ACL2011読み会 EXPLOITING WEB-DERIVED SELECTIONAL PREFERENCE TO IMPROVE STATISTICAL DEPENDENCY PARSING

1. ACL2011ಡΈձ EXPLOITING WEB-DERIVED SELECTIONAL PREFERENCE TO IMPROVE STATISTICAL DEPENDENCY PARSING (1156) GUANGYOU ZHOU, JUN ZHAO, KANG LIU, AND LI CAI 2011/09/03 Yuya Unno (@unnonouno) 2011೥9݄3೔౔༵೔
2. ֓ཁ • ܎Γड͚ͷબ୒બ޷ (selectional preference) Λwebσʔλ ͷස౓Ͱදݱ͠·ͨ͠ • Higher-order EisnerΛϕʔεʹɺ܎Γड͚NάϥϜͷPMI Λಛ௃ྔʹೖΕ·ͨ͠ • ਫ਼౓্͕͕ͬͨΑ :-) 2011೥9݄3೔౔༵೔
3. ࠓߋͰ͕͢܎Γड͚ղੳͱ͸ • ୯ޠ΍จઅͷʮ܎Γड͚ؔ܎ʯΛ౰ͯΔ໰୊ • ʮ܎Γड͚ؔ܎ʯ͕Կͳͷ͔͸ӬԕͷṖ • ۃ࿦͢Ε͹ਖ਼ղίʔύεʹ༩͑ΒΕ্ͨهͷ༷ͳߏ଄ 2011೥9݄3೔౔༵೔
4. બ୒બ޷ͱ͸ʁ • ಛఆͷಈࢺ΍໊ࢺʹ͸܎Γ͍҆ಈࢺ΍໊ࢺ͕͋Δ • ʮ…Λ৯΂Δʯʹ͸ɺ৯΂෺͕དྷ΍͍͢ • ʮ෕ͷߴ͍ϫΠϯʯͱʮ஋ஈͷߴ͍ϫΠϯʯͷߏ଄ ͷҧ͍Λॠ࣌ʹ൑அͰ͖Δͷ͸ԿނͩΖ͏ʁ • http://d.hatena.ne.jp/usata3/20070416/p2 2011೥9݄3೔౔༵೔
5. ΞΠσΞɿΑ͘ڞى͢Δ୯ޠؒ ʹબ୒બ޷ੑ͕͋ΔΜͰ͸ʁ • ʮ෕ͷߴ͍ʯʻʮ஋ஈͷߴ͍ʯ • ܎Γड͚ϖΞީิͷ̎୯ޠͷڞى͠΍͢͞Λಛ௃ʹೖ ΕΑ͏ • ैདྷ͸୯ޠͱΫϥεؒͰ΍͍͚ͬͯͨͲɺWebεέʔ ϧͷσʔλ͕͋Ε͹୯ޠͱ୯ޠͰ΋େৎ෉ͩΑͶʂ 2011೥9݄3೔౔༵೔
6. ڞى۩߹͸PMIͰଌΖ͏ • ࣗݾ૬ޓ৘ใྔ (Point-wise Mutual Information) • ̎ͭͷ֬཰ม਺͕ͲΕ͘Β͍ಠཱͰͳ͍͔ • PMI(x, y) = log (P(x, y) / P(x)P(y)) • ۮવͷڞىͰͳ͚Ε͹ߴ͍஋Λ࣋ͭ • ڞىස౓͑͞Θ͔Ε͹ਪఆͰ͖Δ 2011೥9݄3೔౔༵೔
7. ස౓ͷΧ΢ϯτํ๏͸ೋछྨ • Google V1 ίʔύε • Webεέʔϧͷେن໛ίʔύε • Google API • ̎ͭͷΩʔϫʔυΛೖΕͯάάͬͨ࣌ͷස౓ 2011೥9݄3೔౔༵೔
8. ̏୯ޠͷ৔߹΋ߟ͑Δ • ̏୯ޠ΋΍ͬͯΈΔ • PMI(x, y, z) = log ( P(x, y, z) / P(x, y)P(y, z) ) • x, yͱy, zͷPMIͱ͍͏͜ͱ͔ͳɾɾɾ 2011೥9݄3೔౔༵೔
