Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Distance-Free Modeling of Multi-Predicate Inte...

Yuichiroh
August 24, 2018
0
58

Distance-Free Modeling of Multi-Predicate Interactions in End-to-End Japanese Predicate-Argument Structure Analysis

Yuichiroh Matsubayashi and Kentaro Inui. Presented at COLING 2018.

Capturing interactions among multiple predicate-argument structures (PASs) is a crucial issue in the task of analyzing PAS in Japanese. In this paper, we propose new Japanese PAS analysis models that integrate the label prediction information of arguments in multiple PASs by extending the input and last layers of a standard deep bidirectional recurrent neural network (bi-RNN) model. In these models, using the mechanisms of pooling and attention, we aim to directly capture the potential interactions among multiple PASs, without being disturbed by the word order and distance. Our experiments show that the proposed models improve the prediction accuracy specifically for cases where the predicate and argument are in an indirect dependency relation and achieve a new state of the art in the overall F1 on a standard benchmark corpus.

Yuichiroh

August 24, 2018
Tweet

More Decks by Yuichiroh

See All by Yuichiroh
Beyond Accuracy: Behavioral Testing of NLP Models with CheckList
yuichiroh
0
250
Visually Grounded Neural Syntax Acquisition
yuichiroh
0
240
Group Reading in SNLP2018 - Let’s do it “again”: A First Computational Approach to Detecting Adverbial Presupposition Triggers
yuichiroh
1
270

Featured

See All Featured
Being A Developer After 40
akosma
86
590k
GitHub's CSS Performance
jonrohan
1030
460k
Fashionably flexible responsive web design (full day workshop)
malarkey
404
65k
Writing Fast Ruby
sferik
626
61k
StorybookのUI Testing Handbookを読んだ
zakiyama
26
5.2k
Navigating Team Friction
lara
183
14k
Rebuilding a faster, lazier Slack
samanthasiow
79
8.6k
The Success of Rails: Ensuring Growth for the Next 100 Years
eileencodes
43
6.6k
How STYLIGHT went responsive
nonsquared
95
5.2k
CSS Pre-Processors: Stylus, Less & Sass
bermonpainter
355
29k
It's Worth the Effort
3n
183
27k
For a Future-Friendly Web
brad_frost
175
9.4k

Transcript

  1. %JTUBODF'SFF.PEFMJOHPG.VMUJ1SFEJDBUF *OUFSBDUJPOT JO&OEUP&OE+BQBOFTF 1SFEJDBUF"SHVNFOU4USVDUVSF "OBMZTJT :VJDIJSPI .BUTVCBZBTIJ–‖ BOE,FOUBSP *OVJ–‖ \ZNBUTV

    JOVJ^!FDFJUPIPLVBDKQ –5PIPLV6OJWFSTJUZ ‖3*,&/$FOUFS GPS"EWBODFE*OUFMMJHFODF1SPKFDU 

  2. 1SFEJDBUF"SHVNFOU 4USVDUVSF 1"4 "OBMZTJT 2 !" !# !$ !% !&

    '" '# '$ '% '&

  3. 1SFEJDBUF"SHVNFOU 4USVDUVSF 1"4 "OBMZTJT 3 !" !# !$ !% !&

    DAT NOM

  4. 0WFSWJFX 4 1SFEJDBUF"SHVNFOU 4USVDUVSF 1"4 BOBMZTJT +PJOUQSFEJDUJPOPG NVMUJQMF1"4TJOB TFOUFODF %JTUBODFGSFF

    NPEFMJOHPG JOUFSBDUJPOT 4UBUFPGUIFBSUQFSGPSNBODFJOBTUBOEBSE CFODINBSLDPSQVT !" 0 0 !# 0 0 !$ 1 0 !% 0 0 !& 1 1 ,MBZFS#J3// NVMUJQMF1"4T *OUFSBDUJPOMBZFST !" 0 0 !# 0 0 !$ 1 1 !% 0 0 !& 1 0 ,MBZFS#J3// !" !# !$ !% !& 1SFEJDUJPO GPSFBDI QSFEJDBUF $PNCJOJOH PVUQVUT 1PPMJOH "UUFOUJPO 0 0 0 0 1 1 0 0 1 1 ,MBZFS#J(36 0 0 0 0 1 1 0 0 1 0 ,MBZFS#J(36 0 0 0 0 1 1 0 0 1 0 ,MBZFS#J(36 º … … … … !" 0 !# 0 !$ 0 !% 0 1 !" !# !$ !% !& DAT NOM !& ,MBZFS#J3//

  5. #BDLHSPVOE .BKPS$IBMMFOHFJO+BQBOFTF1"4"OBMZTJT • "SHVNFOU0NJTTJPO JNQMJDJUBSHVNFOUT ౴͑ ͨ ɻ هऀ [ஂ]

