Tweet
Tweet

Slide 1

Slide 1 text

No content

Slide 2

Slide 2 text

ுഓೇ 3FTFBSDI4DJFOUJTU $ZCFS"HFOU"*-BC w ೥ʹ"*-BCʹத్ೖࣾ w ۃϓϩμΫτͱ࿈ܞ͠ͳ͕Βɺ޿ࠂςΩετͷࣗಈੜ੒΍ ޿ࠂޮՌͷ༧ଌͳͲɺࣗવݴޠॲཧٕज़ͷ޿ࠂ෼໺ద༻ʹ ͍ͭͯͷݚڀ։ൃʹैࣄ

Slide 3

Slide 3 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ

Slide 4

Slide 4 text

എܠ  ੍࡞෺ͷधཁ֦େ  ੍࡞Ϧιʔεͷރׇ  ਓ޻஌ೳٕज़ͷ୆಄

Slide 5

Slide 5 text

എܠ੍࡞෺ͷधཁ֦େ ࢢ৔ن໛ͷ֦େ Πϯλʔωοτ޿ࠂࢢ৔͸͜͜೥Ͱ໿ഒ΋ͷن໛ʹ੒௕ ग़యɿ೥ͷΠϯλʔωοτ޿ࠂഔମඅ͸ஹԯԁʹɻϞόΠϧʴಈը޿ࠂͷ৳ͼʹ஫໨

Slide 6

Slide 6 text

ݕࡧ࿈ಈܕ޿ࠂʴσΟεϓϨΠ޿ࠂͰ૯޿ࠂඅͷ͏ͪ ໿ׂͷγΣΞΛތΔ ग़యɿʮ೥Πϯλʔωοτ޿ࠂഔମඅʯղઆɻϚεഔମͱ΄΅ฒΜͩʮஹԁ௒ʯͷ ಺༁͸ʁ എܠ੍࡞෺ͷधཁ֦େ

Slide 7

Slide 7 text

σΟεϓϨΠ޿ࠂͱ͸ w 8FCϖʔδͳͲͷ޿ࠂ࿮ʹදࣔ͞ΕΔ޿ࠂ w ߦಈཤྺͳͲ͔Βझຯᅂ޷ʹ߹͏Α͏ͳλʔήςΟ ϯά͕͞ΕΔ͜ͱ΋͋Δ w ը૾ɺςΩετɺಈըͳͲ͞·͟·ͳഔମ͔Βߏ੒ ͞ΕΔ

Slide 8

Slide 8 text

ݕࡧ࿈ಈܕ޿ࠂͱ͸ w ݕࡧΤϯδϯͰ࢖༻͞ΕΔ޿ࠂ w ϢʔβͷೖྗΩʔϫʔυͱ޿ࠂओͷઃఆΩʔϫʔυ͕Ϛονͨ͠৔߹ʹදࣔ͞ΕΔ w جຊతʹςΩετͷΈͰߏ੒͞ΕΔ

Slide 9

Slide 9 text

എܠ੍࡞Ϧιʔεͷރׇ ݕࡧΫΤϦ਺͸૿ՃͷҰ్ΛͨͲΔ w ຖ೥લ೥ͷ໿લޙͰ૿͑ଓ͚Δͱ༧૝͞ΕΔ w ೥࣌఺Ͱ໿ஹҎ্΋ͷݕࡧΩʔϫʔυʹ౸ୡ͢Δͱ ͍ΘΕΔग़యɿ(PPHMF4FBSDI4UBUJTUJDTBOE'BDUT :PV.VTU,OPX  ͜ΕΒશͯͷΩʔϫʔυʹ ରͯ͠ɺਓखʹΑͬͯߴ඼࣭ͳ ޿ࠂΛ੍࡞͢Δͷ͸೉͍͠