9. PP-ATACHEMENT΋΍ͬͯΈΔ • લஔࢺͷલޙͷ୯ޠͷબ୒બ޷΋ߟ͍͑ͨ • PMIIN(x, z) = P(x, IN, z) / P(x) • ΋͸΍PMIͰ͸ͳ͍ɾɾɾ • ઌͷʮ3୯ޠͷ৔߹ʯͰଌΕΔؾ͕͢Δɾɾɾʁ 2011೥9݄3೔౔༵೔
10. ϕʔε͸ߴ֊EISNERͷLOG- LINEARϞσϧ • Eisner๏ͱ͸ඇަࠩ܎Γड͚ղੳͷղ๏ͷҰͭ • ܎Γड͚෦෼ߏ଄ͷείΞͷ૯࿨Λ࠷େԽͤ͞Δ • CKYΞϧΰϦζϜͱີ઀ͳؔ܎͕͋Δ • ৄ͘͠͸ࢲͷνϡʔτϦΞϧΛࢀর • http://www.slideshare.net/unnonouno/ss-5724050 • PMIಛ௃͸ଞͷόΠφϦಛ௃ͱεέʔϧΛ߹ΘͤΔͨΊɺฏۉɾ෼ࢄͰਖ਼نԽ͢Δ (PMI - μ) / σ Ҏ߱ɺਤ͸͍ͣΕ΋ݪ࿦จΑΓ ρ͸ʮ෕ͷʯ→ʮϫΠϯʯ ͷΑ͏ͳ܎Γड͚෦෼ߏ଄ 2011೥9݄3೔౔༵೔
11. ಛ௃ͷ༗ແͰ݁Ռͷൺֱ • ͍ͣΕʹ͠Ζબ୒બ޷ಛ௃ΛೖΕͨ΄͏͕ྑ͍ • Google APIʹΑΔώοτ਺Λ࢖ͬͨ΄͏͕एׯΑ͍ • dep1: 1֊Eisner, dep2: 2֊Eisner, +hits: Google API, +V1: Google V1 Corpus 2011೥9݄3೔౔༵೔
12. ࠷৽ͷख๏ͱͷൺֱ • ࠷৽ͷ൒ڭࢣ͋Γ ʹ͸ෛ͚Δ • ڭࢣͳ͠σʔλΛ ࢖Θͳ͍ख๏ʹ͸ উͯΔ 2011೥9݄3೔౔༵೔
13. σʔλͷҧ͍͸͋·ΓӨڹ͕ͳ ͍ • ϩΠλʔͱGigword͔Β ࡞ͬͨίʔύε(NEWS)Ͱ ΋࣮ݧͨ͠ • Unigramͷ਺ͷΈʹਫ਼౓͕ ґଘ • σʔλͷछྨʹґଘ͠ͳ ͍Α͏ͩ 2011೥9݄3೔౔༵೔
14. ԕ͍܎Γड͚ͷਫ਼౓޲্͕େ͖ ͍ • ԕ͍܎Γड͚ਫ਼౓͕ ͪΌΜͱ্͕͍ͬͯΔ 2011೥9݄3೔౔༵೔
15. ෼໺దԠ͍ͨ࣌͠΋WEB NGRAMΛೖΕͨํ͕Α͍ • શ͘ͷ෼໺֎ʹྑ͍Өڹ Λ༩͑Δ͔ʁ • BioͷσʔλͰධՁ • ݁Ռతʹɺ෼໺֎ͷσʔ λͷਫ਼౓΋্͕Δ͜ͱ͕ Θ͔ͬͨ 2011೥9݄3೔౔༵೔
16. ॴײ • ಉ͡Α͏ͳ͜ͱ͸Ҏલߟ͑ͯগ࣮͠ݧ͕ͨ͠͏·͘ߦ ͔ͳ͔ͬͨ • ൒ڭࢣ͋ΓΛؚΊɺϥϕϧͳ͠σʔλΛ࢖ͬͯείΞ Λ্͛Δํ޲͕͠͹Β͘ଓ͖ͦ͏ • ҰํͰ܎Γड͚ύε͸ࣗવͰͳ͍͜ͱ΋͋Γɺ͏·͘ ͍͔ͳ͍έʔε΋͋Γͦ͏ 2011೥9݄3೔౔༵೔
17. ·ͱΊ • ܎Γड͚ղੳͷબ୒બ޷ΛWebσʔλͷස౓Ͱදݱ͠·ͨ͠ • ಛ௃ʹೖΕΔͱ͖͸܎Γड͚ύε্ͷྡ઀୯ޠରͷPMIΛಛ ௃ྔͷ஋ʹͯ͠ɺී௨ʹֶश͢Δ͚ͩ • طଘͷख๏ͱൺ΂ͨͱ͖ɺԕ͍܎Γड͚ɺ෼໺֎ͷσʔλɺ ͍ͣΕ΋ྑ޷ͷ݁Ռ • ࠷৽ͷ൒ڭࢣ͋Γʹ͸ෛ͚Δ 2011೥9݄3೔౔༵೔