$30 off During Our Annual Pro Sale. View Details »

20190407 第7章 事例研究:自然言語処理における素性

20190407 第7章 事例研究:自然言語処理における素性

本郷言語処理勉強会 #3 で用いた資料。
https://holy.connpass.com/event/122295/
「ゴールドバーグ, 『自然言語処理のための深層学習』, 共立出版, 2019」輪読会。

kmotohas

April 07, 2019
Tweet

More Decks by kmotohas

Other Decks in Technology

Transcript

  1. !LNPUPIBT
    ຊڷݴޠॲཧษڧձ
    ୈ7ষ ࣄྫݚڀɿࣗવݴޠॲཧʹ͓͚Δૉੑ

    View Slide

  2. ࣗݾ঺հ
    !2
    ຊڮ࿨و !LNPUPIBT

    ‣ ౦޻େجૅ෺ཧֶઐ߈1I%

    ‣ ιϑτόϯΫגࣜձࣾ

    ‣ %&&1$03&*OD

    ‣ εΧΠϚΠϯυגࣜձࣾ

    w ਂ૚ֶशϞσϧͷ։ൃɾӡ༻ϓϥοτϑΥʔϜl4,*-z࡞ͬͯΔձࣾ
    w +BWB4DBMBͰͰ͖ΔEFFQMFBSOJOHKͷ։ൃ
    w ,FSBTOE UI
    DPOUSJCVUPST͕͍ͨΓ͢Δ
    w ๻͸ϓϦηʔϧεΤϯδχΞ

    View Slide

  3. ໨࣍
    !3
    ୈ̎ฤɹࣗવݴޠσʔλͷѻ͍
    ‣ ୈ̒ষɹςΩετσʔλͷͨΊͷૉੑ
    ‣ ୈ̓ষɹࣄྫݚڀɿࣗવݴޠॲཧʹ͓͚Δૉੑ
    w จॻ෼ྨɿݴޠಉఆ
    w จॻ෼ྨɿτϐοΫ෼ྨ
    w จॻ෼ྨɿஶऀಛఆ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠɿ඼ࢺλά෇͚
    w จ຺ʹຒΊࠐ·Εͨ୯ޠɿݻ༗දݱೝࣝ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠͱݴޠֶతૉੑɿલஔࢺҙຯᐆດੑղফ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠͷؒͷؔ܎ɿΞʔΫΛ୯Ґͱͨ͠ύʔδϯά

    View Slide

  4. จॻ෼ྨɿݴޠಉఆ
    !4
    จࣈόΠάϥϜͷόοάʢCBHPGMFUUFSCJHSBNTʣ͕ڧྗ
    @btsmith #nlp
    ▪ Character n-gram frequencies for English
    Language Identification
    28
    e 12.6%
    t 9.1%
    a 8.0%
    o 7.6%
    i 6.9%
    n 6.9%
    s 6.3%
    h 6.2%

    th 3.9%
    he 3.7%
    in 2.3%
    er 2.2%
    an 2.1%
    re 1.7%
    nd 1.6%
    on 1.4%

    the 3.5%
    and 1.6%
    ing 1.1%
    her 0.8%
    hat 0.7%
    his 0.6%
    tha 0.6%
    ere 0.6%

    From Cryptograms.org, derived from English documents at Project Gutenberg
    https://www.slideshare.net/LithiumTech/lightweight-natural-language-processing-nlp

    View Slide

  5. จॻ෼ྨɿݴޠಉఆ
    !5
    จࣈූ߸ԽํࣜಉఆʢFODPEJOHEFUFDUJPOʣʹ͸όΠτόΠάϥϜͷόοά͕༗ޮ
    Figure 2: Byte-based method vs. character-based method
    – ISO-2022-{JP,KR} [ja,ko]
    – UTF-8 [universal]
    or characters (unigram
    models can use two or
    ters (bigrams, trigram
    parameter space is exp
    between the accuracy
    ing, computation and s
    tant as the size of the
    Asian charsets with ch
    3.3 Algorithm
    Our first choice was N
    http://cs229.stanford.edu/proj2007/KimPark-AutomaticDetectionOfCharacterEncodingAndLanguages.pdf

    View Slide

  6. จॻ෼ྨɿτϐοΫ෼ྨ
    !6
    จষ㱨\ܦࡁ੓࣏εϙʔπϨδϟʔΰγοϓੜ׆ͦͷଞ^ʁ
    ҙຯతͳجຊతͳ୯Ґ͸୯ޠ

    ʢޠॱ͸͋·ΓҙຯΛ࣋ͨͳ͍ʣ
    ‣ ୯ޠͷόοά CBDLPGXPSET

    ‣ ୯ޠόΠάϥϜͷόοά CBHPGXPSECJHSBNT

    w ϓϥεͰɺϨϯϚԽɺ୯ޠຒΊࠐΈϕΫτϧɺ5'*%'ॏΈ෇͚ͳͲ΋ޮՌ͋Γ
    https://aylien.com/text-api/text-classification/

