Pro Yearly is on sale from $80 to $50! »

A Simple Approach to Case-Based Reasoning in Knowledge Bases

C13bf50541b0143c6baa3a14fb1cc381?s=47 Hiroki Ouchi
September 23, 2020

A Simple Approach to Case-Based Reasoning in Knowledge Bases

2019-09-25 第12回最先端NLP勉強会

C13bf50541b0143c6baa3a14fb1cc381?s=128

Hiroki Ouchi

September 23, 2020
Tweet

Transcript

  1. 2020-09-25 @ ୈ12ճ࠷ઌ୺ NLP ษڧձ ಡΉਓɿେ಺ܒथ ʢཧݚʣ ※ͱ͘ʹ஫ऍ͕ͳ͍ݶΓਤද͸࿦จ͔ΒͷҾ༻Ͱ͢ Paper: https://openreview.net/pdf?id=AEY9tRqlU7

    Code: https://github.com/rajarshd/CBR-AKBC
  2. 2020/9/18  ͜ͷ࿦จΛબΜͩཧ༝ n ࣄྫ (Case/Example/Instance) ʹ΋ͱͮ͘ਪ࿦ʹڵຯ Ø ࣗ෼΋ࣅͨݚڀΛ͓ͯ͠Γɼ͜ͷํ޲ੑͷએ఻Λ͍ͯ͠Δ ”Instance-Based

    Learning of Span Representations,” Ouchi+’20 Ø ͜ͷํ޲ੑ͕ࠓޙ૿͍͔͑ͯ͘΋ ”All the reviewers think this is a strong paper and would lay out a solid framework for future work in this direction.” (Meta-Reviewer, OpenReview) n Best Paper Runner-Up (@AKBC) ʹબग़ Ø ϨϏϡʔ఺਺(10఺ຬ఺): 8, 8, 7 From http://www.akbc.ws/2020/awards/
  3. 2020/9/25  Take-Home Message n ࣄྫʹ΋ͱͮ͘ਪ࿦͸໘ന͍ n ະ஌ͷ໰୊(ςετσʔλ)Λղ͘ͱ͖ʹ΋ɼֶशσʔλ Λࢀর͠ͳ͕Βղ͘ͱ͍͍͜ͱ͕͋Δ n

    ͍ΖΜͳλεΫʹԠ༻Մೳ n ࣄྫʹ΋ͱͮ͘ਪ࿦ͷೖ໳ॻ
  4. 2020/9/25  ຊ࿦จͷ·ͱΊ n ࣄྫʹ΋ͱ͍ͮͯਪ࿦ (Case-Based Reasoning) ͢Δ γϯϓϧͳख๏ΛఏҊ n

    Ϟσϧύϥϝʔλͷֶश͕͍Βͳ͍ n 2ͭͷϕϯνϚʔΫσʔληοτ(NELL-995ͱFB-122) Ͱ࠷ߴਫ਼౓Λୡ੒ n ఏҊख๏ʹΑͬͯɼਖ਼ղʹḷΓͭ͘ଟ༷ͳਪ࿦ύεΛ ൃݟͰ͖Δ͜ͱ͕Θ͔ͬͨ
  5. 2020/9/25  Case-Based Reasoning (CBR) ͱ͸ʁ n աڈʹܦݧͨ͠ྨࣅ໰୊ͷղ๏ʹج͍ͮͯ৽ͨͳ໰୊ Λղ͘ਪ࿦ख๏·ͨ͸ͦͷաఔ [Aamodt

    and Plaza, 1994] n ྫʣࣗಈं੔උ࢜͸ɼҎલʹܦݧͨ͜͠ͱͷ͋ΔࣅͨΑ͏ͳ ंͷނোΛࢥ͍ग़ͯ͠ɼंͷमཧΛ͢Δ n ͜Ε·Ͱଟ͘ͷݚڀ͕ೝ஌৺ཧֶ෼໺Ͱͳ͞Ε͖ͯͨ n ଟ͘ͷݹయతͳਓ޻஌ೳͷݚڀͰ΋ CBR Λ AI γεςϜʹ औΓೖΕΔ͜ͱʹ஫ྗ͖ͯͨ͠ [Schank, 1982, Kolodner, 1983, Rissland, 1983, Aamodt and Plaza, 1994, Leake, 1996] n ͓͓·͔ͳΠϝʔδ n Ұൠతͳ෼ྨثϕʔεͷख๏͸ʮ࣋ͪࠐΈෆՄࢼݧʯ Ø ͋Β͔͡Ίֶशσʔλ͔Βඞཁͳ஌ࣝΛهԱ͠ɼ هԱͨ͠஌ࣝΛ༻͍ͯະ஌ͷ໰୊Λղ͘ n CBR ͳͲͷࣄྫϕʔεͷख๏͸ʮ࣋ͪࠐΈՄೳࢼݧʯ Ø ֶशσʔλΛࢀর͠ͳ͕Βະ஌ͷ໰୊Λղ͘
  6. 2020/9/25  Case-Based Reasoning (CBR) ͱ͸ʁ n CBR ͸4ͭͷεςοϓ͔Βߏ੒͞ΕΔ [Aamodt

