Entity Set Search of Scientific Literature: An unsupervised Ranking Approach

論文の15分紹介

Transcript

1. Wataru Hirota (M1) w-hirota@ist.osaka-u.ac.jp Entity Set Search of
 Scientific Literature:
 An Unsupervised Ranking Approach 4*(*3
    Shen, J., Xiao, J., He, X., Shang, J., Sinha, S., & Han, J. (University of Illinois)
2. ֓ཁ ! Entity Set Search Ͱߴ͍ਫ਼౓ΛތΔݕࡧख๏
 SetRank ΛఏҊ ! ϥϕϧͳ͠Ͱ࠷దͳύϥϝʔλΛબఆ 2
3. Entity ͱ͸? ! ≒ʮ࣮ମʯ ! ≒ Wikipedia ͷλΠτϧ - ʮླ໦Ұ࿠ʯʮσʔλϕʔεʯ - ֊૚తͳ type Λ࣋ͭ 3 5IJOH 1FSTPO 3FBM1FSTPO "UIMFUF *DIJSP
   4V[VLJ &OUJUZ
   ֊૚
4. Entity Set Search ͱ͸? ! ΫΤϦதʹෳ਺ͷ entity ؚ͕·ΕΔΑ͏ͳݕࡧ ! ͢΂ͯͷ entity ؚ͕·Ε͍ͯΔΑ͏ͳ
 ݕࡧ݁Ռ͕΄͍͠ ! e.g. knowledge base for document retrieval - knowledge base ͚ͩͰ΋ document retrieval ͚ͩͰ΋
 ෆे෼ 4
5. ख๏ͷ֓ཁ 5 1. Entity Extraction 2. Generation of Query Graph ˠݕࡧΛάϥϑͷඃ෴໰୊ͱͯ͠ղ͘
6. 1. Entity Extraction 6 ! ΫΤϦ͔Β entity Λநग़ ! ಉ࣌ʹநग़ͨ͠ entity ಉ࢜ͷ
 LCA (࠷খڞ௨૆ઌ)΋
 ܭࢉ
7. 2. Generation of Query Graph ! ΫΤϦΛ2ͭͷάϥϑͰදݱ͢Δ - word graph, entity graph 7 XPSE
   HSBQI
   ลͷॏΈ
   
   ྡ઀͢Δ୯ޠ͚ͩ XPSE
   HSBQI
   ลͷॏΈ
   -$"
   ྡ઀͢Δ୯ޠ͚ͩ
8. ݕࡧείΞ ! ʹ͓͚ΔϊʔυͷϙςϯγϟϧͱΤοδͷ
 ϙςϯγϟϧͷ࿨ - … ΫΤϦάϥϑ q ͷ͏ͪจষ di ͕ඃ෴͢Δ෦෼ ! ϊʔυͷϙςϯγϟϧ: ! Τοδͷϙςϯγϟϧ : - t ͸ தͷ word ·ͨ͸ entity - a(.) ͸׆ੑԽؔ਺ 8 Σt a(P(t|di )) Gq|di Gq|di Gq|di Σ(t1 ,t2 ) a(P(t1 |di )) a(P(t2 |di ))
9. ݕࡧείΞ 9 XPSE
   HSBQI
   ͷඃ෴౓ FOUJUZ
   HSBQI
   ͷඃ෴౓ a( ⋅ ) &OUJUZ
   HSBQI
   ͱ
   8PSE
   HSBQI
   ͷॏΈ
10. ύϥϝʔλબ୒ ! ຊ࿦จͷ2൪໨ͷ contribution ! Unsupervised - Պֶจݙݕࡧͷϥϕϧ෇͚͸ඇৗʹίετ͕͔͔Δ 10
11. ύϥϝʔλબఆ ! ؊ͱͳΔΞΠσΞ͸
 ͋Δݕࡧ݁Ռ͕, ͢΂ͯͷύϥϝʔλͷ
 ݕࡧ݁Ռͱ߹ࢉͨ͠΋ͷͱࣅ͍ͯΔ΄Ͳ
 ͦΕ͸ྑ͍ݕࡧ݁ՌͰ͋Δ 11 ύϥϝʔλ ύϥϝʔλ ύϥϝʔλO ݕࡧ݁Ռ
    ߹ࢉͨ݁͠Ռ ύϥϝʔλN ࣅ͍ͯΔ΄Ͳ
    HPPE ஫
    ύϥϝʔλू߹͸༗ݶ
 ͔ͭखಈͰ༩͑Δ
12. ࣮ݧ (ධՁࢦඪ) ! ࢦඪʹ͸ nDCG[1] Λ࢖༻ - ߴ͍ॱҐͷจষͷείΞ (≒ద߹౓߹͍) ͕ߴ͍΄Ͳ
 nDCG ͸ߴ͘ͳΔ - ৘ใݕࡧʹ͓͚ΔϝδϟʔͳධՁࢦඪ 12 nDCG = ΣK i=2 scorei log i [1] Järvelin, Kalervo, et al. "Cumulated gain-based evaluation of IR techniques. TOIS. 2012.
13. ࣮ݧઃఆ ! Ωʔϫʔυݕࡧͷఆ൪3ͭͱ nDCG Λൺֱ - BM25[1] - Likelihood Model[2] ▪ Dirichlet prior smoothing (LM-DIR) ▪ Jelinek Mercer smoothing (LM-JM) - Information-Based Model[3] (IB) 13 [1] Stephen E, et al. The Probabilistic Relevance Framework: BM25 and Beyond. Foundations and Trends in Information
 Retrieval. 2009. [2] ChengXiang Zhai and John D. Lafferty. A Study of Smoothing Methods for Language Models Applied to Ad Hoc
 Information Retrieval. SIGIR. 2001. [3] Stéphane Clinchant and Éric Gaussier. Information-based models for ad hoc IR. SIGIR. 2010.
14. ࣮ݧ݁Ռ (ϕʔεϥΠϯͱͷൺֱ) ! 4ͭͷϕʔεϥΠϯʹൺ΂ͯ nDCG ͕༗ҙʹ޲্ - tݕఆͰ p-value < 0.05 14
15. ࣮ݧ݁Ռ (άϥϑߏ଄ͷ༗༻ੑ) ! -t … entity ͷΤοδͷॏΈΛ LCA Ͱ͸ͳ͘ 1 ʹݻఆ ! -s … ΤοδΛ࡟আ ! -t, -ts ૒ํΑΓ΋ߴ͍ nDCG
 → ఏҊͨ͠άϥϑߏ଄͸༗ޮ 15
16. ࣮ݧ݁Ռ (ύϥϝʔλબఆ) ! ਖ਼ղσʔλͰ validation ͨ͠ύϥϝʔλͱ
 ύϥϝʔλͷ஋, nDCG ͕ͲͪΒ΋͍ۙ
 → ͏·͘ύϥϝʔλΛબఆͰ͖͍ͯΔ 16 ˞
    ,5
    QPT,5
    ͸ͦΕͧΕڑ཭ई౓
17. ·ͱΊ ! ΫΤϦΛ entity ͱ୯ޠͷάϥϑߏ଄ʹม׵ͯ͠ݕࡧ ! ύϥϝʔλ΋ࣗಈͰબఆ ! طଘͷΩʔϫʔυݕࡧΑΓ nDCG ͕ߴ͍ 17