    View Slide

  7. จॻ෼ྨɿஶऀಛఆ
    !7
    ஶऀ໊ɾੑผɾ೥ྸɾ฼ޠͳͲɺஶऀͷಛ௃Λਪఆ
    จষͷελΠϧʹؔ͢Δಛ௃͕༗༻
    ‣ ඼ࢺλάͱػೳޠ GVODUJPOXPSE

    w ػೳޠɿPO PG UIF BOE CFGPSF ʜIF TIF * UIFZ ʜ
    w ͦΕࣗ਎͸಺༰Λ఻͑ͣɺ಺༰Λ఻͑Δ୯ޠͱ݁ͼ͍ͭͯҙຯΛׂΓ౰ͯΔ
    w େن໛ίʔύεͷ࠶සग़୯ޠ্Ґޠఔ౓͕ۙࣅతʹػೳޠͷϦετʹͳΔ
    w ͦΕͧΕͷCJHSBN USJHSBN HSBN ػೳޠͷີ౓ͳͲ͕࢖͑Δ

    View Slide

  8. จॻ෼ྨɿஶऀಛఆ
    !8
    ྫɿl#VJMEJOH.BDIJOF-FBSOJOH4ZTUFNTXJUI1ZUIPOzͷ֤ষͷஶऀಛఆ
    ‣ CZ8JMMJ3JDIFSUBOE-VJT1FESP$PFMIP
    ڭࢣͳֶ͠शͷख๏Λར༻ͯ͠શষͷஶऀಛఆͰશ໰ਖ਼ղ
    ‣ Ұจ͋ͨΓͷฏۉ୯ޠ਺
    ‣ Ұจͷ௕͞ͷ෼ࢄ
    ‣ ޠኮ਺
    ‣ Ұจ͋ͨΓͷΧϯϚɺηϛίϩϯɺίϩϯͷฏۉ਺
    ‣ ػೳޠͷόοά
    ‣ ֤඼ࢺλάͷग़ݱׂ߹
    ͜ΕΒͷಛ௃ྔΛ࢖ͬͯLNFBOT L
    ΫϥελϦϯά
    http://www.aicbt.com/authorship-attribution/

    View Slide

  9. จ຺ʹຒΊࠐ·Εͨ୯ޠɿ඼ࢺλά෇͚
    !9
    ඼ࢺλά෇͚ QBSUTPGTQFFDIUBHHJOH

    w ྫ 6OJWFSTBM5SFFCBOL1SPKFDU
    ܗ༰ࢺɺ઀ஔࢺɺ෭ࢺɺॿಈࢺɺ౳Ґ઀ଓࢺɺݶఆࢺɺؒ౤ࢺɺ
    ໊ࢺɺ਺ࢺɺෆมԽࢺɺ୅໊ࢺɺݻ༗໊ࢺɺ۟ಡ఺ɺैଐ઀ଓࢺɺه߸ɺಈࢺɺͦͷଞ
    ‣ ߏ଄໰୊ɺ·ͨ͸྆ଆ̎୯ޠͷ૭ʹ͓͚Δ඼ࢺλά෼ྨͷλεΫʹۙࣅ
    w JOUSJOTJDʢ୯ޠͦΕࣗମʹجͮ͘ʣख͕͔Γ
    ୯ޠͦΕࣗ਎ɺ઀಄ࣙɺ઀ඌࣙɺܗঢ়ʢFEɺVOɺେจࣈʣɺग़ݱස౓
    w FYUSJOTJDʢͦͷจ຺ʹجͮ͘ʣख͕͔Γ
    पΓͷ୯ޠͷɺ୯ޠͦΕࣗ਎ɺ઀಄ࣙɺ઀ඌࣙɺલޙͷ඼ࢺ༧ଌ݁Ռ

    View Slide

  10. จ຺ʹຒΊࠐ·Εͨ୯ޠɿݻ༗දݱೝࣝ
    !10
    タスク 説明: 固有表現抽出
    • テキスト 中から人名や組織名 ような固有表現 (NE) を抽出する
    • 固有表現 定義 ほしい情報に基づいて柔軟に定義する
    • 情報抽出や検索エンジン インデキシングなどで有用
    https://explosion.ai/demos/displacy-ent 5

    https://speakerdeck.com/himkt/neural-named-entity-recognition

    View Slide

  11. จ຺ʹຒΊࠐ·Εͨ୯ޠɿݻ༗දݱೝࣝ
    !11
    ܥྻϥϕϦϯά໰୊ TFRVFODFMBCFMJOHUBTL
    ͱͯ͠ఆࣜԽ
    ‣ ϥϕϧ෇͚ͷεΩʔϚ#*0λά #FHJO*OTJEF0VUTJEFFODPEFEUBH