Slide 10

Slide 10 text

എܠਓ޻஌ೳٕज़ͷ୆಄ ۙ೥ɺଟ͘ͷ෼໺ʹͯϒϨΠΫεϧʔΛى͜͢ IUUQTXXXXJSFEDPNTUPSZBJUFYUHFOFSBUPSHQUMFBSOJOHMBOHVBHFpUGVMMZ IUUQTEFFQNJOEDPNSFTFBSDIDBTFTUVEJFTBMQIBHPUIFTUPSZTPGBS ೥)JOUPOΒͷݚڀνʔϜ͕0$3ͷίϯϖςΟγϣϯͰ ਂ૚ֶशΛ࢖ͬͨख๏Ͱطଘख๏ʹେࠩΛ͚ͭͯ༏উ <,SJ[IFWTLZFUBM /FVS*14> ೥%FFQ.JOEͷڧԽֶशϞσϧʮ"MQIB(Pʯ͕ ғޟͰ౰࣌ͷੈքع࢜ϨʔτҐͷᐬܿʹউར <4JMWFSFUBM /BUVSF> ೥0QFO"*ͷݴޠϞσϧʮ(15ʯʹΑͬͯ ·ΔͰਓ͕ؒॻ͍ͨΑ͏ͳߴਫ਼౓ͳจষΛੜ੒Մೳʹ <#SPXOFUBM /FVS*14>

Slide 11

Slide 11 text

എܠ  ੍࡞෺ͷधཁ֦େ  ੍࡞Ϧιʔεͷރׇ  ਓ޻஌ೳٕज़ͷ୆಄

Slide 12

Slide 12 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ

Slide 13

Slide 13 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ w w w w w w

Slide 14

Slide 14 text

ͲͪΒ͕ߴ඼࣭ʁ 74

Slide 15

Slide 15 text

74 DMJDLT DMJDLT ͲͪΒ͕ߴ඼࣭ʁ

Slide 16

Slide 16 text

74 DMJDLT DMJDLT ͲͪΒ͕ߴ඼࣭ʁ

Slide 17

Slide 17 text

ͲͪΒ͕ΑΓΫϦοΫ͞ΕΔʁ 74 74

Slide 18

Slide 18 text

74 74 ࣗવɾྲྀெͳ೔ຊޠ ෆࣗવͳ೔ຊޠ ݕࡧΩʔϫʔυʹϚον͍ͯ͠Δ ݕࡧΩʔϫʔυʹϚον͍ͯ͠ͳ͍ ͲͪΒ͕ΑΓΫϦοΫ͞ΕΔʁ

Slide 19

Slide 19 text

ߴ඼࣭ͳ޿ࠂจͱ͸ ҎԼͷ߲໨Λ඼࣭֬ೝͷͨΊͷ൑அࢦඪͱߟ͑Δ ͜ͱ͕Ͱ͖Δ w޿ࠂ഑৴࣮੷ wࣗવ͞ɺྲྀெ͞ wݕࡧΩʔϫʔυͱͷؔ࿈ੑ

Slide 20

Slide 20 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ w w w w w

Slide 21

Slide 21 text

جຊతͳ޿ࠂӡ༻ͷϑϩʔ ޿ࠂςΩετΛࣗಈͰ࡞Δ

Slide 22

Slide 22 text

جຊతͳ޿ࠂӡ༻ͷϑϩʔ ޿ࠂςΩετΛࣗಈͰ࡞Δ ͜͜ΛࣗಈԽ͍ͨ͠

Slide 23

Slide 23 text

ೖྗΩʔϫʔυɺ-1৘ใͳͲ ग़ྗ޿ࠂจ Ωʔϫʔυ ΢Ϛ່ɼ%..ɼ1$ɼը໘ λΠτϧ ΢Ϛ່ϓϦςΟʔμʔϏʔ%..(".&4൛ެࣜαΠτ ʛ$ZHBNFT આ໌จ ήʔϜʮ΢Ϛ່ϓϦςΟʔμʔϏʔʯ%..(".&4൛͕ ग़૸தʂ1$ͷେը໘Ͱഭྗຬ఺ͷϥΠϒɾϨʔεγʔϯ Λָ͠΋͏ʂεϚʔτϑΥϯ൛ͱσʔλ࿈ܞͯ͠༡΂Δʂ ޿ࠂςΩετΛࣗಈͰ࡞Δ

Slide 24

Slide 24 text

ςϯϓϨʔτϕʔε ޿ࠂจςϯϓϨʔτʹରͯ͠ɺద੾ʹΩʔϫʔυΛૠೖ͢Δ <#BSU[FUBM &$><'VKJUBFUBM *$&$> ܎Γड͚ؔ܎ͳͲͷߏจ৘ใͷར༻ ঎඼ͳͲͷ৘ใΛઆ໌ͨ͠௕͍จ௕ͷߏจ໦Λ࡞੒͠ɺద੾ʹ ࢬמΓ͢Δ͜ͱͰ୹͍޿ࠂจΛ࡞੒<'VKJUBFUBM *$&$> ݴޠϞσϧ -1͔Β୯ޠͷ࿈ͳΓΛநग़͠ɺۃੑ൑ఆثͰϙδςΟϒ͞Λ ߟྀͯ͠ɺݴޠϞσϧΛ࢖ͬͯ޿ࠂจੜ੒<5IPNBJEPVFUBM $*,. > ͍Ζ͍ΖͳΞϓϩʔν

