Errudite: Scalable, Reproducible, and Testable Error Analysis - PowerPoint PPT Presentation

Errudite: Scalable, Reproducible, and Testable Error Analysis Tongshuang (Sherry) Wu @tongshuangwu University of Washington Marco Tulio Ribeiro Microsoft Research Jeffrey Heer @jeffrey_heer Daniel S. Weld @dsweld University of Washington � 1

Motivation & Contributions � 2

Error analysis is important for… Uncovering bugs Improving the state-of-art Safeguarding deployments � 3

Where We Are Fader et tal. We performed an error analysis on a sample of 100 questions ACL’13 Chen et al. We randomly select 50 incorrect questions and categorize ACL’16 them into 6 classes. We sample 100 incorrect predictions and try to find common Wadhwa et al. error categories. ACL’18 � 4

Where We Are Fader et tal. We performed an error analysis on a sample of 100 questions ACL’13 “We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Chen et al. We randomly select 50 incorrect questions and categorize ACL’16 them into 6 classes. We sample 100 incorrect predictions and try to find common Wadhwa et al. error categories. ACL’18 � 5

Where We Are “We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Subjectively defined hypotheses + Small samples + Focus exclusively on errors + No Test on true cause � 6

Where We Are & Our Contribution “We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Principles & Errudite Subjectively defined hypotheses Precise & reproducible h ypotheses + + Small samples Scale up to the entire dev set + + Focus exclusively on errors Cover errors & correct instances + + No Test on true cause Test via counterfactual analysis 7 �

A E C B D F � 8

A E C B D F Video demo: https://tinyurl.com/errudite-video � 9

Core Design Precise & Reproducible Domain Specific Language � 10

Precise DSL (Domain Specific Language) DSL = + + Target Attribute Extractor Operators length(q) > 20 Extract A E C E Instance Attribute B B D Filter F Instance Groups 11 �

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Subjectively defined hypotheses Subjectively defined hypotheses Precise & reproducible h ypotheses + + Too ambiguous to reproduce Small samples Scale up to the entire dev set + + Focus exclusively on errors Cover errors & correct instances + + No Test on true cause Test via counterfactual analysis � 12

User study: What is imprecise answer boundaries? “The model is making predictions with missing or additional words…?” D1 D2 No exact match, but high overlap O ff by at most 2 tokens both on the left and right exact_match(p(m)) == 0 exact_match(p(m)) == 0 and abs(answer_offset(p(m),"left")) <= 2 and f1(p(m)) > 0.7 and abs(answer_offset(p(m),"right")) <= 2 � 13

User study: What is imprecise answer boundaries? “The model is making predictions with missing or additional words…?” D1 No exact match, but high overlap D2 O ff by at most 2 tokens both on the left and right exact_match(p(m)) == 0 exact_match(p(m)) == 0 and abs(answer_offset(p(m),"left")) <= 2 and f1(p(m)) > 0.7 and abs(answer_offset(p(m),"right")) <= 2 � 14

User study: What is imprecise answer boundaries? D1 No exact match, but high overlap D2 O ff by at most 2 tokens both on the left and right D1 D2 groundtruth …the polynomial time hierarchy collapses. …believed that the polynomial hierarchy does.. prediction � 15

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Errudite Subjectively defined hypotheses Subjectively defined hypotheses Precise & reproducible h ypotheses Precise & reproducible h ypotheses + + Quantify instances with a domain Small samples Scale up to the entire dev set specific language + + Focus exclusively on errors Cover errors & correct instances + + No Test on true cause Test via counterfactual analysis 16 �

Design & Use Scenario Examine the distractor hypothesis on BiDAF (Seo et al., 2016), with SQuAD (10570 instances; Rajpurkar et al., 2016) Independently tested by 4 (out of 10) participants in the user study � 17