    $"4& 456%*&4 0' /-1 '&"563&4
    ɩF TFOUFODF BCPWF XPVME CF UBHHFE BT
    CQ?Mf"@S1_ aKBi?fA@S1_ -fP T`2bB/2MifP Q7fP J+*Q`KBFf"@P_: AM/mbi`B2bfA@P_:
    pBbBi2/fP ?BbfP MB2+2fP S`Bbf"@S1_ BMfP JBHMf"@GP* -fP `2TQ`i2`bfP bvfP XfP
    ɩF USBOTMBUJPO GSPN OPOPWFSMBQQJOH TFHNFOUT UP #*0 UBHT BOE CBDL JT TUSBJHIUGPSXBSE
    -JLF 104UBHHJOH UIF /&3 UBTL JT B TUSVDUVSFE POF BT UBHHJOH EFDJTJPOT GPS EJĊFSFOU
    XPSET JOUFSBDU XJUI FBDI PUIFS JU JT NPSF MJLFMZ UP SFNBJO XJUIJO UIF TBNF FOUJUZ UZQF UIBO UP
    TXJUDI JU JT NPSF MJLFMZ UP UBH i+PIO 4NJUI *ODw BT "@P_: A@P_: A@P_: UIBO BT "@S1_ A@S1_
    "@P_:
    )PXFWFS XF BHBJO BTTVNF JU DBO CF BQQSPYJNBUFE SFBTPOBCMZ XFMM VTJOH JOEFQFOEFOU
    DMBTTJmDBUJPO EFDJTJPOT
    ɩF DPSF GFBUVSF TFU GPS UIF /&3 UBTL JT TJNJMBS UP UIBU PG UIF 104UBHHJOH UBTL BOE SFMJFT
    PO XPSET XJUIJO B XPSET XJOEPX UP FBDI TJEF PG UIF GPDVT XPSE *O BEEJUJPO UP UIF GFBUVSFT PG
    FOUJUJFT TVDI BT .JMBO +PIO 4NJUI .D$PSNJL *OEVTUSJFT BOE 1BSJT BT XFMM BT UP DBUFHPSJ[F UIFN
    JOUP B QSFEFmOFE TFU PG DBUFHPSJFT TVDI BT -ĠĔĒĥĚĠğ 0ģĘĒğĚīĒĥĚĠğ 1ĖģĤĠğ PS 0ĥęĖģ /PUF
    UIBU UIJT UBTL JT DPOUFYU EFQFOEFOU BT .JMBO DBO CF B MPDBUJPO UIF DJUZ
    PS BO PSHBOJ[BUJPO B TQPSUT
    UFBN i.JMBO QMBZFE BHBJOTU #BSTB 8FEOFTEBZ FWFOJOHw
    BOE 1BSJT DBO CF UIF OBNF PG B DJUZ PS
    B QFSTPO
    " UZQJDBM JOQVU UP UIF QSPCMFN XPVME CF B TFOUFODF TVDI BT
    CQ?M aKBi? - T`2bB/2Mi Q7 J+*Q`KBF AM/mbi`B2b pBbBi2/ ?Bb MB2+2 S`Bb BM JBHM - `2TQ`i2`b
    bv X
    BOE UIF FYQFDUFE PVUQVU XPVME CF
    (S1_ CQ?M aKBi? ) - T`2bB/2Mi Q7 (P_: J+*Q`KBF AM/mbi`B2b ) pBbBi2/ ?Bb MB2+2 (S1_ S`Bb )
    BM (GP* JBHM )- `2TQ`i2`b bv X
    8IJMF /&3 JT B TFRVFODF TFHNFOUBUJPO UBTL‰JU BTTJHOT MBCFMFE CSBDLFUT PWFS OPO
    PWFSMBQQJOH TFOUFODF TQBOT‰JU JT PGUFO NPEFMFE BT B TFRVFODF UBHHJOH UBTL MJLF 104UBHHJOH
    ɩF VTF PG UBHHJOH UP TPMWF TFHNFOUBUJPO UBTLT JT QFSGPSNFE VTJOH #*0 FODPEFE UBHTņ &BDI XPSE
    JT BTTJHOFE POF PG UIF GPMMPXJOH UBHT BT TFFO JO 5BCMF
    5BCMF #*0 UBHT GPS OBNFE FOUJUZ SFDPHOJUJPO
    Tag Meaning
    O Not part of a named entity
    B-PER
    I-PER
    First word of a person name
    Continuation of a person name
    B-LOC First word of a location name
    ظ଴ग़ྗ
    #*0λά෇͚
    •IOB (or IOB1): CoNLL 2003 データセットで使われている
    • 接頭辞 I を付与; NEが連続する場合,2番目以降 NE 開始単語に
    I-PER E-PER S-PER O S-LOC O
    B-PER
    I-PER I-PER B-PER O I-LOC O
    B-PER
    I-PER I-PER B-PER O I-LOC O
    I-PER
    IOB
    BIO
    BIOES
    系列タグスキーマ
    https://speakerdeck.com/himkt/neural-named-entity-recognition
    #-0$