    ͕ ਘͶΔ ͱ [ट૬] ͕ "SHNFOU 4USVDUVSFT 4ZOUBDUJD %FQFOEFODZ 5 [reporters]-NOM ask -then [prime minister]-NOM answer-PAST ‘The prime minister answered when the reporters asked.’ DAT DAT NOM NOM • 3FQPSUFSTBTLFE UIFQSJNFNJOJTUFSBRVFTUJPO • 5IFQSJNFNJOJTUFSBOTXFSFE SFQPSUFST Main clause Subordinate clause 8FDBOJNQMJDJUMZLOPXUIFTF BSHVNFOUT

  6. #BDLHSPVOE .BKPS$IBMMFOHFJO+BQBOFTF1"4"OBMZTJT • "SHVNFOU0NJTTJPO JNQMJDJUBSHVNFOUT – -PXQSFEJDUJPOBDDVDBDZJO ' <.BUTVCBZBTIJ *OVJ

     TJOHMFNPEFM> •  GPSBSHVNFOUTJOEJSFDUEFQFOEFODZ SFMBUJPOUPBQSFEJDBUF – PGUIFPCMJHBUPSZBSHVNFOUTBSF FMMJQUJDBM JO+BQBOFTFOFXTQBQFSBSUJDMFT 6 8FDBOOPUJHOPSF

  7. "QQSPBDI .PEFMJOH.VMUJQMF1"4*OUFSBDUJPOT • 6TJOHQSFEJDUJPOGPSFBDI QSFEJDBUF  BSHVNFOU QBJSUPUIFPUIFSQBJST • 1FSTPOXIPBTLFE㱠1FSTPOXIPBOTXFSFE

    • 1FSTPOXIPJTBTLFE 1FSTPOXIPBOTXFSFE • 1FSTPOXIPJTBOTXFSFE 1FSTPOXIPBTLFE 7 ౴͑ ͨ ɻ هऀ [ஂ] ͕ ਘͶΔ ͱ [ट૬] ͕ [reporters]-NOM ask -then [prime minister]-NOM answer-PAST ‘The prime minister answered when the reporters asked.’ DAT DAT NOM NOM

  8. 0VS"QQSPBDI EFFQCJ3// .PEFMJOH*OUFSBDUJPOT PO.VMUJQMF1"4T ʴ 8 1 1 0 0

    0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 Word emb. Predicates Target predicate Input sentence ℎ"," $ … % &'," |), p, w 4PGUNBY MBZFS … … … … % &',, |), p, w %" = 3 %/ = 5 %1 = 7 3, 1 0 3" 0 0 3/ 0 0 31 1 1 34 0 0 35 1 0 36 0 0 ,MBZFS#J(36 ℎ/," $ … ℎ1," $ … *OUFSBDUJPOMBZFST … ℎ1,, $ … ℎ/,, $ … ℎ",, $ ℎ'," ℎ',, … … … … QSFEJDUJPOGPS FBDIQSFEJDBUF 4UBUFPGUIFBSU BQQSPBDI !",$ ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & ℎ",$ *",(,$ *",",$ *",),$ B 100- ℎ",$ & ℎ(,( & … … ℎ(,+ & ,",(,$ ℎ",(,$ - ℎ",",$ - ℎ",),$ - ℎ",$ & ℎ",$ & ℎ",$ & *",(,$ *",",$ *",),$ ℎ",$ .BYQPPMJOH "UUFOUJPO C "55100- ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & !",$ *",(,$ * ",",$ *",),$ !",( … … !",+ ,",$ ℎ′",$ 4FMG"UUFOUJPO D 100-4&-'"55 ℎ",$

  9. 3FMBUFE8PSL • (SJE3//CBTFE NPEFM <0VDIJ > • 1SPQPTFENFUIPE *OQVUTFOUFODF 1SFEJDUJPO

    GPSFBDI QSFEJDBUFT 4FRVFOUJBMMZ QSPQBHBUJOH JOUFSNFEJBUFQSFEJDUJPOJOGPSNBUJPO GPSNVMUJQSFEJDBUFTUISPVHI3//T ˠJOGMVFODFECZXPSEEJTUBODF • -45.-.GPSHFUTMPOHEJTUBODFEFQFOEFODZ <-JO[FO > 9 1 1 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 Word emb. Predicates Target predicate Input sentence ℎ"," % … & '(," |*, p, w 4PGUNBY MBZFS … … … … & '(,- |*, p, w &" = 3 &0 = 5 &2 = 7 4- 1 0 4" 0 0 40 0 0 42 1 1 45 0 0 46 1 0 47 0 0 ,MBZFS#J(36 ℎ0," % … ℎ2," % … *OUFSBDUJPOMBZFST … ℎ2,- % … ℎ0,- % … ℎ",- % ℎ(," ℎ(,- … … … … .VMUJQSFEJDBUF *OQVU !",$ ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & ℎ",$ * ",(,$ * ",",$ * ",),$ B 100- ℎ",$ & ℎ(,( & … … ℎ(,+ & ,",(,$ ℎ",(,$ - ℎ",",$ - ℎ",),$ - ℎ",$ & ℎ",$ & ℎ",$ & * ",(,$ * ",",$ * ",),$ ℎ",$ .BYQPPMJOH "UUFOUJPO C "55100- ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & !",$ * ",(,$ * ",",$ * ",),$ !",( … … !",+ ,",$ ℎ′",$ 4FMG"UUFOUJPO D 100-4&-'"55 ℎ",$ %JSFDUMZBTTPDJBUJOHQSFEJDUJPO JOGPSNBUJPOGPSFBDI QSFE BSH  QBJSVTJOHQPPMJOHBOEBUUFOUJPO BDSPTTQSFEJDBUFT BDSPTT UPLFOT *OQVUTFOUFODF 1SFEJDUJPO GPSFBDI QSFEJDBUFT

