Group Reading in SNLP2018 - Let’s do it “again”: A First Computational Approach to Detecting Adverbial Presupposition Triggers

Andre Cianflone, Yulan Feng, Jad Kabbara, and Jackie Chi Kit Cheung.
Presented in ACL 2018.
Best Long Papers

Transcript

1. Let’s do it “again”: A First Computational Approach to Detecting Adverbial Presupposition Triggers ಡΉਓ
   দྛ༏Ұ࿠ !࠷ઌ୺/-1ษڧձ   1SFTFOUFE
   BU
   "$-
    FRVBMMZ
   DPOUSJCVUFE
2. ֓ཁ • ͋ΔจʹલఏͷଘࡏΛࣔ͢෭ࢺͷϚʔΧʔ͕ଘࡏ ͢Δ͔Λ༧ଌ͢Δɺͱ͍͏৽͍͠ςετλεΫΛ ࡞ͬͨ • σʔλ΋࡞ͬͨ • "UUFOUJPOΛ༻͍ͨڧ͍ϕʔεϥΠϯ΋࡞ͬͨ – 8FJHIUFE
   1PPMJOH
   BUUFOUJPOPWFSBUUFOUJPO
   
   TFMGBUUFOUJPO • -45.
   ͷ্ʹ৽ͨͳύϥϝʔλΛಋೖ͢Δ͜ͱͳ͘௥ՃͰ ͖ɺਫ਼౓΋্͕Δ • 0OF
   PG
   UIF
   #FTU
   -POH
   1BQFST
   JO
   "$-
    