    View Slide

  12. จ຺ʹຒΊࠐ·Εͨ୯ޠͱݴޠֶతૉੑɿલஔࢺҙຯᐆດੑղফ
    !12
    લஔࢺҙຯᐆດੑղফ QSFQPTJUJPOTFOTFEJTBNCJHVBUJPO

    ‣ ໨త 163104&
    डӹऀ #&/&'*$*"3:
    ظؒ %63"5*0/
    ৔ॴ -0$"5*0/

    ‣ લஔࢺपΓͷ૭Λૉੑͱͯ͠༻͍Δͷ͸ཧ૝తͰ͸ͳ͍
    ‣ Ή͠ΖώϡʔϦεςΟοΫʢࠨଆͷ࠷ॳͷಈࢺˍӈଆͷ࠷ॳͷ໊ࢺʣ͕༗ޮ
    w લஔࢺͷࢧ഑ཁૉ HPWFSOPS
    ͱ໨తޠ PCKFDU
    Λ֫ಘͰ͖Δ
    803% */ $0/5&95 -*/(6*45*$ '&"563&4
    13&104*5*0/ 4&/4& %*4".#*(6"5*0/
    1SFQPTJUJPOT XPSET MJLF PO JO XJUI BOE GPS TFSWF GPS DPOOFDUJOH QSFEJDBUFT XJUI UIFJS BSHV
    NFOUT BOE OPVOT XJUI UIFJS QSFQPTJUJPOBM NPEJmFST 1SFQPTJPOT BSF WFSZ DPNNPO BOE BMTP WFSZ
    BNCJHVPVT $POTJEFS GPS FYBNQMF UIF XPSE GPS JO UIF GPMMPXJOH TFOUFODFT

    B 8F XFOU UIFSF GPS MVODI
    C )F QBJE GPS NF
    D 8F BUF GPS UXP IPVST
    E )F XPVME IBWF MFGU GPS IPNF CVU JU TUBSUFE SBJOJOH
    ɩF XPSE GPS QMBZT B EJĊFSFOU SPMF JO FBDI PG UIFN JO B
    JU JOEJDBUFT B 1ĦģġĠĤĖ JO C
    B #ĖğĖėĚ
    ĔĚĒģĪ JO D
    B %ĦģĒĥĚĠğ BOE JO E
    B -ĠĔĒĥĚĠğ
    *O PSEFS UP GVMMZ VOEFSTUBOE UIF NFBOJOH PG B TFOUFODF POF TIPVME BSHVBCMMZ LOPX UIF
    DPSSFDU TFOTFT PG UIF QSFQPTJUJPOT XJUIJO JU ɩF QSFQPTJUJPOTFOTF EJTBNCJHVBUJPO UBTL EFBMT XJUI
    BTTJHOJOH UIF DPSSFDU TFOTF UP B QSFQPTJUJPO JO DPOUFYU GSPN B mOJUF JOWFOUPSZ PG TFOTFT 4DIOFJEFS
    FU BM < > EJTDVTT UIF UBTL QSFTFOU B VOJmFE TFOTF JOWFOUPSZ UIBU DPWFST NBOZ QSFQPTJ
    803%*/$0/5&95 -*/(6*45*$'&"563&413&104*5*0/4&/4&%*4".#*(6"5*0/
    UJPOT BOE QSPWJEF B TNBMM BOOPUBUFE DPSQVT PG TFOUFODFT GSPN POMJOF SFWJFXT DPWFSJOH
    QSFQPTJUJPO NFOUJPOT FBDI BOOPUBUFE XJUI JUT TFOTFŇ
    8IJDI BSF B HPPE TFU PG GFBUVSFT GPS UIF QSFQPTJUJPO TFOTF EJTBNCJHVBUJPO UBTL 8F GPMMPX
    IFSF UIF GFBUVSF TFU JOTQJSFE CZ UIF XPSL PG )PWZ FU BM <>
    0CWJPVTMZ UIF QSFQPTJUJPO JUTFMG JT B VTFGVM GFBUVSF UIF EJTUSJCVUJPO PG QPTTJCMF TFOTFT GPS
    JO JT WFSZ EJĊFSFOU GSPN UIF EJTUSJCVUJPO PG TFOTFT GPS XJUI PS BCPVU GPS FYBNQMF
    #FTJEFT UIBU
    XF XJMM MPPL JO UIF DPOUFYU JO XIJDI UIF XPSE PDDVST " mYFE XJOEPX BSPVOE UIF QSFQPTJUJPO
    NBZ OPU CF JEFBM JO UFSNT PG JOGPSNBUJPO DPOUFOU UIPVHIU $POTJEFS GPS FYBNQMF UIF GPMMPXJOH
    TFOUFODFT