  10. 3FMBUFE8PSL • (SJE3//CBTFE NPEFM <0VDIJ > • 1SPQPTFENFUIPE *OQVUTFOUFODF 1SFEJDUJPO

    GPSFBDI QSFEJDBUFT 4FRVFOUJBMMZ QSPQBHBUJOH JOUFSNFEJBUFQSFEJDUJPOJOGPSNBUJPO GPSNVMUJQSFEJDBUFTUISPVHI3//T ˠJOGMVFODFECZXPSEEJTUBODF • -45.-.GPSHFUTMPOHEJTUBODFEFQFOEFODZ <-JO[FO > 10 1 1 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 Word emb. Predicates Target predicate Input sentence ℎ"," % … & '(," |*, p, w 4PGUNBY MBZFS … … … … & '(,- |*, p, w &" = 3 &0 = 5 &2 = 7 4- 1 0 4" 0 0 40 0 0 42 1 1 45 0 0 46 1 0 47 0 0 ,MBZFS#J(36 ℎ0," % … ℎ2," % … *OUFSBDUJPOMBZFST … ℎ2,- % … ℎ0,- % … ℎ",- % ℎ(," ℎ(,- … … … … .VMUJQSFEJDBUF *OQVU !",$ ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & ℎ",$ * ",(,$ * ",",$ * ",),$ B 100- ℎ",$ & ℎ(,( & … … ℎ(,+ & ,",(,$ ℎ",(,$ - ℎ",",$ - ℎ",),$ - ℎ",$ & ℎ",$ & ℎ",$ & * ",(,$ * ",",$ * ",),$ ℎ",$ .BYQPPMJOH "UUFOUJPO C "55100- ℎ",$ & .BYQPPMJOH ℎ",$ & ℎ(,$ & ℎ",$ & ℎ),$ & ℎ",$ & !",$ * ",(,$ * ",",$ * ",),$ !",( … … !",+ ,",$ ℎ′",$ 4FMG"UUFOUJPO D 100-4&-'"55 ℎ",$ %JSFDUMZBTTPDJBUJOHQSFEJDUJPO JOGPSNBUJPOGPSFBDI QSFE BSH  QBJSVTJOHQPPMJOHBOEBUUFOUJPO BDSPTTQSFEJDBUFT BDSPTT UPLFOT *OQVUTFOUFODF 1SFEJDUJPO GPSFBDI QSFEJDBUFT "JNUPDBQUVSFUIFJOUFSBDUJPOT XJUIPVUCFJOHEJTUVSCFECZUIF XPSEPSEFSBOEEJTUBODF

  11. #BTF.PEFM 11 !" !# !$ !% !& !' !( )&

    )% )$ 1SFEJDU BOBSHVNFOUMBCFMTFRVFODF GPSFBDIQSFEJDBUFJOEFQFOEFOUMZ *$,$ *$,% *$,& *$,' *$,( *$,# *$," *%,$ *%,% *%,& *%,' *%,( *%,# *%," *&,$ *&,% *&,& *&,' *&,( *&,# *&," * ∈ NOM, ACC, DAT, NONE

  12. #BTF.PEFM 12 !" !# !$ !% !& !' !( )%

    4PGUNBY MBZFS ,MBZFS #J(36 … … … … 0 0 0 0 0 0 1 *%,$ *%,% *%,& *%,' *%,( *%,# *%," 8PSEFNC 5BSHFUQSFEJDBUF • 4UBDLFECJ3// <;IPVBOE9V > • 5BSHFUQSFEJDBUF QPTJUJPO<)F > • 3FTJEVBMDPOOFDUJPOT <)F  0VDIJ > 1SFEJDUBSHVNFOUMBCFMTFRVFODF GPSFBDIQSFEJDBUFJOEFQFOEFOUMZ 4UBUFPGUIFBSU "SDIUFDUVSF

  13. &YUFOEFE.PEFMT 5XPFYUFOTJPOTPO *OQVUBOEMBTUMBZFST $PNCJOBUJPOTPG QPPMJOHBOEBUUFOUJPO • WBSJBOUT 13 1 1

    0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ"," $ … 4PGUNBY MBZFS %" = 3 %( = 5 %* = 7 ,- ," ,( ,* ,. ,/ ,0 ℎ(," $ … ℎ*," $ … *OUFSBDUJPOMBZFST … ℎ*,- $ … ℎ(,- $ … ℎ",- $ ℎ1," ℎ1,- … … … … 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 .VMUJQSFEJDBUF QPTJUJPOT 4FOUFODF 1SFEJDBUFQPTJUJPO %" = 3, %( = 5, %* = 7 8PSEFNC 5BSHFUQSFEJDBUF "MMQSFEJDBUFT QSFEJDUJPOGPS FBDIQSFEJDBUF 21," 21,( 21,* 21,. 21,/ 21,0 21,-