Slide 25

Slide 25 text

ςΩετ͔ΒςΩετΛੜ੒͢Δܥྻม׵ϞσϧʢTFRTFRʣ <4VUTLFWFSFUBM /FVS*14>ͷొ৔ 4FRTFRϞσϧ͸ػց຋༁ɺࣗಈཁ໿ɺର࿩ॲཧͳͲͷ ࣗવݴޠੜ੒෼໺Ͱ਺ʑͷ੒ޭ ήʔϜ ΢Ϛ່ ʜ ༡΂Δʂ ΢Ϛ່ ϓϦςΟ ʜ ޷ධൃചத ΢Ϛ່ ϓϦςΟ ʜ χϡʔϥϧωοτϫʔΫ΁

Slide 26

Slide 26 text

͔͠͠ैདྷͷTFRTFRख๏͸޿ࠂͷޮՌΛ௚઀ߟྀͰ͖ͳ͍ w ޿ࠂޮՌͷ஋͸ඍ෼Ͱ͖ͳ͍ͷͰɺܭࢉάϥϑʹ૊ΈೖΕΒ Εͳ͍ w ௚઀తͳ޿ࠂޮՌҎ֎ʹ΋ɺྲྀெ͞ɺΩʔϫʔυؔ࿈౓ͳͲ ͷࢦඪ΋ߟྀ͍ͨ͠ ޿ࠂςΩετΛࣗಈͰ࡞Δ ήʔϜ ΢Ϛ່ ʜ ༡΂Δʂ ΢Ϛ່ ϓϦςΟ ʜ ޷ධൃചத ΢Ϛ່ ϓϦςΟ ʜ ޿ࠂ഑৴࣮੷ɺࣗવ͞ɾྲྀெ͞ɺݕࡧΩʔϫʔυͱͷؔ࿈ੑ

Slide 27

Slide 27 text

ڧԽֶशͷऔΓೖΕ 4FMGDSJUJDBM4FRVFODF5SBJOJOH 4$45 <3FOOJFFUBM $713> ैདྷͷTFRTFRͰܭࢉ͞ΕΔଛࣦؔ਺ͱɺαϯϓϦϯάʹΑΓಘΒΕͨτʔΫϯʹରͯ͠ใु Λܭࢉͯ͠ଛࣦؔ਺ʹՃ͑ͯ࠷దԽ͢Δ

Slide 28

Slide 28 text

ڧԽֶशͷऔΓೖΕ ڧԽֶश 4$45 Λಋೖͨ͠޿ࠂจͷࣗಈੜ੒Ϟσϧͷશମਤ<,BNJHBJUPFUBM /""$-> If you are nding the most popular insurance ... Decoder Output by Sampling!y! Decoder Output by MLE!y* Input!x BiLSTM Layer Attention Layer Context Vector c Calculating Rewards r Advertisement Quality score calculated with GBRT Fluency score calculated with LSTM Language Model Relevance score calculated with Keyword Matching Lmle : Loss of Maximum Likelihood Estimation Lrl : Loss of Reinforcement Learning Model Parameters Reference y Which insurance is the best ... Checkout the most popular insurance ... Document Tag Keywords Contents of a web-page r = rF + rR + rQ    "*-BCɺࣗવݴޠॲཧ෼໺ͷτοϓΧϯϑΝϨϯεʮ/""$-)-5ʯʹͯڞஶ࿦จ࠾୒ɹʕ޿ࠂޮՌΛߟྀͨ͠޿ࠂจੜ੒ख๏ΛఏҊʕ cגࣜձࣾαΠόʔΤʔδΣϯτ