    and Plaza, 1994] ᶃ Retrieve: ະ஌ͷ໰୊͕༩͑ΒΕͨͱ͖ɼͦΕͱྨࣅ͢Δ ࣄྫΛݕࡧɽ1ͭͷࣄྫ͸໰୊ͱͦͷղ๏͔ΒͳΔͱ͞ΕΔ ᶄ Reuse: ݕࡧͨ͠ࣄྫͷղ๏Λ༩͑ΒΕͨ໰୊΁ద༻ ᶅ Revise: ᶄͷղ๏͕ద༻Ͱ͖ͳ͍৔߹͸ɼղ๏Λվྑɾमਖ਼ ᶆ Retain: ໰୊͕ղ͚ͨΒɼͦͷࣄྫΛ৽ͨͳܦݧͱͯ͠อ࣋ ຊݚڀͷఏҊख๏΋͜ͷ4εςοϓʹґڌ
  7. 2020/9/24  औΓ૊ΉλεΫ n Query Answering (Link Prediction) n ೖྗ:

    ΫΤϦ(eq , rq , ?) n ͜͜Ͱ eq ͸ΤϯςΟςΟɼrq ͸ؔ܎Λද͢ n ग़ྗ: ? ʹ͋ͯ͸·ΔΤϯςΟςΟ rq eq
  8. 2020/9/25  ͳͥ CBR Λ࠾༻͢Δͷ͔ʁ n طଘख๏ͷ໰୊఺ n ΤϯςΟςΟؒͷਪ࿦ϧʔϧΛϞσϧԽ n

    ྫ: ceo(X, Y)⋀ headquatered(Y, Z) ⟹ works_in_city(X, Z) n ͢΂ͯͷਪ࿦ϧʔϧΛϞσϧύϥϝʔλʹຒΊࠐΉͷ͸ࠔ೉ ਐΉํ޲͸Ϟσϧ ͷग़ྗ͢ΔείΞ Ͱຖ࣌ࠁܾΊΔ
  9. 2020/9/25  ͳͥ CBR Λ࠾༻͢Δͷ͔ʁ ਐΉํ޲͸Ϟσϧ ͷग़ྗ͢ΔείΞ Ͱຖ࣌ࠁܾΊΔ n ఏҊख๏ʹΑΔ໰୊఺ͷ؇࿨ࡦ

    n ࣄྫʹ΋ͱͮ͘ϊϯύϥϝτϦοΫΞϓϩʔνΛ࠾༻ n ྨࣅࣄྫ͔Βਪ࿦ϧʔϧΛऔಘͯ͠ར༻ n طଘख๏ͷ໰୊఺ n ΤϯςΟςΟؒͷਪ࿦ϧʔϧΛϞσϧԽ n ྫ: ceo(X, Y)⋀ headquatered(Y, Z) ⟹ works_in_city(X, Z) n ͢΂ͯͷਪ࿦ϧʔϧΛϞσϧύϥϝʔλʹຒΊࠐΉͷ͸ࠔ೉
  10. 2020/9/18  ఏҊख๏ͷ֓ཁ

  11. 2020/9/24  εςοϓᶃ ྨࣅΤϯςΟςΟͷݕࡧ n ΫΤϦ eq ͱྨࣅͷΤϯςΟςΟ e’ Λݕࡧ

    n ͨͩ͠ e’ ͸ؔ܎ rq ͷ outgoing edge Λ࣋ͭ eq e’ ※ΤϯςΟςΟؒྨࣅ౓ͷܭࢉํ๏͸ Section 2.2 Λࢀর
  12.  εςοϓᶄ ਪ࿦ύεͷऔಘ n ྨࣅΤϯςΟςΟ e’ ͱؔ܎ rq ͷΤϯςΟςΟ e’’

    Λ ͭͳ͙ਪ࿦ύεΛऔಘ rq rq e’ e’ e’’ e’’ 2020/9/21
  13.  εςοϓᶅ ਪ࿦ύεΛద༻ n औಘͨ͠ਪ࿦ύεΛΫΤϦ (eq , rq , ?)