  14. .VMUJQSFEJDBUFQPTJUJPOT • $POUFYUJOGPSNBUJPO IBECFFOJOQVUJOFBDI UJNFTUFQ • *OGPSNBUJPOPGUIF QSFEJDBUFQSPQBHBUFT UISPVHIEFFQCJ3//T 14

    <;IPV9V > <)F >

  15. .VMUJQSFEJDBUFQPTJUJPOJOQVU 15 1 1 0 0 0 0 1 1

    0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ"," $ … 4PGUNBY MBZFS %& %" %' %( %) %* %+ ℎ'," $ … ℎ(," $ … *OUFSBDUJPOMBZFST … ℎ(,& $ … ℎ',& $ … ℎ",& $ ℎ,," ℎ,,& … … … … 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 .VMUJQSFEJDBUF QPTJUJPOT 4FOUFODF 8PSEFNC 5BSHFUQSFEJDBUF "MMQSFEJDBUFT -,," -,,' -,,( -,,) -,,* -,,+ -,,& "JNUPQSPQBHBUF NVMUJQMFQSFEJDBUF JOGPSNBUJPOVTJOH 3//MBZFST

  16. *OUFSBDUJPOMBZFST 16 1PPMJOH "UUFOUJPO 1 1 0 0 0 0

    1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ",$ % .BYQPPMJOH ℎ&,$ % ℎ&,$ % ℎ&,$ % ℎ',$ % ℎ&,$ % ℎ&,$ (&,&,$ (&,",$ (&,',$ )",&,$ ℎ",&,$ * "UUFOUJPO 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ&,& % … … … ℎ&,+ % ,& ," 1 0 0 0 0 0 1 0 0 0 1 1 0 0 ,MBZFS#J(36 ℎ",$ % º

  17. 1PPMJOH • .PEFMJOHBSHVNFOUTIBSJOHBNPOHQSFEJDBUFT 1FSTPOXIPBTLFE㱠1FSTPOXIPBOTXFSFE 1FSTPOXIPJTBTLFE 1FSTPOXIPBOTXFSFE 1FSTPOXIPJTBOTXFSFE 1FSTPOXIPBTLFE ౴͑ ͨ

    ɻ هऀ [ஂ] ͕ ਘͶΔ ͱ [ट૬] ͕ [reporters]-NOM ask -then [prime minister]-NOM answer-PAST DAT DAT NOM NOM

  18. 1PPMJOH • .PEFMJOHBSHVNFOUTIBSJOHBNPOHQSFEJDBUFT *OUFHSBUJOHMBCFMQSFEJDUJPO GPSFBDIQSFEJDBUF POFBDIUPLFO 1FSTPOXIPBTLFE㱠1FSTPOXIPBOTXFSFE 1FSTPOXIPJTBTLFE 1FSTPOXIPBOTXFSFE 1FSTPOXIPJTBOTXFSFE

    1FSTPOXIPBTLFE 1 1 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 .BYQPPMJOH ℎ",$ ౴͑ ͨ ɻ هऀ [ஂ] ͕ ਘͶΔ ͱ [ट૬] ͕ [reporters]-NOM ask -then [prime minister]-NOM answer-PAST DAT DAT NOM NOM ℎ%,$ & ℎ",$ & ℎ',$ &

  19. "UUFOUJPO • $BQUVSJOHBSFMBUJPOCFUXFFO QSFE BSH QBJST "MVODICPYFBUFO CZUIFTPO JT MJLFMZUPCFNBEF

    CZIJTQBSFOUT "SHVNFOUSFMBUJPOJOBQSFEJDBUF .VMUJQSFEJDBUFSFMBUJPO !",$,% ℎ",$,% ' 1SFEJDBUFQBJS "UUFOUJPO 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ$,$ ( … … … ℎ$,) ( *$ *" 1 0 0 0 0 0 1 0 0 0 1 1 0 0 ,MBZFS#J(36 ℎ",% ( !",$,% ℎ",$,% ' 4FMG"UUFOUJPO 1 0 0 0 0 0 1 0 0 0 1 1 0 0 ,MBZFS#J(36 ℎ$,$ ( … … … ℎ$,) ( *" ℎ",% ( )VNBO UBLFT BCSFBE $PNQBOZ UBLFT BOFXFNQMPZFF

  20. 1PPMJOHº "UUFOUJPO 20 "UUFOUJPOUIFO1PPMJOH (Att-Pool) 1PPMJOHUIFO4FMG"UUFOUJPO (Pool-SelfAtt) 1PPMJOHˠ4FMG"UUFOUJPOPO BTJOHMFVOJGJFETFRVFODF .FSHJOHUIFQSFEJDUJPOTGPSB