3. ͕͜͜Ғ͍ • ܭࢉݴޠֶతʹك༗ͳண؟఺ • ͖ͪΜͱίϯτϩʔϧ͞ΕͨλεΫઃܭ – ໰୊ઃܭͷݟຊ • େن໛ͳ࣮ݧσʔλͷ࡞੒ • ߟ͑͏Δ࠷ߴΫϥεͷϕʔεϥΠϯϞσϧΛ࣮૷ ͠ɺͦΕΑΓ΋ऑ͍ෳ਺ͷϞσϧͱൺֱ  "MMJOPOF 1BDLBHF ৽ͨͳํ޲ੑͷఏࣔ ໰୊ͷܗࣜԽ ίϛϡχςΟ΁ͷ ࣮ݧ؀ڥͷ੔උ ڧྗͳϞσϧͷఏҊͱ े෼ͳൺֱ࣮ݧ
4. લఏͷଘࡏΛࣔ͢෭ࢺϚʔΧʔ • UPP
   BHBJO
   BMTP
   TUJMM
   BOE
   ZFU – ࿦จͰ͸ QSFTVQQPTJUJPO USJHHFST ͱݺͿ  John is going to the restaurant again. QSFTVQQPTFT John has been to the restaurant.
5. લఏͷଘࡏΛࣔ͢෭ࢺϚʔΧʔ • ෭ࢺͷલఏϚʔΧʔ͸ఆףࢺʹΑΔલఏ ϚʔΧʔͷ࣍ʹස౓͕ߴ͍ • ܭࢉݴޠֶͰ͸ຆͲݚڀͷલྫ͕ͳ͍ • ͲͷΑ͏ͳίϯςΫετͰ͜ΕΒͷϚʔ Χʔ͕࢖ΘΕΔ͔Λཧղ͍ͨ͠ – ޠ༻࿦తਪ࿦ͷྑ͍ςετʹͳΓ͏Δ – จষੜ੒ʹ͓͚Δ SFBEBCJMJUZ
   ޲্ͷͨΊʹ 
6. ఏҊλεΫɿ 1SFEJDUJOH
   QSFTVQQPTJUJPO USJHHFST • ͋Δจ຺Ͱɺλʔήοτͷޠ͕෭ࢺͷલఏϚʔ ΧʔͰम০͞ΕΔ͔Ͳ͏͔Λೋ஋෼ྨ – લఏͷ෭ࢺ͕෇͘Մೳੑͷ͋Δޠ – λʔήοτͷޠΛத৺ͱͯ͠ɺલ̑̌ޠɺޙΖ͸จ ຤·ͰΛจ຺ͱͯ͠ೖྗ  ...Taped just as the market closed yesterday, it offers Ms. Farrell advising , " We view the market here as going through a relatively normal cycle ... . We continue to feel that the stock market is still the @@@@ place to be for long-term appreciation ೖྗ ࣄલ࣮ݧͰ࠷΋ద੾ͳ௕͞ ʢจ຺ͷτʔΫϯྻɺจ຺ͷ104ྻʣ NNʹλʔήοτΛڭ͑ΔͨΊͷಛघτʔΫϯ ग़ྗ 
   ϚʔΧʔ͕͋ͬͨɺ
   ͳ͔ͬͨ λʔήοτपΓͷલఏͷ෭ࢺΛ ফͯ͋͠ΔจষΛ༻ҙ
7. σʔλ࡞੒ • λʔήοτͷ෭ࢺ – UPP
   BHBJO
   BMTP
   TUJMM
   BOE
   ZFU • ίʔύε – 1FOO
   5SFFCBOL – (JHBXPSE DPSQVT
   TFDUJPOT
    • ̎ͭͷ໰୊ઃఆ – ֤λʔήοτޠ͝ͱͷೋ஋෼ྨ໰୊ (JHBXPSE͚ͩ – ͍ͣΕ͔ͷ෭ࢺ͕ೖ͍ͬͯΔ͔Λ౴͑Δ໰୊ 
8. σʔλ࡞੒ํ๏ • લํจ຺Ҏ֎ͷ෼ྨʹӨڹ͢ΔϑΝΫλʔΛ ίϯτϩʔϧ͍ͨ͠ – ਖ਼ྫͱෛྫͷ਺ – λʔήοτ෭ࢺ͝ͱͷ೉қ౓ – ෭ࢺͷ܎Γઌͷޠʹؔ͢Δ܏޲ • αϯϓϦϯάͷΞϧΰϦζϜ – ·ͣਖ਼ྫΛΈ͚ͭΔ – ෭ࢺͷ܎Γઌͷޠ͕ೖ͍ͬͯͯɺλʔήοτ෭ࢺ Ͱम০͞Ε͍ͯͳ͍ࣄྫΛෛྫʹ͢Δ 
9. αϯϓϦϯάͷΞϧΰϦζϜ • ·ͣਖ਼ྫΛΈ͚ͭΔ • ෭ࢺͷ܎Γઌͷޠ͕ೖ͍ͬͯͯɺλʔήοτ෭ ࢺͰम০͞Ε͍ͯͳ͍ࣄྫΛෛྫʹ͢Δ  ...Taped just as the market closed yesterday, it offers Ms. Farrell advising , " We view the market here as going through a relatively normal cycle ... . We continue to feel that the stock market is still the place to be for long-term appreciation ...More people are remaining independent longer presumably because they are better off physically and financially . Careers count most for the well-to-do many affluent people place personal success and money above family ਖ਼ྫ ෛྫ ಉ͡ޠͰ෭ࢺम০͞Εͯ ͍ͳ͍ྫΛαϯϓϦϯά
10. σʔλͷ౷ܭ • ςετηοτ͸ಛఆͷηΫγϣϯ͔Β࡞੒ – 15# TFD
     – (JHBXPSE TFD
     