    ʹద༻ n ਪ࿦ύεద༻ͷࡍ͸ؔ܎໊ͷจࣈྻ׬શҰகΛ࠾༻ works_in_city(x,z) ⟸ ceo(x,y)∧ headquatered(y,z) MATCH MATCH औಘͨ͠ਪ࿦ύε e’’ eq rq
  14. 2020/9/21  εςοϓᶆ ਪ࿦ύεͷอଘ n ͏·͘ద༻Ͱ͖ͨ(ΫΤϦΤϯςΟςΟ eq ͱؔ܎ rq ͷ

    ΤϯςΟςΟ e’’ ʹḷΓ͍ͭͨ)ਪ࿦ύεΛอଘ
  15. 2020/9/23  ӅΕεςοϓᶇ ౴͑ͷϥϯΩϯά n ਪ࿦ύεΛ࢖ͬͯḷΓ͍ͭͨΤϯςΟςΟΛΧ΢ϯτ • works_in_city(x,z) ⟸ ceo(x,y)

    ∧ headquatered(y,z) Ø z = Seattle Ø Seattle += 1 • works_in_city(x,z) ⟸ ceo(x,y) ∧ based_in(y,z) Ø z = Washington Ø Washington += 1 ɾ ɾ Seattle 7 Washington 2 U.S. 1 New York 1 n Χ΢ϯτͷଟ͍ॱʹΤϯςΟςΟΛϥϯΩϯά ※͜ͷεςοϓ͸ஶऀΒͷҎԼͷهड़Λ΋ͱʹେ಺͕ਪଌ ``we rank the entities based on the number of reasoning paths that lead to them”
  16. 2020/9/24  ࣮ݧɿओͳ݁Ռ Ø ଞͷख๏ΑΓྑ͍݁Ռ͕ಘΒΕͨ n ఏҊख๏ͷੑೳΛݕূ͢ΔͨΊɼඪ४తͳϕϯνϚʔ Ϋσʔληοτ্Ͱطଘख๏ͱൺֱ͢Δ ※3ͭͷσʔληοτͰ࣮ݧΛߦͳ͍ͬͯΔ͕ɼ͜͜Ͱ͸NELL-995ͷ ݁ՌͷΈΛܝࡌ͢ΔɽଞͷσʔληοτͰ΋ࣅͨ܏޲͕ݟΒΕΔ

  17. 2020/9/24  ࣮ݧɿFew-Shot Learning ʹ͓͚Δ݁Ռ n ఏҊख๏͸ֶश͕ෆཁͰɼগ਺ͷྨࣅΤϯςΟςΟʹ ؔ࿈͢Δਪ࿦ύεΛ༻͍ͯ༧ଌ͢Δ Ø ֶशࣄྫ͕ݶΒΕ͍ͯΔؔ܎(few-shot

    relations)ʹ ର͢Δੑೳ͕ྑ͍ͷͰ͸ͳ͍͔ʁ n طଘݚڀ [Lv+’19] ʹ͕͍ͨ͠ɼNELL-995 தͷස౓ 114ະຬͷؔ܎Λ few-shot relations ͱఆ࣮ٛͯ͠ݧ
  18. 2020/9/24  ࣮ݧɿFew-Shot Learning ʹ͓͚Δ݁Ռ Ø ଞͷख๏ΑΓ few-shot relations ʹڧ͍܏޲

    n ఏҊख๏͸ֶश͕ෆཁͰɼগ਺ͷྨࣅΤϯςΟςΟʹ ؔ࿈͢Δਪ࿦ύεΛ༻͍ͯ༧ଌ͢Δ Ø ֶशࣄྫ͕ݶΒΕ͍ͯΔؔ܎(few-shot relations)ʹ ର͢Δੑೳ͕ྑ͍ͷͰ͸ͳ͍͔ʁ
  19. 2020/9/24  ෼ੳɿͳͥ CBR ͷੑೳ͸ྑ͍ͷ͔ʁ n ͻͱͭͷؔ܎͕ଟ༷ͳ;ͨͭͷΤϯςΟςΟΛͭͳ͙ Ø ͨͱ͑ಉؔ͡܎Ͱ΋ҟͳΔਪ࿦ύεΛֶश͢Δඞཁ͕͋Δ GEORGE

    BUSH HOUSE OF REPUBLICANS agent_belongs_to _organization VANCOUVER CANUCKS NHL agent_belongs_to _organization A B X Z Y
  20. 2020/9/24  ෼ੳɿͳͥ CBR ͷੑೳ͸ྑ͍ͷ͔ʁ n CBR ͸ΫΤϦΤϯςΟςΟͱࣅ͍ͯΔΤϯςΟςΟΛ ݕࡧ͠ɼͦͷपลͷਪ࿦ύεΛར༻͢Δ Ø

