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

文献紹介: A Persona-Based Neural Conversation Model

Yumeto Inaoka
February 28, 2018
Science
0
190

文献紹介: A Persona-Based Neural Conversation Model

2018/02/28の文献紹介で発表

Yumeto Inaoka

February 28, 2018
Tweet

More Decks by Yumeto Inaoka

See All by Yumeto Inaoka
文献紹介: Quantity doesn’t buy quality syntax with neural language models
yumeto
1
50
文献紹介: Open Domain Web Keyphrase Extraction Beyond Language Modeling
yumeto
0
75
文献紹介: Self-Supervised_Neural_Machine_Translation
yumeto
0
57
文献紹介: Comparing and Developing Tools to Measure the Readability of Domain-Specific Texts
yumeto
0
56
文献紹介: PAWS: Paraphrase Adversaries from Word Scrambling
yumeto
0
32
文献紹介: Beyond BLEU: Training Neural Machine Translation with Semantic Similarity
yumeto
0
120
文献紹介: EditNTS: An Neural Programmer-Interpreter Model for Sentence Simplification through Explicit Editing
yumeto
0
150
文献紹介: Decomposable Neural Paraphrase Generation
yumeto
0
92
文献紹介: Analyzing the Limitations of Cross-lingual Word Embedding Mappings
yumeto
0
100

Other Decks in Science

See All in Science
Presenting Effectively with Data (in a Hurry)
thomaselove
1
110
実践 時系列解析 輪読会 #6
tok41
0
110
回転座標系での古典場の方程式について。微分形式からの導出
deepflow
0
150
不審なURLの見つけ方
secchick
1
250
計算量理論
hn410
0
390
プロバスケットボール・B.LEAGUEにおけるインパクトメトリクスの提案 / Proposal of an impact metrics in B.LEAGUE
konakalab
0
120
実験データ解析再入門：論文を「フェイクニュース」にしないために
ykamit
18
9.4k
Tokyo.R コレスポンデンス分析の正しい使い方
bob3bob3
1
1.3k
SNLP2022:What does the sea say to the shore? A BERT based DST style approach for speaker to dialogue attribution in novels
yukizenimoto
0
110
四元数と回転 / Quaternion Rotation
usamik26
0
410
Pythonで学ぶSynthetic Difference in Differences
masa_asa
0
1.3k
Causal Inference: What If, Chapter1
norihirosuzuki
0
850

Featured

See All Featured
In The Pink: A Labor of Love
frogandcode
131
21k
Teambox: Starting and Learning
jrom
124
7.8k
Put a Button on it: Removing Barriers to Going Fast.
kastner
56
2.4k
BBQ
matthewcrist
74
8k
VelocityConf: Rendering Performance Case Studies
addyosmani
316
22k
JavaScript: Past, Present, and Future - NDC Porto 2020
reverentgeek
37
3.4k
How to name files
jennybc
44
66k
How GitHub Uses GitHub to Build GitHub
holman
465
280k
The Web Native Designer (August 2011)
paulrobertlloyd
75
2k
JazzCon 2018 Closing Keynote - Leadership for the Reluctant Leader
reverentgeek
174
8.7k
Code Review Best Practice
trishagee
45
10k
Cheating the UX When There Is Nothing More to Optimize - PixelPioneers
stephaniewalter
270
12k

Transcript

  1. A Persona-Based
 Neural Conversation Model Jiwei Li, Michel Galley, Chris

    Brockett, Georgios Spithourakis,
 Jianfeng Gao, and Bill Dolan. Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 994 - 1003, 2016. จݙ঺հ` ௕Ԭٕज़ՊֶେֶɹࣗવݴޠॲཧݚڀࣨɹҴԬເਓ

  2. "CTUSBDU w࿩ऀҰ؏ੑͷ໰୊Λѻ͏ฦ౴ੜ੒Ϟσϧ wܦྺ΍ελΠϧͷΑ͏ͳݸੑΛೖྗʹ෇Ճ wQFSQMFYJUZ #-&6ͷ྆ํͰੑೳ͕޲্ wਓखධՁͰ΋Ұ؏ੑʹ͓͍ͯੑೳ͕޲্ 2