    B )F MJLFE UIF SPVOE PCKFDU GSPN UIF WFSZ mSTU UJNF IF TBX JU
    C )F TBWFE UIF SPVOE PCKFDU GSPN IJN UIF WFSZ mSTU UJNF UIFZ TBX JU
    ɩF UXP JOTUBODFT PG GSPN IBWF EJĊFSFOU TFOTFT CVU NPTU PG UIF XPSET JO B XJOEPX BSPVOE UIF
    XPSE BSF FJUIFS OPU JOGPSNBUJWF PS FWFO NJTMFBEJOH 8F OFFE B CFUUFS NFDIBOJTN GPS TFMFDUJOH
    JOGPSNBUJWF DPOUFYUT 0OF PQUJPO XPVME CF UP VTF B IFVSJTUJD TVDI BT iUIF mSTU WFSC PO UIF
    MFGUw BOE iUIF mSTU OPVO PO UIF SJHIUw ɩFTF XJMM DBQUVSF UIF USJQMFUT hMJLFE GSPN UJNFi BOE
    hTBWFE GSPN IJNi XIJDI JOEFFE DPOUBJO UIF FTTFODF PG UIF QSFQPTJUJPO TFOTF *O MJOHVJTUJD UFSNT
    XF TBZ UIBU UIJT IFVSJTUJD IFMQT VT DBQUVSF UIF HPWFSOPS BOE UIF BOE PCKFDU PG UIF QSFQPTJUJPO
    #Z LOPXJOH UIF JEFOUJGZ PG UIF QSFQPTJUJPO BT XFMM BT JUT HPWFSOPS BOE PCKFDUT IVNBOT DBO JO

    View Slide

  13. จ຺ʹຒΊࠐ·Εͨ୯ޠͱݴޠֶతૉੑɿલஔࢺҙຯᐆດੑղফ
    !13
    લஔࢺҙຯᐆດੑղফ QSFQPTJUJPOTFOTFEJTBNCJHVBUJPO

    ‣ ໨త 163104&
    डӹऀ #&/&'*$*"3:
    ظؒ %63"5*0/
    ৔ॴ -0$"5*0/

    ‣ લஔࢺपΓͷ૭Λૉੑͱͯ͠༻͍Δͷ͸ཧ૝తͰ͸ͳ͍
    ‣ Ή͠ΖώϡʔϦεςΟοΫʢࠨଆͷ࠷ॳͷಈࢺˍӈଆͷ࠷ॳͷ໊ࢺʣ͕༗ޮ
    w લஔࢺͷࢧ഑ཁૉ HPWFSOPS
    ͱ໨తޠ PCKFDU
    Λ֫ಘͰ͖Δ
    w ґଘߏ଄ύʔβʢޙड़ʣͷ݁Ռ΋༻͍ͯϩόετʹ͢Δ
    803% */ $0/5&95 -*/(6*45*$ '&"563&4
    13&104*5*0/ 4&/4& %*4".#*(6"5*0/
    1SFQPTJUJPOT XPSET MJLF PO JO XJUI BOE GPS TFSWF GPS DPOOFDUJOH QSFEJDBUFT XJUI UIFJS BSHV
    NFOUT BOE OPVOT XJUI UIFJS QSFQPTJUJPOBM NPEJmFST 1SFQPTJPOT BSF WFSZ DPNNPO BOE BMTP WFSZ
    BNCJHVPVT $POTJEFS GPS FYBNQMF UIF XPSE GPS JO UIF GPMMPXJOH TFOUFODFT