    ྨࣅΤϯςΟςΟಉ࢜ͳΒɼࣅͨ(͋Δ͍͸ಉ͡)ਪ࿦ύεΛ ࢖͑Δ৔߹͕ൺֱతଟ͍ [େ಺ิ଍] Ø ྨࣅΤϯςΟςΟ͝ͱʹݸผͰਪ࿦͠΍͍͢࿮૊Έʹͳͬͯ ͍Δ͸ͣ (fine-grained contextual encoding) Ø ͜ͷ఺͕(ύϥϝʔλʹ͢΂ͯͷਪ࿦஌ࣝΛຒΊࠐΉܥͷ) طଘϞσϧͱൺ΂ͯ༗རͳ఺ͱߟ͑ΒΕΔ GEORGE BUSH HOUSE OF REPUBLICANS agent_belongs_to _organization A B BARACK OBAMA HOUSE OF REPUBLICANS agent_belongs_to _organization A B ྨࣅ ※͜ͷεϥΠυ͸ Section 3.4 ͷهड़ͷߦؒΛେ಺͕ଟ෼ʹຒΊͨ΋ͷͰ͋Δ
  21. 2020/9/24  ෼ੳɿͳͥ CBR ͷੑೳ͸ྑ͍ͷ͔ʁ n Ծઆɿ͢΂ͯͷਪ࿦஌ࣝ(ύε)ΛϞσϧύϥϝʔλʹ ֮͑ͤ͞Δͷ͸೉͍͠ n ≒

    CBR ͷΑ͏ʹɼ༧ଌʹඞཁͱͳΔਪ࿦ύεΛ໰୊ʹԠͯ͡ దٓݕࡧͯ͠༻͍ͨํ͕൚Խͤ͞΍͍͢ n ݕূํ๏ɿਖ਼ղΤϯςΟςΟʹḷΓͭ͘(Ϟσϧ͕༧ଌ ͨ͠)ਪ࿦ύεͷ਺Λ਺͑ͯൺֱ͢Δ Ø ֮͑ͨਪ࿦஌͕ࣝগͳ͚Ε͹ɼਖ਼ղΤϯςΟςΟʹḷΓͭ͘ ਪ࿦ύε΋ൃݟͰ͖ͳ͍ͱ͍͏લఏ͕͋Δ(ͨͿΜ) [େ಺ิ଍] ※͜ͷεϥΠυ͸ Section 3.4 ͷهड़ͷߦؒΛେ಺͕ଟ෼ʹຒΊͨ΋ͷͰ͋Δ n ݕূ݁Ռ Ø ఏҊख๏(CBR): 306.4 unique paths Ø طଘख๏(MINERVA): 176.83 unique paths n ݁࿦ɿԾઆ͸ࢧ࣋͞Εͨ n ୅දతͳطଘख๏(MINERVA)ΑΓ΋ ఏҊख๏(CBR) ͷ΄͏͕ɼ ਖ਼ղΤϯςΟςΟʹḷΓͭ͘ਪ࿦ύεΛଟ͘ൃݟͰ͖Δ
  22. 2020/9/24  ຊ࿦จͷݶք n ਪ࿦ύεద༻࣌ͷؔ܎໊ͷจࣈྻ׬શҰக Ø ؔ܎ͷ෼ࢄදݱΛ༻͍ͯྨࣅ౓Λܭࢉ͢Δ͜ͱʹ Αͬͯ͜ͷ໰୊Λ؇࿨͍ͨ͠ works_in_city(x,z) ⟸

    ceo(x,y)∧ headquatered(y,z) MATCH MATCH औಘͨ͠ਪ࿦ύε e’’ eq rq
  23. 2020/9/25  ຊ࿦จͷ·ͱΊ & Take-Home Message n γϯϓϧͳ CBR ͷख๏ΛఏҊ

    n 2ͭͷϕϯνϚʔΫσʔληοτͰ࠷ߴਫ਼౓Λୡ੒ n ఏҊख๏ʹΑͬͯɼਖ਼ղʹḷΓͭ͘ଟ༷ͳਪ࿦ύεΛ ൃݟͰ͖Δ͜ͱ͕Θ͔ͬͨ n ࣄྫʹ΋ͱͮ͘ਪ࿦͸໘ന͍ n ະ஌ͷ໰୊(ςετσʔλ)Λղ͘ͱ͖ʹ΋ɼֶशσʔλ Λࢀর͠ͳ͕Βղ͘ͱ͍͍͜ͱ͕͋Δ ·ͱΊ Take-Home Message ஶऀʹΑΔ࣍ճ࡞