    DFSUBJOQSFEJDBUFJOUPBRVFSZ QSFE BSH QSFEJDUJPOˠQPPMJOH "UUFOUJPO 1PPMJOH 1PPMJOH 4FMGBUUFOUJPO !",$,% ℎ",$,% ' ℎ",",% ' ℎ",(,% ' ℎ",% ) ℎ",% ) ℎ",% ) *",$,% *",",% *",(,% ℎ",% 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ$,$ ) … … … ℎ$,+ ) ,$ ," ℎ",% ) 1 0 0 0 0 0 1 0 0 0 1 1 0 0 ,MBZFS#J(36 -$,% -$,% -$,$ … … -$,+ !$,% ℎ′$,% ℎ$,% 1 1 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ",% ) ℎ$,% ) ℎ$,% ) ℎ$,% ) ℎ(,% ) ℎ$,% ) *$,$,% *$,",% *$,(,% ,$ ," ,(

  21. &YQFSJNFOUT • &WBMVBUJOHUIFJNQBDUTPGPVSFYUFOTJPOT – .VMUJQSFEJDBUFJOQVUʢMPʣ – *OUFSBDUJPOMBZFST • Pool, Att-Pool,

    Pool-SelfAtt – #BTFMJOFT • Base, Grid SFJNQMFNFOUBUJPO SelfAtt • $PNQBSJTPOUPSFMBUFEXPSL 21

  22. &YQFSJNFOUBM4FUUJOHT • $PSQVTɿ/"*45 5FYU $PSQVT  – TUBOEBSE53"*/%&75&45EJWJTJPOT<5BJSB > –

       TFOUFODFT • 8PSE7FDUPST – 8PSEWFDWFDUPSTQSFUSBJOFEPO8JLJQFEJB <.BUTVCBZBTIJ *OVJ> – 'JOFUVOFEPOUSBJOJOH • &WBMVBUJPO.FUSJDTɿ' NFBTVSF – "WFSBHFPGNPEFMTGPSFBDITFUUJOH 22

  23. )ZQFSQBSBNFUFST • 4FMFDUFEUPPCUBJOBNBYJNVN'POUIF%&7TFU !",$ !",$ !"," … … !",% &",$

    ℎ′",$ 4FMG"UUFOUJPO ℎ",$ 1 1 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 1 0 0 0 0 0 1 1 0 0 1 0 0 0 ,MBZFS#J(36 ℎ),$ * .BYQPPMJOH ℎ",$ * ℎ",$ * ℎ",$ * ℎ+,$ * ℎ",$ * ,",",$ ,",),$ ,",+,$ 128+128 dim 256+2 dim 256+256 dim 1024 dim 1024+1024 dim 1024 dim ESPQPVU -PTTOFHBUJWFMPHMJLFMJIPPE 0QUJNJ[BUJPO"EBN FBSMZTUPQQJOH ,MBZFST 23

  24. *NQBDUTPGFBDIFYUFOTJPO 24 0WFSBMM • &BDIDPNQPOFOUHJWFTB QPTJUJWFJNQBDU

  25. *NQBDUTPGFBDIFYUFOTJPO 25 0WFSBMM • &BDIDPNQPOFOUHJWFTB QPTJUJWFJNQBDU • "UU1PPMWT1PPM4FMG"UU ˠ1PPM4FMG"UU XPSLTCFUUFS

  26. *NQBDUTPGFBDIFYUFOTJPO 26 0WFSBMM • &BDIDPNQPOFOUHJWFTB QPTJUJWFJNQBDU • "UU1PPMWT1PPM4FMG"UU ˠ1PPM4FMG"UU XPSLTCFUUFS

    • $PNCJOBUJPOTPGUIF DPNQPOFOUTGVSUIFS JNQSPWFUIFQFSGPSNBODF – 1PPMBOE4FMG"UU XPSL DPNQMFNFOUBSJMZXJUIFBDI PUIFS – .11PPM4FMG"UU XPSLTCFTU

  27. $PNQBSJTPOUPSFMBUFEXPSL • MP-PooL-SelfAtt BSDIJWFTUIFTUBUFPG UIFBSUQFSGPSNBODFBT BTJOHMFNPEFM 27 4UBUFPGUIFBSU &OEUP&OENPEFM #FTUNPEFMVTJOH

    TZOUBDUJDUSFFJOQVU 0VSCFTU TJOHMFNPEFM 0WFSBMM *NQMJDJU BSHVNFOUT "SHVNFOUTJO EJSFDU EFQFOEFODZ SFMBUJPO

  28. $PNQBSJTPOUPSFMBUFEXPSL • MP-PooL-SelfAtt BSDIJWFTUIFTUBUFPG UIFBSUQFSGPSNBODFBT BTJOHMFNPEFM • .VDIJNQSPWFNFOUPO QSFEJDUJOHJNQMJDJU BSHVNFOUT

    28 4UBUFPGUIFBSU &OEUP&OENPEFM #FTUNPEFMVTJOH TZOUBDUJDUSFFJOQVU 0VSCFTU TJOHMFNPEFM 0WFSBMM *NQMJDJU BSHVNFOUT "SHVNFOUTJO EJSFDU EFQFOEFODZ SFMBUJPO