    B 8F XFOU UIFSF GPS MVODI
    C )F QBJE GPS NF
    D 8F BUF GPS UXP IPVST
    E )F XPVME IBWF MFGU GPS IPNF CVU JU TUBSUFE SBJOJOH
    ɩF XPSE GPS QMBZT B EJĊFSFOU SPMF JO FBDI PG UIFN JO B
    JU JOEJDBUFT B 1ĦģġĠĤĖ JO C
    B #ĖğĖėĚ
    ĔĚĒģĪ JO D
    B %ĦģĒĥĚĠğ BOE JO E
    B -ĠĔĒĥĚĠğ
    *O PSEFS UP GVMMZ VOEFSTUBOE UIF NFBOJOH PG B TFOUFODF POF TIPVME BSHVBCMMZ LOPX UIF
    DPSSFDU TFOTFT PG UIF QSFQPTJUJPOT XJUIJO JU ɩF QSFQPTJUJPOTFOTF EJTBNCJHVBUJPO UBTL EFBMT XJUI
    BTTJHOJOH UIF DPSSFDU TFOTF UP B QSFQPTJUJPO JO DPOUFYU GSPN B mOJUF JOWFOUPSZ PG TFOTFT 4DIOFJEFS
    FU BM < > EJTDVTT UIF UBTL QSFTFOU B VOJmFE TFOTF JOWFOUPSZ UIBU DPWFST NBOZ QSFQPTJ
    803%*/$0/5&95 -*/(6*45*$'&"563&413&104*5*0/4&/4&%*4".#*(6"5*0/
    UJPOT BOE QSPWJEF B TNBMM BOOPUBUFE DPSQVT PG TFOUFODFT GSPN POMJOF SFWJFXT DPWFSJOH
    QSFQPTJUJPO NFOUJPOT FBDI BOOPUBUFE XJUI JUT TFOTFŇ
    8IJDI BSF B HPPE TFU PG GFBUVSFT GPS UIF QSFQPTJUJPO TFOTF EJTBNCJHVBUJPO UBTL 8F GPMMPX
    IFSF UIF GFBUVSF TFU JOTQJSFE CZ UIF XPSL PG )PWZ FU BM <>
    0CWJPVTMZ UIF QSFQPTJUJPO JUTFMG JT B VTFGVM GFBUVSF UIF EJTUSJCVUJPO PG QPTTJCMF TFOTFT GPS
    JO JT WFSZ EJĊFSFOU GSPN UIF EJTUSJCVUJPO PG TFOTFT GPS XJUI PS BCPVU GPS FYBNQMF
    #FTJEFT UIBU
    XF XJMM MPPL JO UIF DPOUFYU JO XIJDI UIF XPSE PDDVST " mYFE XJOEPX BSPVOE UIF QSFQPTJUJPO
    NBZ OPU CF JEFBM JO UFSNT PG JOGPSNBUJPO DPOUFOU UIPVHIU $POTJEFS GPS FYBNQMF UIF GPMMPXJOH
    TFOUFODFT

    B )F MJLFE UIF SPVOE PCKFDU GSPN UIF WFSZ mSTU UJNF IF TBX JU
    C )F TBWFE UIF SPVOE PCKFDU GSPN IJN UIF WFSZ mSTU UJNF UIFZ TBX JU
    ɩF UXP JOTUBODFT PG GSPN IBWF EJĊFSFOU TFOTFT CVU NPTU PG UIF XPSET JO B XJOEPX BSPVOE UIF
    XPSE BSF FJUIFS OPU JOGPSNBUJWF PS FWFO NJTMFBEJOH 8F OFFE B CFUUFS NFDIBOJTN GPS TFMFDUJOH
    JOGPSNBUJWF DPOUFYUT 0OF PQUJPO XPVME CF UP VTF B IFVSJTUJD TVDI BT iUIF mSTU WFSC PO UIF
    MFGUw BOE iUIF mSTU OPVO PO UIF SJHIUw ɩFTF XJMM DBQUVSF UIF USJQMFUT hMJLFE GSPN UJNFi BOE
    hTBWFE GSPN IJNi XIJDI JOEFFE DPOUBJO UIF FTTFODF PG UIF QSFQPTJUJPO TFOTF *O MJOHVJTUJD UFSNT
    XF TBZ UIBU UIJT IFVSJTUJD IFMQT VT DBQUVSF UIF HPWFSOPS BOE UIF BOE PCKFDU PG UIF QSFQPTJUJPO
    #Z LOPXJOH UIF JEFOUJGZ PG UIF QSFQPTJUJPO BT XFMM BT JUT HPWFSOPS BOE PCKFDUT IVNBOT DBO JO
    NBOZ DBTFT JOGFS UIF TFOTF PG UIF QSFQPTJUJPO VTJOH SFBTPOJOH QSPDFTTFT BCPVU UIF mOFHSBJOFE
    QSFQPTJUJPO NFOUJPOT FBDI BOOPUBUFE XJUI JUT TFOTFŇ
    8IJDI BSF B HPPE TFU PG GFBUVSFT GPS UIF QSFQPTJUJPO TFOTF EJTBNCJHVBUJPO UBTL
    IFSF UIF GFBUVSF TFU JOTQJSFE CZ UIF XPSL PG )PWZ FU BM <>
    0CWJPVTMZ UIF QSFQPTJUJPO JUTFMG JT B VTFGVM GFBUVSF UIF EJTUSJCVUJPO PG QPTTJCMF
    JO JT WFSZ EJĊFSFOU GSPN UIF EJTUSJCVUJPO PG TFOTFT GPS XJUI PS BCPVU GPS FYBNQMF
    #F
    XF XJMM MPPL JO UIF DPOUFYU JO XIJDI UIF XPSE PDDVST " mYFE XJOEPX BSPVOE UIF Q
    NBZ OPU CF JEFBM JO UFSNT PG JOGPSNBUJPO DPOUFOU UIPVHIU $POTJEFS GPS FYBNQMF UIF
    TFOUFODFT