  3. *OUSPEVDUJPO wେྔͷਓؒରਓؒͷର࿩ʹΑΔࣗવͳର࿩γεςϜͷ
 ߏங͕஫໨ΛूΊ͍ͯΔ w܇࿅σʔλ಺ͷ୅දతͳฦ౴Λฦ͢܏޲͕͋Δ
 ˠͦͷΑ͏ͳฦ౴ͷ໬౓͕ߴ͘ͳΓ΍͍ͨ͢Ί wໃ६ͨ͠ฦ౴Λฦ͢͜ͱ͕͋Δ wຊ࿦จͰ͸Ұ؏ੑͱݸੑͷ໰୊ʹ͍ͭͯऔΓ૊Ή 3

  4. *OUSPEVDUJPO wେྔͷਓؒରਓؒͷର࿩ʹΑΔࣗવͳର࿩γεςϜͷ
 ߏங͕஫໨ΛूΊ͍ͯΔ w܇࿅σʔλ಺ͷ୅දతͳฦ౴Λฦ͢܏޲͕͋Δ
 ˠͦͷΑ͏ͳฦ౴ͷ໬౓͕ߴ͘ͳΓ΍͍ͨ͢Ί wໃ६ͨ͠ฦ౴Λฦ͢͜ͱ͕͋Δ wຊ࿦จͰ͸Ұ؏ੑͱݸੑͷ໰୊ʹ͍ͭͯऔΓ૊Ή 4

  5. *OUSPEVDUJPO wେྔͷਓؒରਓؒͷର࿩ʹΑΔࣗવͳର࿩γεςϜͷ
 ߏங͕஫໨ΛूΊ͍ͯΔ w܇࿅σʔλ಺ͷ୅දతͳฦ౴Λฦ͢܏޲͕͋Δ
 ˠͦͷΑ͏ͳฦ౴ͷ໬౓͕ߴ͘ͳΓ΍͍ͨ͢Ί wໃ६ͨ͠ฦ౴Λฦ͢͜ͱ͕͋Δ wຊ࿦จͰ͸Ұ؏ੑͱݸੑͷ໰୊ʹ͍ͭͯऔΓ૊Ή 5

  6. ؔ࿈ݚڀ w3JUUFSΒ  ͸౷ܭతػց຋༁ͷ໰୊ͱͯ͠औΓ૊Μͩ w4FSCBOΒ  ͸ର࿩ཤྺ΁ͷґଘؔ܎Λิ଍͢Δ͜ͱΛ ໨తͱͨ͠֊૚తFODPEFSEFDPEFSϞσϧΛఏҊ w-JΒ 

    ͸యܕతԠ౴ͷׂ߹ΛݮΒͨ͢Ίʹ
 ࠷େ໬౓ .-& Ͱ͸ͳ͘૬ޓ৘ใྔ ..* Λ໨తؔ਺ͱ͢Δ TFRTFRγεςϜΛఏҊ 6

  7. ఏҊϞσϧ 7

  8. ఏҊϞσϧ wதؒ૚Ϣχοτʹ-45.Λ༻͍ͨ3// w࠷ޙͷग़ྗΛ%FDPEFSʹ౉͢ 8 &ODPEFS

  9. ఏҊϞσϧ wதؒ૚Ϣχοτʹ-45.Λ༻͍ͨ3// w&ODPEFSͷग़ྗΛ%FDPEFSʹೖྗ w4QFBLFS&NCFEEJOHΛ֤ӅΕ૚ͰՃࢉ 9 %FDPEFS

  10. ఏҊϞσϧ w4QFBLFS.PEFM
 ฦ౴ऀͷݸੑͷΈΛߟྀ
 4QFBLFS&NCFEEJOHΛೖྗ w4QFBLFS"EESFTTFF.PEFM
 ฦ౴ऀͱฉ͖खͷ྆ํΛߟྀ
 ԼࣜͰ4QFBLFS&NCFEEJOHΛ߹੒ 10

  11. %FDPEJOHBOE3FSBOLJOH ɹ.ೖྗจɹ3ฦ౴จɹc3cฦ౴จ௕
 ɹW4QFBLFS*%ɹЕ Ѝௐ੔ύϥϝʔλ w#FBN4FBSDI࣌ʹ্ࣜͷධՁؔ਺Ͱ3FSBOLJOHΛߦ͏ wయܕతͰͳ͘௕͍จ͕༏ઌ͞ΕΔ wɹɹɹɹɹ͸3͔Β.Λग़ྗ͢ΔTFRTFRΛֶशͯ͠ܭࢉ 11