Scenario: distractor hypothesis Who created the 2005 theme for Doctor Who? Common belief: BiDAF… … John Debney created a new arrangement Matches entity types of Ron Grainer’s original theme for Doctor Knows to find a PERSON Who in 1996. For the return of the series in 2005, Murray Gold provided a new Finds the exact answer spans Distracted by other PERSON spans arrangement... featured sampled from the 1963 original. � 18

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Subjectively defined hypotheses Precise & reproducible h ypotheses + + Small samples Small samples Scale up to the entire dev set + + 100 << 2000+ errors in total Focus exclusively on errors Cover errors & correct instances + + No Test on true cause Test via counterfactual analysis � 19

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Errudite Subjectively defined hypotheses Precise & reproducible h ypotheses + + Small samples Small samples Scale up to the entire dev set Scale up to the entire dev set + + Focus exclusively on errors Cover errors & correct instances + + No Test on true cause Test via counterfactual analysis 20 �

Build distractor groups with DSL C C D ENT(g) != "" 1 and count(token(c, pattern=ENT(g))) > 2 count(token(g, pattern=ENT(g))) 3 and ENT(g) == ENT(p(m)) 4 and f1(m) == 0 5 � 21

Build distractor groups with DSL ENT(Murray Gold) == PERSON 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 count(token(g, pattern=ENT(g))) 3 and ENT(g) == ENT(p(m)) 4 and f1(m) == 0 5 “The g roundtruth is an ENT ity.” � 22

Build distractor groups with DSL count(PERSON : Murray Gold, John Dubney, Ron Grainer) == 3 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 has_distractor count(token(g, pattern=ENT(g))) 3 and ENT(g) == ENT(p(m)) 4 count(PERSON : Murray Gold) == 1 and f1(m) == 0 5 “There are more tokens matching the ground truth entity type ( ENT(g) ) in the whole c ontext than in the g roundtruth.” 23 �

Build distractor groups with DSL 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 has_distractor count(token(g, pattern=ENT(g))) 3 correct_type and ENT(g) == ENT(p(m)) 4 and f1(m) == 0 ENT(John Debney) == PERSON 5 “The m odel p rediction ENT ity type matches the g roundtruth ENT ity type.” 24 �

Build distractor groups with DSL 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 has_distractor count(token(g, pattern=ENT(g))) 3 correct_type and ENT(g) == ENT(p(m)) 4 is_distracted and f1(m) == 0 5 “The m odel prediction is incorrect.” 25 �

Build distractor groups with DSL 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 has_distractor count(token(g, pattern=ENT(g))) 3 correct_type and ENT(g) == ENT(p(m)) 4 is_distracted and f1(m) == 0 5 Correct Incorrect 5.7% of all BiDAF errors: The distractor hypothesis seems correct! 26 �

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Subjectively defined hypotheses Precise & reproducible h ypotheses + + Small samples Scale up to the entire dev set + + Focus exclusively on errors Focus exclusively on errors Cover errors & correct instances + + Wrongly prioritize groups that are No Test on true cause Test via counterfactual analysis well-handled in average. 27 �

“We randomly select 50-100 incorrect questions (based on EM) and roughly label them into N error groups.” Biased conclusion due to… Biased conclusion due to… Errudite Errudite Subjectively defined hypotheses Precise & reproducible h ypotheses + + Small samples Scale up to the entire dev set + + Focus exclusively on errors Focus exclusively on errors Cover errors & correct instances Cover errors & correct instances + + Wrongly prioritize groups that are No Test on true cause Test via counterfactual analysis well-handled in average. 28 �

Build distractor groups with DSL all_instance 1 ENT(g) != "" is_entity and count(token(c, pattern=ENT(g))) > 2 has_distractor count(token(g, pattern=ENT(g))) 3 correct_type and ENT(g) == ENT(p(m)) 4 is_distracted and f1(m) == 0 5 Correct Incorrect 88% EM > 68% EM: BiDAF performs better when have distractors & entity type is matched, than overall. Reject / revise the hypothesis! 29 