    B )F MJLFE UIF SPVOE PCKFDU GSPN UIF WFSZ mSTU UJNF IF TBX JU
    C )F TBWFE UIF SPVOE PCKFDU GSPN IJN UIF WFSZ mSTU UJNF UIFZ TBX JU
    ɩF UXP JOTUBODFT PG GSPN IBWF EJĊFSFOU TFOTFT CVU NPTU PG UIF XPSET JO B XJOEPX
    XPSE BSF FJUIFS OPU JOGPSNBUJWF PS FWFO NJTMFBEJOH 8F OFFE B CFUUFS NFDIBOJTN GP
    JOGPSNBUJWF DPOUFYUT 0OF PQUJPO XPVME CF UP VTF B IFVSJTUJD TVDI BT iUIF mSTU W
    MFGUw BOE iUIF mSTU OPVO PO UIF SJHIUw ɩFTF XJMM DBQUVSF UIF USJQMFUT hMJLFE GSPN
    hTBWFE GSPN IJNi XIJDI JOEFFE DPOUBJO UIF FTTFODF PG UIF QSFQPTJUJPO TFOTF *O MJOHV
    XF TBZ UIBU UIJT IFVSJTUJD IFMQT VT DBQUVSF UIF HPWFSOPS BOE UIF BOE PCKFDU PG UIF Q
    #Z LOPXJOH UIF JEFOUJGZ PG UIF QSFQPTJUJPO BT XFMM BT JUT HPWFSOPS BOE PCKFDUT IVN
    NBOZ DBTFT JOGFS UIF TFOTF PG UIF QSFQPTJUJPO VTJOH SFBTPOJOH QSPDFTTFT BCPVU UIF m
    TFNBOUJDT PG UIF XPSET ɩF IFVSJTUJD GPS FYUSBDUJOH UIF PCKFDU BOE HPWFSOPS SFRVJSFT
    B 104UBHHFS JO PSEFS UP JEFOUJGZ UIF OPVOT BOE WFSCT *U JT BMTP TPNFXIBU CSJUUMF
    IBSE UP JNBHJOF DBTFT JO XIJDI JU GBJMT 8F DPVME SFmOF UIF IFVSJTUJD XJUI NPSF SVMFT
    SPCVTU BQQSPBDI XPVME CF UP VTF B EFQFOEFODZ QBSTFS UIF HPWFSOPS BOE PCKFDU JOGP
    FBTJMZ SFBEBCMF GSPN UIF TZOUBDUJD USFF SFEVDJOH UIF OFFE GPS DPNQMFY IFVSJTUJDT
    he liked the round object from the very first time he saw it
    nsubj nsubj dobj
    det
    amod amod amod
    det rcmod
    dobj
    prep pobj
    root
    0G DPVSTF UIF QBSTFS VTFE GPS QSPEVDJOH UIF USFF NBZ CF XSPOH UPP 'PS SPCVTUOFTT X

    View Slide

  14. จ຺ʹຒΊࠐ·Εͨ୯ޠͷؒͷؔ܎ɿΞʔΫΛ୯Ґͱͨ͠ύʔδϯά
    !14
    ґଘߏ଄ύʔδϯά EFQFOEFODZQBSTJOH

    ‣ ౷ޠతґଘߏ଄໦ TZOUBDUJDEFQFOEFODZUSFF
    Λฦ͢
    ‣ Oݸͷ୯ޠͱ୯ޠͷؔ܎ʢΞʔΫʣʹείΞ"3$4$03& I N TFOU
    ΛׂΓ౰ͯΔ
    w จTFOUɺओࣙ୯ޠީิͷΠϯσοΫεIɺम০ޠ୯ޠީิͷΠϯσοΫεN
    3&-"5*0/ #&58&&/ 803%4 */ $0/5&95 "3$'"$503&% 1"34*/(
    ɩF QSFQPTJUJPOTFOTF EJTBNCJHVBUJPO UBTL JT BO FYBNQMF PG B IJHIMFWFM TFNBOUJD DMBTTJm
    DBUJPO QSPCMFN GPS XIJDI XF OFFE B TFU PG GFBUVSFT UIBU DBOOPU CF SFBEJMZ JOGFSSFE GSPN UIF TVSGBDF
    GPSNT BOE DBO CFOFmU GSPN MJOHVJTUJD QSFQSPDFTTJOH JF 104UBHHJOH BOE TZOUBDUJD QBSTJOH

