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

論文紹介: Word-node2vec

論文紹介: Word-node2vec

- 首都大 小町研 NAACL 2019 読み会
- Word-Node2Vec: Improving Word Embedding with Document-Level Non-Local Word Co-occurrences. Sen et al., NAACL2019 を紹介
- paper URL: https://www.aclweb.org/anthology/N19-1109.pdf

Avatar for Satoru Katsumata

Satoru Katsumata

December 11, 2023
Tweet

More Decks by Satoru Katsumata

Other Decks in Research

Transcript

  1. ֓ཁ w XPSEWFD (MPWFͳͲ͸͋Δ୯ޠʹରͯ͠ɺ
 ૭෯಺ͷ୯ޠΛڞىͨ͠୯ޠͱͯ͠Δ w จॻ಺ͷ૭෯֎ͷ୯ޠʹରͯ͠΋ڞىΛݟͨϞσϧΛఏҊ
 ˠXPSEOPEFWFD
 MPDBMͱOPOMPDBMͷڞىɺͲͪΒ΋ݟΔ
 ۩ମతʹ͸ɺFEHFͷॏΈ࡞੒࣌ʹ


    MPDBMɺOPOMPDBMΛߟྀ͍ͯ͠Δ w ධՁλεΫXPSETJNJMBSJUZXPSEBOBMPHZ
 DPODFQUDBUFHPSJ[BUJPO
 bXPSEOPEFWFDPVUQFSGPSNTXPSEWFDBOEHMPWFPOBSBOHF PGEJ⒎FSFOUUBTLT` !2
  2. άϥϑͷઃఆ
 ʢஶऀͷ࣮૷Ͱ֬ೝʣ w &NCFEEJOHͷֶशσʔλ
 ˠෳ਺ͷEPDVNFOUTʢEPDVNFOU୯Ґͷ۠੾Γ͕ଘࡏʣ w OPEF୯ޠɺFEHF୯ޠͷڞىɺແ޲άϥϑ w ڞىͷఆٛυΩϡϝϯτ಺Ͱग़ͯདྷͨ΋ͷશͯ
 ˠEPDVNFOU಺ͷग़ݱ୯ޠશ͕ͯ͋Δ୯ޠͷۙ๣OPEFͱͳΔ

    w 4UPQXPSEͰbUIF`ͱ͔͸֎ͯ͠Δ !3 w1 i w1 1 w1 N ʜ EPDMFOHUI/ ؆୯ͷͨΊJUIXPSEʹؔ͢ΔFEHFͷΈهࡌ w2 j w2 1 w2 O ʜ EPDMFOHUI0 if(w1 i = w2 j ) w1 i w1 1 w1 N ʜ w2 1 w2 O ʜ DPNNJUEFGDFFCFGGEEB
  3. 4LJQHSBN
 XJUI/FHBUJWF4BNQMJOH w ఏҊख๏͸4LJQHSBNXJUI/FHBUJWF4BNQMJOH 4(/4 Λجʹͨ͠΋ͷ
 DPOUFYUXPSEͷܾΊํʹάϥϑཁૉ͕ࠞೖ͚ͨͩ͠ w ͱΓ͋͑ͣ4(/4ͷ͓͞Β͍
 ʢXʹର͢Δ4(/4ʣ

    !4 Λ(w) = {i : ti = w} c(w) = ∪i∈Λ(w) ∪k j = − k j ≠ 0 ti+j Ω+ = ∪w∈V {w, c(w)} Ω− = ∪w∈V {w, ∪ {v : v ∼ (V − c(w))}} /PUBUJPO ֶशσʔλશମ
 ʢ5͸૯UPLFO਺ʣ ˡXʹର͢ΔJOEFYू߹ ˡXʹର͢Δपลޠू߹ ਖ਼ྫ ෛྫ 8JOEPXTJ[FL 7PDBCTFU7 ໨తؔ਺     
  4. 4(/4(SBQIͷ৔߹ !5 ֶश࣌ɺೖྗ͸άϥϑ(ʢ1ͷҰ൪ӈͷ΍ͭʣ OPEF FEHF c(w) = ∪i∈Λ(w) ∪k j

    = − k j ≠ 0 ti+j ˢ Дۙ๣ /PUBUJPO IPQ਺ؔ਺ 8JOEPXTJ[FL ʢDPOUFYUXPSET͕࠷େLݸ·Ͱʣ h(u, v) ਖ਼ྫෛྫ ໨తؔ਺       
  5. ࣮ݧઃఆ w %BUB%#1FEJB
 ස౓ҎԼͷ୯ޠ͸ແࢹTUPQXPSETऔΓআ͖ w ख๏
 MPDBM4(/4(MPWF'BTUUFYU&MNP
 %PDVNFOUMFWFMDPPDDVS-%"
 DPNCJOBUJPOMPDBMBOEOPOMPDBM4(/4ͱ-%"ͷઢܗิؒ
 OPEFWFD΋CBTFMJOF

    w ධՁͱλεΫ
 XPSETJNJMBSJUZ.&/84*.
 XPSEBOBMPHZ(PPHMF.434FN&WBM
 DPODFQUDBUFHPSJ[BUJPO"1#-&44&44-@C w .FUSJD
 TJNJMBSJUZ4QFBSNBO`TЛ
 ͦΕҎ֎BDDVSBDZ w ύϥϝʔλTJNJMBSJUZͷ.&/͕࠷େԽ͢ΔΑ͏ʹ
 ଞͷλεΫ͸ͦͷύϥϝʔλΛ࢖͍·Θ͠ !9 DPODFQUDBUPHPSJ[BUJPO  OPVOʹରͯ͠DPODFQUΛ౰ͯΔ bDFSFNPOZ` bHSBEVBUJPO`
 ˠb4PDJBM0DDBTJPO`
 ʢ෼ྨʹ࢖༻ͨ͠Ϟσϧ͸ෆ໌ʣ