20180414_WSDM2018_reading_YoheiKIKUTA

HP: https://atnd.org/events/95510

Transcript

1. Why People Search for Images
   using Web Search Engines
   WSDM 2018 ࿦จಡΈձ
   20180414
   ٠ా ངฏ (@yohei_kikuta)
   Event URL: https://atnd.org/events/95510, paper: https://arxiv.org/abs/1711.09559
   

   View Slide

2. ·ͱΊ
   1. text base ͷ΢Σϒը૾ݕࡧͷҙਤ͸෼ྨՄೳ͔ʁ
   → YES. 3ͭʹ෼ྨ: Entertain, Explore/Learn, Locate/Acquire
   2. औಘՄೳͳಛ௃ྔ͔ΒҙਤΛ൑ผͰ͖Δ͔ʁ
   → YES. ଺ཹ࣌ؒ΍Ϛ΢εδΣενϟ
   3. ηογϣϯॳظͰݕࡧҙਤΛ༧ଌͰ͖Δ͔
   → MAYBE. ಛ௃ྔΛ࢖ͬͯϞσϧΛ࡞੒ͯ͠Ұఆ౓ͷੑೳ
   2
   

   View Slide

3. എܠ
   3
   

   View Slide

4. ݕࡧͷҙਤΛ஌Γ͍ͨ
   Ϣʔβͷݕࡧߦಈͷཪʹ͋ΔҙਤΛ஌Δ͜ͱ͸ॏཁ
   → Ϣʔβͷຬ଍౓޲্ʢsuggestion, recommendation, ...ʣ
   ΢Σϒݕࡧͷݚڀ͸ͳ͞Ε͖͕ͯͨɺը૾ݕࡧʹؔͯ͠͸ݶఆత
   → ΫΤϦϕʔε
   → ͔͠͠ը૾ݕࡧͷΫΤϦ͸୹͘ͳΓ͕ͪͰෆ࣮֬ੑ͕େ͖͍
   ຊ࿦จͰ͸ηογϣϯ৘ใΛѻͬͯը૾ݕࡧͷҙਤΛݚڀ
   4
   

   View Slide

5. ຊ࿦จʹ͓͚ΔϦαʔνΫΤενϣϯ
   1. text base ͷ΢Σϒը૾ݕࡧͷҙਤ͸෼ྨՄೳ͔ʁ
   2. औಘՄೳͳಛ௃ྔ͔ΒҙਤΛ൑ผͰ͖Δ͔ʁ
   3. ηογϣϯॳظͰݕࡧҙਤΛ༧ଌͰ͖Δ͔
   5
   

   View Slide

6. ઌߦݚڀ
   6
   

   View Slide

7. ΢Σϒݕࡧʹ͓͚Δҙਤͷ taxonomy
   A taxonomy of web search (2002) ͰҙਤΛ3ͭʹ෼ྨ
   1. Navigational: ಛఆͷαΠτ΁౸ୡ
   2. Informational: ৘ใͷऔಘ
   3. Transactional: ΢ΣϒΛഔհͱͨ͠׆ಈ
   Ref: https://dl.acm.org/citation.cfm?id=792552 7
   

   View Slide

8. ΢Σϒݕࡧʹ͓͚Δҙਤͷ taxonomy
   Task Behaviors During Web Search: The Difficulty of Assigning
   Labels (2009) Ͱ͸ݕࡧλεΫΛ7ͭʹ෼ྨ
   » Navigate, Find-Simple, Find-Complex, Locate/Acquire,
   Explore/Learn, Play, Meta
   ຊ࿦จ͸͜ͷઌߦݚڀΛ౿ऻͭͭ͠ը૾ݕࡧʹൃలͤͨ͞΋ͷɺͱ͍͏
   ৭߹͍͕ڧ͍
   Ref: http://ieeexplore.ieee.org/document/4755491/ 8
   

   View Slide

9. ը૾ݕࡧͷҙਤΛ෼ྨ
   9
   

   View Slide

10. Ξϓϩʔν
    σʔλΛूΊͯͦΕΛجʹ3ਓͷ΢Σϒݚڀऀ͕෼ྨ
    » ϢʔβͷΞϯέʔτσʔλ
    » ੑผ৘ใͳͲΛऔಘ
    » ࠷ۙͷݕࡧʹؔ͢ΔৄࡉʢಈػͳͲʣɺ࢖༻ͨ͠ΫΤϦ
    » ద੾ͳճ౴Λͨ͠211ਓ͕ର৅
    10
    

    View Slide

11. Ξϓϩʔν
    σʔλΛूΊͯͦΕΛجʹ3ਓͷ΢Σϒݚڀऀ͕෼ྨ
    » ϩάσʔλ
    » https://www.sogou.com/ ͷϩάσʔλ
    » 30෼Ҏ಺ʹ࿈ଓతͳΫΤϦΛ༩͍͑ͯΔ475ηογϣϯʢআ͘Ξμ
    ϧτʣ
    11
    

    View Slide

12. Ξϓϩʔν
    σʔλΛूΊͯͦΕΛجʹ3ਓͷ΢Σϒݚڀऀ͕෼ྨ
    » ϩάσʔλʢlength ͸ΫΤϦ਺ʣ
    Ref: https://arxiv.org/abs/1711.09559 12
    