Slide 29

Slide 29 text

ڧԽֶशͷऔΓೖΕ ڧԽֶश 4$45 ʹΑΓ޿ࠂޮՌ΍ͦͷଞͷࢦඪΛใुͱͯ͠ѻ͏͜ͱͰֶशՄೳʹ If you are nding the most popular insurance ... Decoder Output by Sampling!y! Decoder Output by MLE!y* Input!x BiLSTM Layer Attention Layer Context Vector c Calculating Rewards r Advertisement Quality score calculated with GBRT Fluency score calculated with LSTM Language Model Relevance score calculated with Keyword Matching Lmle : Loss of Maximum Likelihood Estimation Lrl : Loss of Reinforcement Learning Model Parameters Reference y Which insurance is the best ... Checkout the most popular insurance ... Document Tag Keywords Contents of a web-page r = rF + rR + rQ    "*-BCɺࣗવݴޠॲཧ෼໺ͷτοϓΧϯϑΝϨϯεʮ/""$-)-5ʯʹͯڞஶ࿦จ࠾୒ɹʕ޿ࠂޮՌΛߟྀͨ͠޿ࠂจੜ੒ख๏ΛఏҊʕ cגࣜձࣾαΠόʔΤʔδΣϯτ If you are nding the most popular insurance ... Decoder Output by Sampling!y! Decoder Output by MLE!y* Input!x BiLSTM Layer Attention Layer Context Vector c Calculating Rewards r Advertisement Quality score calculated with GBRT Fluency score calculated with LSTM Language Model Relevance score calculated with Keyword Matching Lmle : Loss of Maximum Likelihood Estimation Lrl : Loss of Reinforcement Learning Model Parameters Reference y Which insurance is the best ... Checkout the most popular insurance ... Document Tag Keywords Contents of a web-page r = rF + rR + rQ    ࣗવ͞ɺྲྀெ͞ Flu  ੜ੒݁Ռʹରͯ͠ɺݴޠϞσϧͰ ࢉग़͞ΕΔʮΒ͠͞ʯ ݕࡧΩʔϫʔυͱͷؔ࿈ੑ Rel  ੜ੒݁ՌͷΩʔϫʔυʹର͢ΔΧόʔ཰ͱ ΩʔϫʔυͷҐஔ ޿ࠂ഑৴࣮੷ QS  ੜ੒݁Ռʹରͯ͠ɺաڈͷ޿ࠂ഑৴ͰಘΒΕͨ σʔλͰ܇࿅ͨ͠ճؼϞσϧʹΑͬͯࢉग़͞ΕΔ ਪఆ඼࣭஋

Slide 30

Slide 30 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ w w w w w w w

Slide 31

Slide 31 text

ੜ੒͞Εͨ޿ࠂͷධՁ ධՁ͢ΔͨΊͷํ๏ wఆΊΒΕͨࢦඪͰࣗಈͰධՁʢࣗಈධՁʣ wਓ͕ؒ௚઀ݟͯධՁʢਓखධՁʣ w࣮ࡍʹ഑৴ͯ͠ධՁʢ഑৴ධՁʣ ΦϑϥΠϯධՁ ΦϯϥΠϯධՁ

Slide 32

Slide 32 text

ΦϑϥΠϯධՁ ਓखධՁͰ͸ɺ޿ࠂ੍࡞ऀͱΤϯυϢʔβʔΛ૝ఆͨ͠ Ϋϥ΢υιʔγϯάͦΕͧΕʹΑΔԼͷ߲໨ͷධՁ w ྲྀெੑ Fluency  w ັྗ Attractive.  w Ωʔϫʔυͱͷؔ࿈ੑ Relevance Model Copywriter Crowdsourcing Fluency Attractive. Relevance Fluency Attractive. Relevance Reference 87.5 25.5 24.4 75.6 26.8 29.1 Seq2Seq 83.3 25.1 23.7 64.5 23.8 26.1 + Flu, QS 81.7 25.3 22.8 64.3 24.4 26.6 + Flu, Rel 77.5 24.2 23.7 60.9 24.8 26.2 + Flu, Rel, QS 81.2 23.9 24.3 62.7 25.4 26.9

Slide 33

Slide 33 text

ΦϯϥΠϯධՁ ࣗಈੜ੒͞Εͨ޿ࠂจΛ࣮ࡍʹ഑৴ͯ͠ҎԼͷ߲໨Λɺ ਓख੍࡞޿ࠂͱͷഒ཰ΛධՁ w දࣔճ਺ Impression  w ΫϦοΫ཰ CTR  w ফඅ༧ࢉ Cost Model Impression CTR Cost Seq2Seq 3.54x 0.66x 3.31x + Flu, Rel 3.80x 0.52x 3.62x + Flu, Rel, QS 1.32x 0.71x 2.58x