    BT XFMM BT GSPN TFMFDUFE QJFDFT PG JOGPSNBUJPO GSPN NBOVBMMZ DVSBUFE TFNBOUJD MFYJDPOT
    3&-"5*0/ #&58&&/ 803%4 */ $0/5&95
    "3$'"$503&% 1"34*/(
    *O UIF EFQFOEFODZ QBSTJOH UBTL XF BSF HJWFO B TFOUFODF BOE OFFE UP SFUVSO B TZOUBDUJD EFQFOEFODZ
    USFF PWFS JU TVDI BT UIF USFF JO 'JHVSF &BDI XPSE JT BTTJHOFE B QBSFOU XPSE FYDFQU GPS UIF
    NBJO XPSE PG UIF TFOUFODF XIPTF QBSFOU JT B TQFDJBM _PPh TZNCPM
    the boy with the black shirt opened the door with a key
    det prep
    prep
    amod
    det
    pobj
    nsubj
    root
    dobj pobj
    det det
    'JHVSF %FQFOEFODZ USFF
    'PS NPSF JOGPSNBUJPO PO UIF EFQFOEFODZ QBSTJOH UBTL JUT MJOHVJTUJD GPVOEBUJPOT BOE BQ
    QSPBDIFT UP JUT TPMVUJPO TFF UIF CPPL CZ ,àCMFS FU BM <>
    0OF BQQSPBDI UP NPEFMJOH UIF UBTL JT UIF BSDGBDUPSFE BQQSPBDI <.D%POBME FU BM >
    XIFSF FBDI PG UIF QPTTJCMF n2 XPSEXPSE SFMBUJPOT BSDT
    JT BTTJHOFE B TDPSF JOEFQFOEFOU PG UIF
    PUIFST BOE UIFO XF TFBSDI GPS UIF WBMJE USFF XJUI UIF NBYJNBM PWFSBMM TDPSF ɩF TDPSF BTTJHONFOU
    JT NBEF CZ B USBJOFE TDPSJOH GVODUJPO "ģĔ4ĔĠģĖ.h; m; sent/ SFDFJWJOH B TFOUFODF BT XFMM BT UIF
    JOEJDFT h BOE m PG UXP XPSET XJUIJO JU UIBU BSF DPOTJEFSFE BT DBOEJEBUFT GPS BUUBDINFOU h JT
    UIF JOEFY PG UIF DBOEJEBUF IFBEXPSE BOE m JT UIF JOEFY PG UIF DBOEJEBUF NPEJmFS
    5SBJOJOH UIF
    TDPSJOH GVODUJPO TVDI UIBU JU XPSLT XFMM XJUI UIF TFBSDI QSPDFEVSF XJMM CF EJTDVTTFE JO $IBQUFS

    View Slide

  15. จ຺ʹຒΊࠐ·Εͨ୯ޠͷؒͷؔ܎ɿΞʔΫΛ୯Ґͱͨ͠ύʔδϯά
    !15
    Α͘࢖ΘΕΔૉੑ
    ‣ ओࣙ୯ޠͷޠܗͱ඼ࢺλά
    w DBLFBUF
    ‣ म০ޠͷޠܗͱ඼ࢺλάʢUIF Bʣ
    ‣ ओࣙ୯ޠͷ྆ଆ̎ޠͷ૭
    ‣ म০ޠͷ྆ଆ̎ޠͷ૭
    ‣ ୯ޠΫϥελɾ୯ޠຒΊࠐΈͳͲͷ෼෍࿦త৘ใ
    w Ұൠʹґଘߏ଄໦ͷ܇࿅ίʔύε͸͋·Γେ͖͘ͳ͍
    ‣ ୯ޠͱ୯ޠͷڑ཭EJTUcINc
    ‣ ୯ޠؒͷํ޲
    w XNUIF XICPZͷͱ͖ɺNIͩͱΞʔΫ͕షΒΕΔՄೳੑߴ͍͕ɺNIͩͱ΄΅ͳ͍
    ‣ ୯ޠؒʹݱΕΔ୯ޠͷޠܗ
    w UIFʜBʜCPZͷΑ͏ʹ UIF CPZ
    ͷؒʹݶఆࢺ͕͋Δ৔߹ɺUIF͕CPZʹ͔͔ΔՄೳੑ͸௿͍

    View Slide

  16. ໨࣍
    !16
    ୈ̎ฤɹࣗવݴޠσʔλͷѻ͍
    ‣ ୈ̒ষɹςΩετσʔλͷͨΊͷૉੑ
    ‣ ୈ̓ষɹࣄྫݚڀɿࣗવݴޠॲཧʹ͓͚Δૉੑ
    w จॻ෼ྨɿݴޠಉఆ
    w จॻ෼ྨɿτϐοΫ෼ྨ
    w จॻ෼ྨɿஶऀಛఆ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠɿ඼ࢺλά෇͚
    w จ຺ʹຒΊࠐ·Εͨ୯ޠɿݻ༗දݱೝࣝ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠͱݴޠֶతૉੑɿલஔࢺҙຯᐆດੑղফ
    w จ຺ʹຒΊࠐ·Εͨ୯ޠͷؒͷؔ܎ɿΞʔΫΛ୯Ґͱͨ͠ύʔδϯά

    View Slide