    View Slide

13. ࡞੒ͨ͠൑அج४
    1. Ϣʔβͷݕࡧߦಈ͸໌֬ͳ໨తʹґΔ΋ͷ͔ʁ
    2. ޙͷར༻ͷͨΊʹը૾Λμ΢ϯϩʔυ͢Δඞཁ͕͋Δ͔ʁ
    13
    

    View Slide

14. 3ͭͷݕࡧҙਤ
    1. Explore/Learn (1-yes, 2-no)
    ྫʣΰϦϥͱϘϊϘͷݟͨ໨ͷҧ͍ΛνΣοΫ
    2. Locate/Acquire (1-yes, 2-yes)
    ྫʣϨϙʔτ࡞੒Ͱ࢖͏ΰϦϥͷը૾Λ୳ͯ͠μ΢ϯϩʔυ
    3. Entertain (1-no, 2-yes or no)
    ྫʣΰϦϥͷ໘നը૾ΛோΊΔ
    14
    

    View Slide

15. 3ͭͷݕࡧҙਤʢྫʣ
    Ref: https://arxiv.org/abs/1711.09559 15
    

    View Slide

16. ଥ౰ੑͷݕূʢ3ਓͷେֶӃੜʹΑΔҙਤ෼ྨʣ
    » ϢʔβͷΞϯέʔτσʔλ
    Fleiss' kappa: 0.673
    Explore/Learn: 27%, Locate/Acquire: 66%, Entertain: 7%
    » ϩάσʔλʢΫΤϦͷΈΛ࢖༻ʣ
    Fleiss' kappa: 0.375
    Explore/Learn: 56%, Locate/Acquire: 39%, Entertain: 5%
    ͏·͘෼͚Εͦ͏͕ͩΫΤϦͷΈͰ͸ҙਤΛ἞Ήͷ͸೉͍͠
    16
    

    View Slide

17. औಘՄೳͳಛ௃ྔͰҙਤΛ൑ผ
    17
    

    View Slide

18. 35ਓͷֶ෦ੜʹΑΔ12ݸͷը૾ݕࡧλεΫ
    ྫʣPCͷഎܠΛ੨ۭͱ৿ͷը૾ʹมߋʢLocate/Acquireʣ
    ͦͷࡍʹҎԼͷಛ௃ྔΛऔಘ
    Ref: https://arxiv.org/abs/1711.09559 18
    

    View Slide

19. ҙਤʹΑͬͯ༗ҙͳ͕ࠩग़ΔͷͰ൑ผՄೳ
    ఀཹ࣌ؒ͸ E/L ͕ଟ͍ɺϚ΢εΫϦοΫ਺͸ E/L < L/A < EɺͳͲ
    ʢৄࡉ͸࿦จΛࢀরʣ
    Ref: https://arxiv.org/abs/1711.09559 19
    

    View Slide

20. ηογϣϯॳظͰͷҙਤͷ༧ଌ
    20
    

    View Slide

21. ໰୊ઃఆ
    ηογϣϯॳظͱ͸ʮ࠷ॳͷϚ΢εεΫϩʔϧ͕͋Δ·Ͱʯ
    ༧ଌͰ࢖͏ feature ͱͯ͠ҎԼͷ஫ҙ
    - ΫϦοΫ਺ͱ࠷ॳͷϚ΢εΦʔόʔ࣌ؒ͸࢖Θͳ͍
    - ΫΤϦϕʔεͰ΍Γ͍ͨͷͰ query reformulation ͸࢖Θͳ͍
    ֶ෦ੜʹղ͔ͤͨը૾ݕࡧλεΫʹରͯ͠ GBDT Ͱ 10-fold CV
    21
    

    View Slide

22. ༧ଌੑೳ͸ߴ͘͸ͳ͍͕ෆՄೳͰ͸ͳͦ͞͏
    Baseline ͸ majority ʹશ෦دͤΔͱ͍͏΋ͷ
    Ref: https://arxiv.org/abs/1711.09559 22
    

    View Slide

23. ·ͱΊͱॴײ
    23
    

    View Slide

24. ·ͱΊʢ࠶ܝʣ
    1. text base ͷ΢Σϒը૾ݕࡧͷҙਤ͸෼ྨՄೳ͔ʁ
    → YES. 3ͭʹ෼ྨ: Entertain, Explore/Learn, Locate/Acquire
    2. औಘՄೳͳಛ௃ྔ͔ΒҙਤΛ൑ผͰ͖Δ͔ʁ
    → YES. ଺ཹ࣌ؒ΍Ϛ΢εδΣενϟ
    3. ηογϣϯॳظͰݕࡧҙਤΛ༧ଌͰ͖Δ͔
    → MAYBE. ಛ௃ྔΛ࢖ͬͯϞσϧΛ࡞੒ͯ͠Ұఆ౓ͷੑೳ
    24
    

    View Slide

25. ॴײ
    » γϯϓϧͳج४Ͱ෼ྨΛ͍ͯ͠Δͱ͍͏ͷ͸ྑ͍
    » ࢐৽ͱ͍͏Θ͚Ͱ͸ͳ͍͕ҰͭҰͭ͸͔ͬ͠Γௐ΂͍ͯΔ
    » ࣮αʔϏε΁ͷԠ༻ʹ͸Ұาඈ༂͕ඞཁͦ͏ʢ༧ଌੑೳͳͲʣ
    » ٱʑʹ਺ࣜΛશવ࢖Θͳ͍࿦จΛಡΜͰ৽઱ͩͬͨ
    25
    

    View Slide