrieve association information from long-term memory Recognition Identify previously learned information Overview 6

Memory Recall and Recognition Recall Retrieve association information Recall my friend’s name from long-term memory Recognition Recognize my friend’s face Identify previously learned information Example from http://webdesign-review.blogspot.com/2016/04/recognition-vs-recall-in-mobile-web.html Overview 7

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 CogStage-3 Assumptions Associations Recall Recognition Entity Pair Pred. Memory Memory Rationalized By Overview 8

Rationalizing Medical Relation Prediction Entity Pair Aspirin Headache Overview 9

Rationalizing Medical Relation Prediction CogStage-1 Recall Entity Pair Memory Aspirin Headache Overview 10

Rationalizing Medical Relation Prediction CogStage-1 Associations Recall Entity Pair Memory Caffeine Migraine Pain Relief Aspirin Headache Fever Overview 11

Rationalizing Medical Relation Prediction CogStage-1 Associations Recall Entity Pair Memory Caffeine Migraine 353 Pain Relief 428 463 Aspirin Headache �84 83 385 123 146 Fever 130 Corpus-level Statistics Overview 12

Rationalizing Medical Relation Prediction Medical Records CogStage-1 Associations Aspirin Recall Entity Pair Memory Headache Caffeine Migraine 353 Pain Relief 428 463 Aspirin Headache �84 83 385 123 146 Fever 130 Corpus-level Statistics Overview 13

Rationalizing Medical Relation Prediction CogStage-1 Associations Recall Entity Pair Memory Caffeine Migraine Association Strength 353 Pain Relief 428 463 Aspirin Headache �84 83 385 123 146 Fever 130 Corpus-level Statistics Overview 14

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 Associations Recall Recognition Entity Pair Memory Memory Caffeine Migraine Pain Relief Aspirin Headache Fever Overview 15

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 Associations Recall Recognition Entity Pair Memory Memory Caffeine Migraine Pain Relief may_cause Aspirin Headache Fever Overview 16

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 Assumptions Associations Recall Recognition Entity Pair Memory Memory Caffeine Migraine Pain Relief may_cause Aspirin Headache Fever Overview 17

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 CogStage-3 Assumptions Associations Recall Recognition Entity Pair Pred. Memory Memory Caffeine Migraine Migraine Pain Relief may_cause Aspirin Headache Fever Overview 18

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 CogStage-3 Assumptions Associations Recall Recognition Entity Pair Pred. Memory Memory Rationalized By Caffeine Migraine Pain Relief may_cause Aspirin Headache Fever Overview 19

Rationalizing Medical Relation Prediction CogStage-1 CogStage-2 CogStage-3 Assumptions Associations Recall Recognition Entity Pair Pred. Memory Memory Rationalized By Caffeine Migraine may_treat Pain Relief may_cause Aspirin Headache Fever may_treat Overview 20

Framework Overview Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics 21 Model

CogStage-1: Global Association Recall Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics 22 Model

CogStage-1: Global Association Recall Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics 4�3 ��4 �3 123 14� Empirical Context Distribution 23 Model

CogStage-1: Global Association Recall Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics ⇣ ⌘ v 0 T exp e j · v e i p ( e j | e i ) = 4�3 ⇣ ⌘ P |V| v T k =1 exp k · v e i ��4 e 0 �3 123 14� Empirical Context Distribution 24 Model

CogStage-1: Global Association Recall Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics ⇣ ⌘ v 0 T exp e j · v e i p ( e j | e i ) = 4�3 ⇣ ⌘ P |V| v T k =1 exp k · v e i ��4 e 0 �3 123 14� Estimated Context Distribution Empirical Context Distribution 25 Model

CogStage-1: Global Association Recall Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics ⇣ ⌘ v 0 T exp e j · v e i p ( e j | e i ) = 4�3 ⇣ ⌘ P |V| v T k =1 exp k · v e i ��4 e 0 �3 123 14� Estimated Context Distribution Empirical Context Distribution 26 Model

CogStage-2: Assumption Formation and Representation Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics $ hold a relationship 𝑠 " and 𝑏 # Assumption : Recognize whether context 𝑏 ! % ? 27 Model

CogStage-2: Assumption Formation and Representation Gl�bal�A���cia�i���Recall A������i���F���a�i���&�Re��e�e��a�i�� Deci�i���Maki�g 353 428 463 �84 OWA� Rela.� 83 �e�a.��ec. head��ec. �ai���ec. �P�ed.�� �a�i��ale� 385 123 146 130 Corpus-level�Statistics $ hold a relationship 𝑠 " and 𝑏 # Assumption : Recognize whether context 𝑏 ! % ? Closed World Assumption (CWA): Open World Assumption (OWA): 28 Model