scott wen en tau au yih
play

Scott Wen en-tau au Yih Who is Justin Biebers sister? Jazmyn - PowerPoint PPT Presentation

Scott Wen en-tau au Yih Who is Justin Biebers sister? Jazmyn Bieber semantic parsing Knowledge . sister_of(justin_bieber, ) Base query matching . sibling_of(justin_bieber, x ) gender( x , female) Who is Justin


  1. Scott Wen en-tau au Yih

  2. Who is Justin Bieber’s sister? Jazmyn Bieber semantic parsing Knowledge πœ‡π‘¦. sister_of(justin_bieber, 𝑦) Base query matching πœ‡π‘¦. sibling_of(justin_bieber, x ) ∧ gender( x , female)

  3. Who is Justin Bieber’s sister? Jazmyn Bieber Knowledge semantic parsing Base query πœ‡π‘¦. sibling_of(justin_bieber, x ) ∧ gender( x , female)

  4. β€œWhat was the date that Minnesota became a state?” β€œWhen was the state Minnesota created?” β€œMinnesota's date it entered the union?” location.dated_location.date_founded

  5. Basic idea directly grows staged

  6. Addresses Key Challenges 52.5

  7. β€’ Introduction β€’ Background β€’ Graph knowledge base β€’ Query graph

  8. 12/26/1999 β€’ from Mila Kunis β€’ cvt1 Meg Griffin β€’ Family Guy Lacey Chabert cvt2 series cvt3 1/31/1999

  9. constraints Meg Griffin argmin topic entity core inferential chain Family Guy cast y x

  10. β€’ Introduction β€’ Background β€’ Staged Query Graph Generation (Our Approach) β€’ Link topic entity β€’ Identify core inferential chain β€’ Augment constraints

  11. Staged Meg Family Guy (1) Link Topic Entity s 1 Family Guy s 0 Ο• s 2 Meg Griffin

  12. Staged Meg Family Guy (2) Identify Core Inferential Chain s 3 Family Guy cast actor y x s 1 s 4 Family Guy Family Guy writer start y x s 5 Family Guy genre x

  13. Staged Meg Family Guy (3) Augment Constraints s 3 Family Guy cast actor y x s 6 Meg Griffin Family Guy cast actor y x s 7 argmin Meg Griffin Family Guy y x

  14. s 1 Family Guy s 0 Ο• s 2 Meg Griffin

  15. s 3 Family Guy cast actor y x s 1 s 4 Family Guy Family Guy writer start y x s 5 Family Guy genre x Who first voiced Meg on Family Guy? { castβˆ’actor, writerβˆ’start, genre }

  16. β€’ Input is mapped to two 𝑙 -dimensional vectors 300 exp cos(𝑧 𝑆 , 𝑧 𝑄 ) 𝑄 𝑆 𝑄 = 𝑆 β€² exp cos(𝑧 𝑆 β€² , 𝑧 𝑄 ) 300 ... 𝑧 𝑄 ∈ R 𝑙 𝑧 𝑆 ∈ R 𝑙 ... max max max ... ... ... ... 1000 1000 1000 ... 15K 15K 15K 15K 15K castβˆ’actor who voiced meg on 𝑓 <s> w 1 w 2 w T </s>

  17. s 3 Family Guy cast actor y x s 6 Meg Griffin β€’ Who voiced Family Guy Family Guy cast actor y x s 7 argmin Meg Griffin s 3 Family Guy cast actor Family Guy y x y x cast FamilyGuy 𝑧 actor 𝑧 𝑦 β€’ One or more constraint nodes can be added to 𝑧 or 𝑦 β€’ 𝑧 : Additional property of this event (e.g., character 𝑧 MegGriffin ) β€’ 𝑦 : Additional property of the answer entity (e.g., gender)

  18. Who first voiced Meg on Family Guy? s 3 Family Guy cast actor y x s 4 Family Guy writer start y x

  19. Who first voiced Meg on Family Guy? s 7 argmin Meg Griffin Family Guy y x s 3 Family Guy cast actor y x

  20. π‘Ÿ = Who first voiced Meg on Family Guy? Meg Griffin argmin 𝑑 = Family Guy cast y x

  21. β€’ Introduction β€’ Background β€’ Staged Query Graph Generation (Our Approach) β€’ Experiments β€’ Data & evaluation metric β€’ Creating training data from Q/A pairs β€’ Results

  22. β€’ What character did Natalie Portman play in Star Wars? Padme Amidala β€’ What currency do you use in Costa Rica? Costa Rican colon β€’ What did Obama study in school? political science β€’ What do Michelle Obama do for a living? writer, lawyer β€’ What killed Sammy Davis Jr? throat cancer [Examples from Berant]

  23. Relation Matching (Identifying Core Inferential Chain) Pattern Inferential Chain what was <e> known for people.person.profession what kind of government does <e> have location.country.form_of_government what year were the <e> established sports.sports_team.founded what city was <e> born in people.person.place_of_birth what did <e> die from people.deceased_person.cause_of_death who married <e> people.person.spouse_s people.marriage.spouse

  24. Reward Function 𝛿

  25. Avg. F1 (Accuracy) on WebQuestions Test Set 60 52.5 50 45.3 44.3 41.3 39.9 39.2 40 37.5 35.7 33 30 20 10 0 Yao-14 Berant-13 Bao-14 Bordes-14b Berant-14 Yang-14 Yao-15 Wang-14 Yih-15

  26. Method #Entities Covered Ques. Labeled Ent. Freebase API 19,485 98.8% 81.2% Yang & Chang, ACL-15 9,147 99.8% 87.8% 52.5% 48.4%

  27. 49.6 52.5

  28. A random sample of 100 incorrectly answered questions

  29. directly

  30. http://aka.ms/sent2vec http://aka.ms/codalab-webq http://aka.ms/stagg

Recommend


More recommend