  29. $PNQBSJTPOUPSFMBUFEXPSL • MP-PooL-SelfAtt BSDIJWFTUIFTUBUFPG UIFBSUQFSGPSNBODFBT BTJOHMFNPEFM • .VDIJNQSPWFNFOUPO QSFEJDUJOHJNQMJDJU BSHVNFOUT

    • &OTFNCMFNPEFMTIPXT BGVSUIFSJNQSPWFNFOU 29 4UBUFPGUIFBSU &OEUP&OENPEFM #FTUNPEFMVTJOH TZOUBDUJDUSFFJOQVU 0VSCFTU TJOHMFNPEFM 0WFSBMM *NQMJDJU BSHVNFOUT "SHVNFOUTJO EJSFDU EFQFOEFODZ SFMBUJPO

  30. $PODMVTJPO 30 1SFEJDBUF"SHVNFOU 4USVDUVSF 1"4 BOBMZTJT +PJOUQSFEJDUJPOPG NVMUJQMF1"4TJOB TFOUFODF %JTUBODFGSFF

    NPEFMJOHPG JOUFSBDUJPOT $PNCJOJOHFBDIQSPQPTFEDPNQPOFOUT JNQSPWFTUIFQSFEJDUJPOˠ4UBUFPGUIFBSU !" 0 0 !# 0 0 !$ 1 0 !% 0 0 !& 1 1 ,MBZFS#J3// NVMUJQMF1"4T *OUFSBDUJPOMBZFST !" 0 0 !# 0 0 !$ 1 1 !% 0 0 !& 1 0 ,MBZFS#J3// !" !# !$ !% !& 1SFEJDUJPO GPSFBDI QSFEJDBUF $PNCJOJOH PVUQVUT 1PPMJOH "UUFOUJPO 0 0 0 0 1 1 0 0 1 1 ,MBZFS#J(36 0 0 0 0 1 1 0 0 1 0 ,MBZFS#J(36 0 0 0 0 1 1 0 0 1 0 ,MBZFS#J(36 º … … … … !" 0 !# 0 !$ 0 !% 0 1 !" !# !$ !% !& DAT NOM !& ,MBZFS#J3//

  31. 'VUVSF%JSFDUJPOT • 1"4JOUFSBDUJPOT DPOUFYUTIBSJOH BDSPTTNVMUJQMF TFOUFODFT – 4JNQMZ UIFNPEFMDBOCFFYUFOEFECZDPODBUFOBUJOH TFWFSBMTFOUFODFT

    • -FBSOJOHJOUFSBDUJPOTCFUXFFOWFSCBMBOEOPNJOBM QSFEJDBUFT – 7FSCBMBOEOPNJOBMQSFEJDBUFTCFIBWFEJGGFSFOUMZ – -FBSOJOHTFQBSBUFNPEFMTJNQSPWFUIFFOUJSFBDDVSBDZ – )PXDBOUIFTFNPEFMTCFJODPSQPSBUFEJOBOFGGJDJFOUXBZ • 007IBOEMJOH – DIBSBDUFSCBTFEBQQSPBDI#1& DIBSBDUFSDPOWPMVUJPOT • 3PCVTUNPEFMJOHGPSNVMUJQMFEPNBJO – *OQVUTQSFFODPEJOHEPNBJOTQFDJGJDQBSBNFUFST – 3PCVTUNPEFMGPSUIFTFEPNBJOQBSBNFUFST 31

  32. 32

  33. %FUBJMFE"OBMZTJT SelfAtt, MP-SelfAtt, MP-Pool-SelfAtt 33 Instance that only SelfAtt failed

    to predict the answer. stretch respond MP-Pool-SelfAtt MP-SelfAtt SelfAtt thinking Tanigawa thinking Tanigawa thinking Tanigawa thinking Tanigawa thinking Tanigawa thinking Tanigawa stretch respond stretch respond spine ACC stretch little thinking of after , respond Tanigawa . !"#$ %&"'()*+"$ എے spine Λ ACC [৳͹͠ PRED ] stretch গ little [ߟ NOM FALSE ] thinking ͷ of ޙ after (nominal) ɺ , Ԡ͡Δ respond [୩઒ NOM GOLD ] Tanigawa ɻ . ‘[Tanigawa NOM GOLD ], responding after briefly [thinking NOM FALSE ] and [stretching PRED ] his spine.’ େֶ university ڭत professor [Β NOM GOLD ] PLURAL ͕ NOM ஍Լ underground ʹ DAT જͬͯ dive ɺ , ϑΝΫε fax Ͱ by ࿈བྷ contacting Λ AC ׆ಈ activitZ Λ ACC [ࢧ͑ͯ PRED ] support ͍Δ PROGRESSIVE ɻ [NONENOM FALSE ] ‘The university [professors NOM GOLD ] went into hiding and are [supporting PRED ] the activities of the local organ தԝ central [লி NOM FALSE ] ministries ͕ NOM [৬һ NOM GOLD ] staff ʹ DAT ର͠ against ʮ “ ໷ night ͷ in ઀଴ entertainment ͸ TOP ͢Ε͹ VERBALIZERCONDITIONAL ࡁΉ finish ͜ͱ NOMINALIZER ͩ COPULA. ɻ ‘It is sufficient enough if the central [ministries NOM FALSE ] tell the [staffNOM GOLD ] “Do not [accept PRED ] a busin ेࡾ 13 ೔ day ޕޙ afternoon ʹ DAT ͸ TOP ຍ࠿ bonsai ࡞Ո artist ɺ , ໦ଜ Kimura ਖ਼඙ Masahiko [ࢯ NOM GOLD ] Mr. ʹ {by- ΑΔ -by} ࣮ٕ tec ΕΔ PASSIVE ɻ . [NONENOM FALSE ] • 5IFQSPQPTFENPEFMT TVDDFTTGVMMZDBQUVSF UIFBSHVNFOUTIBSJOH SFMBUJPOT NPSFEJGGJDVMUQBJST FBTZUPQSFEJDUQBJST