Slide 34

Slide 34 text

੍࡞͠ଓ͚ΔͨΊʹ ੜ੒⁶ධՁͷ܁Γฦ͠ ͜ͷϧʔϓΛΑΓߴ଎ʹɺΑΓޮ཰తʹߦ͏ඞཁ

Slide 35

Slide 35 text

'"45<,BXBNPUPFUBM &./-1> ਓखධՁΛΑΓޮ཰తʹߦ͏πʔϧͷ։ൃ $ZCFS"HFOUGBTUBOOPUBUJPOUPPM ࣗવݴޠॲཧ෼໺ͷτοϓΧϯϑΝϨϯεʮ&./-1ʯͷ4ZTUFN%FNPOTUSBUJPO5SBDLʹͯ࿦จ ࠾୒ʔϞόΠϧ୺຤༻ͷޮ཰తͳΞϊςʔγϣϯπʔϧΛఏҊʔcגࣜձࣾαΠόʔΤʔδΣϯτ

Slide 36

Slide 36 text

'"45<,BXBNPUPFUBM &./-1> ଞπʔϧͱͷ࢖༻ײͷධՁ w ࡞ۀޮ཰ Efficiency  w ࡞ۀਫ਼౓ Quality  w ࢖༻ײ Usability Tool Efficiency (!) Quality (") Usability (!) Baseline (Mobile) 6.7 0.98 5.82 FAST (Mobile, Card UI) 4.6 0.97 2.03 FAST (Mobile, Multi-label UI) 4.9 0.97 3.15

Slide 37

Slide 37 text

ݚڀ੒ՌͷϓϩμΫτ΁ͷ׆༻ ࠓճ঺հͨ͠಺༰ΛؚΊɺଞʹ΋ଟ͘ͷख๏΍πʔϧ͕ݚڀ։ൃ͞Εɺ ࣮ࡍʹϓϩμΫτʹಋೖ͞Εɺݕূ͞Ε͍ͯΔ

Slide 38

Slide 38 text

ߴ඼࣭ͳ޿ࠂΛ ࣗಈͰ੍࡞͠ଓ͚͍ͨ ͜Ε͔Β΋

Slide 39

Slide 39 text

͜Ε͔Β ͞ΒͳΔੜ੒඼࣭ͷ޲্ w ྲྀெੑ w ଟ༷ੑ w ஧࣮ੑ ΑΓଟछͳσʔλͷ׆༻ w ϚϧνϞʔμϧʢը૾ɺ-1ϨΠΞ΢τʣ <.VSBLBNJFUBM ݴޠॲཧֶձ> ࢼߦࡨޡͷͨΊͷ؀ڥ w ֤छධՁج൫ͷ੔උ w σʔλऩूͷͨΊͷج൫੔උ

Slide 40

Slide 40 text

3FGFSFODFT w<,SJ[IFWTLZFUBM /FVS*14>*NBHF/FU$MBTTJpDBUJPOXJUI%FFQ$POWPMVUJPOBM/FVSBM/FUXPSLT w<4JMWFSFUBM /BUVSF>.BTUFSJOHUIFHBNFPG(PXJUIEFFQOFVSBMOFUXPSLTBOEUSFFTFBSDI w<#SPXOFUBM /FVS*14>-BOHVBHF.PEFMTBSF'FX4IPU-FBSOFST w<#BSU[FUBM &$>/BUVSBMMBOHVBHFHFOFSBUJPOGPSTQPOTPSFETFBSDIBEWFSUJTFNFOUT w<'VKJUBFUBM *$&$>"VUPNBUJDHFOFSBUJPOPGMJTUJOHBETCZSFVTJOHQSPNPUJPOBMUFYUT w<5IPNBJEPVFUBM $*,.>"VUPNBUFETOJQQFUHFOFSBUJPOGPSPOMJOFBEWFSUJTJOH w<4VUTLFWFSFUBM /FVS*14>4FRVFODFUP4FRVFODF-FBSOJOHXJUI/FVSBM/FUXPSLT w<3FOOJFFUBM $713>4FMGDSJUJDBM4FRVFODF5SBJOJOHGPS*NBHF$BQUJPOJOH w<,BNJHBJUPFUBM /""$->"O&NQJSJDBM4UVEZPG(FOFSBUJOH5FYUTGPS4FBSDI&OHJOF"EWFSUJTJOH w<,BXBNPUPFUBM &./-1>'"45'BTU"OOPUBUJPOUPPMGPS4NBS5EFWJDFT w<.VSBLBNJFUBM ݴޠॲཧֶձ>-1UP5FYUϚϧνϞʔμϧ޿ࠂจੜ੒