11. ղੳϞσϧ XFJHIUFEQPPMJOH
    81
    BSDIJUFDUVSF  w2v attention-over-attention method (Cui et al., 2017) + Self-Attention
12. ղੳϞσϧ XFJHIUFEQPPMJOH
    81
    BSDIJUFDUVSF  w2v ֤୯ޠʹ ॏΈΛ͚ͭͯ ͍ΘΏΔ #J-45.
    ͍ΘΏΔ 4FMG"UUFOUJPO ͍ΘΏΔ୯ޠ FNCFEEJOHͷྻ ଍ͯ͠ɺ ͔ͦ͜ΒFFͰ ೋ஋෼ྨ attention-over-attention method (Cui et al., 2017) + Self-Attention XFJHIUFEQPPMJOH
13. ղੳϞσϧ XFJHIUFEQPPMJOH
    81
    BSDIJUFDUVSF  w2v ! = #$# attention-over-attention method (Cui et al., 2017) + Self-Attention ୯ޠ਺º୯ޠ਺ͷߦྻ ͋ΔΫΤϦʹ ؔ͢Δ ֤όϦϡʔ ͷॏΈ ͋ΔόϦϡʔ ʹؔ͢Δ ֤ΫΤϦͷॏΈ ֤ΫΤϦͷॏΈ ֤όϦϡʔͷॏΈ !% !& ' ( = !)' LSTM͔Βઌ͸৽͍͠ ύϥϝʔλ͕૿͑ͳ͍ 2VFSZ
    7BMVF ॏΈ෇͖τʔΫϯྻ
14. ࣮ݧ • ఏҊλεΫʹ͓͍ͯఏҊϞσϧͷੑೳΛ֬ೝ • ઃఆ – ܇࿅σʔλͷແ࡞ҝΛ։ൃ༻σʔλʹ͢Δ • ൺֱର৅ – ࠷සग़Ϋϥε – ର਺ճؼ෼ྨث XJUI
    CJHSBN – #J-45.
    XJUI
    BWFSBHF
    QPPMJOH – $//
    TFOUFODF
    DMBTTJGJDBUJPO
    NPEFM
    • <,JN
    
    &./-1> 
15. ݁Ռ 
16. ղੳྫ • ਖ਼͘͠෼ྨͰ͖ͨྫ • ෼ྨʹࣦഊͨ͠ྫ  ...Taped just as the market closed yesterday, it offers Ms. Farrell advising , " We view the market here as going through a relatively normal cycle ... . We continue to feel that the stock market is still the place to be for long-term appreciation ...More people are remaining independent longer presumably because they are better off physically and financially . Careers count most for the well-to-do many affluent people place personal success and money above family ਖ਼ྫ ෛྫ TUJMM
    ͕͋ͬͨͱͯ͠΋ಡΊͳ͘͸ͳ͍ ˠ
    ʮજࡏతʹ͋ͬͯ΋Α͍͔ʯͱ ʮ࣮ࡍ͔͋ͬͨʯ͸ผ໰୊
17. ·ͱΊ • ͋ΔจʹલఏͷଘࡏΛࣔ͢෭ࢺͷϚʔΧʔ͕ଘࡏ͢Δ ͔Λ༧ଌ͢Δɺͱ͍͏৽͍͠ςετλεΫΛ࡞ͬͨ • ஸೡͳ໰୊ઃܭ • େن໛ͳσʔλ • 8FJHIUFE
    1PPMJOHΛ࢖ͬͨڧ͍ղੳϞσϧ – "UUFOUJPOPWFSBUUFOUJPO
    
    TFMGBUUFOUJPO – -45.
    ͷ্ʹ৽ͨͳύϥϝʔλΛಋೖ͢Δ͜ͱͳ͘௥Ճ Ͱ͖ɺਫ਼౓΋্͕Δ 
18. ײ૝ʢࠓޙͷల։ͳͲʣ • ྫ͑͹ਓؒͳΒͲͷ͘Β͍༧ଌͰ͖Δ͔ • ࣈ໘্͸ϚʔΧʔ͕ͳ͍͕ɺ৺ͷதͰ͸લఏ Λ࣋ͬͯಡΜͰ͍Δέʔε΋͋Δɻͦͷ͋ͨ ΓΛ্ख͘໰୊ઃఆʹऔΓࠐΊͳ͍͔ɻ • lDPOUJOVF
    UP
    GFFMzͷΑ͏ʹख͔͕ΓϑϨʔ ζ͕͋Δέʔεͱɺͦ͏Ͱ͸ͳ͍೉͍͠έʔ εͷ෼ੳͳͲɺ͍Ζ͍Ζௐ΂Δͱɺʮલఏͷ ಡΈʯΛඥղ͘໘ന͍͜ͱ͕٧·͍ͬͯͦ͏