  34. 34 എے spine Λ ACC [৳͹͠ PRED ] stretch গ

    little [ߟ NOM FALSE ] thinking ͷ of ޙ after (nominal) ɺ , Ԡ͡Δ respond [୩઒ NOM GOLD ] Tanigawa ɻ . ‘[Tanigawa NOM GOLD ], responding after briefly [thinking NOM FALSE ] and [stretching PRED ] his spine.’ େֶ university ڭत professor [Β NOM GOLD ] PLURAL ͕ NOM ஍Լ underground ʹ DAT જͬͯ dive ɺ , ϑΝΫε fax Ͱ by ࿈བྷ contacting Λ ACC औΓ take ߹͍ each-other ɺ , ஍ํ district ͷ of ૊৫ organization ͷ of ׆ಈ activitZ Λ ACC [ࢧ͑ͯ PRED ] support ͍Δ PROGRESSIVE ɻ [NONENOM FALSE ] ‘The university [professors NOM GOLD ] went into hiding and are [supporting PRED ] the activities of the local organizations, contacting each other by fax. [NONE NOM FALSE ] தԝ central [লி NOM FALSE ] ministries ͕ NOM [৬һ NOM GOLD ] staff ʹ DAT ର͠ against ʮ “ ໷ night ͷ in ઀଴ entertainment ͸ TOP [ड͚Δ PRED ] accept ͳ NEGIMPERATIVE ʯ ” ͱ as ௨ୡ notification ͢Ε͹ VERBALIZERCONDITIONAL ࡁΉ finish ͜ͱ NOMINALIZER ͩ COPULA. ɻ ‘It is sufficient enough if the central [ministries NOM FALSE ] tell the [staffNOM GOLD ] “Do not [accept PRED ] a business dinner.”’ ेࡾ 13 ೔ day ޕޙ afternoon ʹ DAT ͸ TOP ຍ࠿ bonsai ࡞Ո artist ɺ , ໦ଜ Kimura ਖ਼඙ Masahiko [ࢯ NOM GOLD ] Mr. ʹ {by- ΑΔ -by} ࣮ٕ technique ͷ of σϞϯετϨʔγϣϯ demonstration ͕ NOM [ߦΘ PRED perform ΕΔ PASSIVE ɻ . [NONENOM FALSE ] ‘On the afternoon of 13th, a practical demonstration by the bonsai artist [Mr. NOM GOLD ] Masahiko Kimura will be [performed PRED ]. [NONE NOM FALSE ]’ %FUBJMFE"OBMZTJT SelfAtt, MP-SelfAtt, MP-Pool-SelfAtt Instance that all the models failed to predict the answer. ‘[Tanigawa NOM GOLD ], responding after briefly [thinking NOM FALSE ] and [stretching PRED ] his spine.’ େֶ university ڭत professor [Β NOM GOLD ] PLURAL ͕ NOM ஍Լ underground ʹ DAT જͬͯ dive ɺ , ϑΝΫε fax Ͱ by ࿈བྷ contacting Λ ACC औΓ take ߹͍ eac ׆ಈ activitZ Λ ACC [ࢧ͑ͯ PRED ] support ͍Δ PROGRESSIVE ɻ [NONENOM FALSE ] ‘The university [professors NOM GOLD ] went into hiding and are [supporting PRED ] the activities of the local organizations, cont தԝ central [লி NOM FALSE ] ministries ͕ NOM [৬һ NOM GOLD ] staff ʹ DAT ର͠ against ʮ “ ໷ night ͷ in ઀଴ entertainment ͸ TOP [ड͚Δ PRED ] accept ͢Ε͹ VERBALIZERCONDITIONAL ࡁΉ finish ͜ͱ NOMINALIZER ͩ COPULA. ɻ ‘It is sufficient enough if the central [ministries NOM FALSE ] tell the [staffNOM GOLD ] “Do not [accept PRED ] a business dinner.”’ ेࡾ 13 ೔ day ޕޙ afternoon ʹ DAT ͸ TOP ຍ࠿ bonsai ࡞Ո artist ɺ , ໦ଜ Kimura ਖ਼඙ Masahiko [ࢯ NOM GOLD ] Mr. ʹ {by- ΑΔ -by} ࣮ٕ technique ͷ of ΕΔ PASSIVE ɻ . [NONENOM FALSE ] ‘On the afternoon of 13th, a practical demonstration by the bonsai artist [Mr. NOM GOLD ] Masahiko Kimura will be [performed accept !"#$ central TOP ministries NOM staff DAT against “ ” night in entertainment accept NEG IMPERATIVE as notification VERBALIZER NOMINALIZER COPULA . finish %&"'()*+"$ notification + VERBALIZER finish ministries staff ministries staff ministries staff Dialog-level relational knowledge is needed • NOM of accept should be the same as the persons who received the order from the ministries.

  35. 35 • Pay little attention to Mr. (head of the

    phrase including person name “Masahiko Kimura”) which is the nominative of “demonstraition “ େֶ university ڭत professor [Β NOM GOLD ] PLURAL ͕ NOM ஍Լ underground ʹ DAT જͬͯ dive ɺ , ϑΝΫε fax Ͱ by ࿈བྷ contacting Λ ACC औΓ take ߹͍ each-other ɺ , ஍ํ district ͷ of ૊৫ organization ͷ of ׆ಈ activitZ Λ ACC [ࢧ͑ͯ PRED ] support ͍Δ PROGRESSIVE ɻ [NONENOM FALSE ] ‘The university [professors NOM GOLD ] went into hiding and are [supporting PRED ] the activities of the local organizations, contacting each other by fax. [NONE NOM FALSE ]’ தԝ central [লி NOM FALSE ] ministries ͕ NOM [৬һ NOM GOLD ] staff ʹ DAT ର͠ against ʮ “ ໷ night ͷ in ઀଴ entertainment ͸ TOP [ड͚Δ PRED ] accept ͳ NEGIMPERATIVE ʯ ” ͱ as ௨ୡ notification ͢Ε͹ VERBALIZERCONDITIONAL ࡁΉ finish ͜ͱ NOMINALIZER ͩ COPULA. ɻ ‘It is sufficient enough if the central [ministries NOM FALSE ] tell the [staffNOM GOLD ] “Do not [accept PRED ] a business dinner.”’ ेࡾ 13 ೔ day ޕޙ afternoon ʹ DAT ͸ TOP ຍ࠿ bonsai ࡞Ո artist ɺ , ໦ଜ Kimura ਖ਼඙ Masahiko [ࢯ NOM GOLD ] Mr. ʹ {by- ΑΔ -by} ࣮ٕ technique ͷ of σϞϯετϨʔγϣϯ demonstration ͕ NOM [ߦΘ PRED ] perform ΕΔ PASSIVE ɻ . [NONENOM FALSE ] ‘On the afternoon of 13th, a practical demonstration by the bonsai artist [Mr. NOM GOLD ] Masahiko Kimura will be [performed PRED ]. [NONE NOM FALSE ]’ େֶ university ڭत professor [Β NOM GOLD ] PLURAL ͕ NOM ஍Լ underground ʹ DAT જͬͯ dive ɺ , ϑΝΫε fax Ͱ by ࿈བྷ contacting Λ ACC औΓ take ߹͍ each-other ɺ , ஍ํ district ͷ of ૊ o ׆ಈ activitZ Λ ACC [ࢧ͑ͯ PRED ] support ͍Δ PROGRESSIVE ɻ [NONENOM FALSE ] ‘The university [professors NOM GOLD ] went into hiding and are [supporting PRED ] the activities of the local organizations, contacting each other by fax. [NO தԝ central [লி NOM FALSE ] ministries ͕ NOM [৬һ NOM GOLD ] staff ʹ DAT ର͠ against ʮ “ ໷ night ͷ in ઀଴ entertainment ͸ TOP [ड͚Δ PRED ] accept ͳ NEGIMPERATIVE ʯ ” ͱ as ͢Ε͹ VERBALIZERCONDITIONAL ࡁΉ finish ͜ͱ NOMINALIZER ͩ COPULA. ɻ ‘It is sufficient enough if the central [ministries NOM FALSE ] tell the [staffNOM GOLD ] “Do not [accept PRED ] a business dinner.”’ ेࡾ 13 ೔ day ޕޙ afternoon ʹ DAT ͸ TOP ຍ࠿ bonsai ࡞Ո artist ɺ , ໦ଜ Kimura ਖ਼඙ Masahiko [ࢯ NOM GOLD ] Mr. ʹ {by- ΑΔ -by} ࣮ٕ technique ͷ of σϞϯετϨʔγϣϯ demonstration ͕ NO ΕΔ PASSIVE ɻ . [NONENOM FALSE ] ‘On the afternoon of 13th, a practical demonstration by the bonsai artist [Mr. NOM GOLD ] Masahiko Kimura will be [performed PRED ]. [NONE NOM FALSE ]’ %FUBJMFE"OBMZTJT SelfAtt, MP-SelfAtt, MP-Pool-SelfAtt !"#$%&'(#) perform artist Masahiko Mr. technique 13 day TOP afternoon DAT bonsai artist -by} , Masahiko Kimura Mr. {by- technique PASSIVE NOM . demonstration *#+) of perform Need to model interactions with mominal predicates Instance that all the models failed to predict the answer.