  12. σʔληοτ w5XJUUFS1FSTPOB%BUBTFU
 ೥݄͔Βϲ݄෼ͷ5XJUUFS'JSF)PTFΛ࢖༻
 ظؒ಺ʹճҎ্λʔϯͷձ࿩Λͨ͠Ϣʔβʹݶఆ
  ϢʔβʹΑΔ  ૊ͷձ࿩ؚ͕·ΕΔ
 ಉϢʔβʹΑΔ೥݄͔Βϲ݄෼ͷձ࿩Λ૊
 ͣͭ։ൃ

    ݕূ ςετηοτͱͯ͠ઃఆ
 ฦ౴ऀͷ4QFBLFS*%ͷΈ͕ೖ͍ͬͯΔͨΊ4QFBLFS.PEFM
 ͷΈʹར༻ 12

  13. σʔληοτ w5XJUUFS4PSEPOJ%BUBTFU 4PSEPOJ  
 ैདྷͷ405"ͱͷൺֱͷͨΊʹ࢖༻
 ςετηοτͷΈ࢖༻
  ૊ͷձ࿩σʔλ
 ͭͷೖྗจʹରͯ͠࠷େݸͷฦ౴


    ˠ5XJUUFS1FSTPOB%BUBTFUͱͷ#-&6ͷൺֱ͸Ͱ͖ͳ͍ 13

  14. σʔληοτ w5FMFWJTJPO4FSJFT5SBOTDSJQUT%BUB
 57γϦʔζl'SJFOETz l5IF#JH#BOH5IFPSZzͷࣈນ
 ਓͷओཁਓ෺ʹΑΔ ૊ͷձ࿩
 ͏ͪ։ൃ ςετηοτͱͯͦ͠ΕͧΕ ૊࢖༻ w0QFO4VCUJUMFT


    ϊΠζΛؚΉ.ʙ.ͷࣈນσʔληοτ
 5FMFWJTJPO4FSJFT5SBOTDSJQUT%BUBͷن໛͕খ͍ͨ͞Ί
 ຊσʔληοτͰυϝΠϯదԠΛߦ͏ 14

  15. ֶशͷৄࡉ wMBZFS-45. w IJEEFODFMMTGPSFBDIMBZFS w#BUDITJ[F w-FBSOJOHSBUF w<>ͷҰ༷෼෍ͰύϥϝʔλΛॳظԽ w5ISFTIPMEGPSHSBEJFOUDMJQQJOH w7PDBCVMBSZTJ[F 

    w%SPQPVUSBUF w#FBNTJ[F 15

  16. ݁Ռ w5XJUUFS4PSEPOJEBUBTFUʹ͓͚ΔධՁ w.5CBTFMJOF͸4.5ʹΑΔख๏ wPVSTZTUFN͸5XJUUFS1FSTPOB%BUBTFUͰֶशͨ͠΋ͷ wֶशίʔύεͷن໛ %SPQPVUͷ࢖༻ ର৅Ϣʔβͷબผ͕
 վળͷཧ༝ͱߟ͑ΒΕΔ 16

  17. ݁Ռ w5XJUUFS1FSTPOBEBUBTFUʹ͓͚ΔධՁ w.-&ͷ৔߹͸  ..*ͷ৔߹͸ ͷվળ wఏҊख๏͸..*ΑΓ.-&ʹΑΓ༗ӹ 17

  18. ݁Ռ w57TFSJFTσʔληοτʹ͓͚ΔධՁ w4QFBLFS.PEFM 4QFBLFS"EESFTTFF.PEFMͷ͍ͣΕ΋
 #-&6είΞΛ޲্ͤ͞Δ wఏҊ͢ΔͭͷϞσϧͷؒʹେ͖ͳҧ͍͸ͳ͍
 ˠਓͷύλʔϯ͕ัଊͰ͖ΔఔσʔλαΠζ͕େ͖͘ͳ͍ 18

  19. ݁Ռ w5XJUUFS1FSTPOB%BUBTFUͷ։ൃσʔλͱ
 57TFSJFTEBUBTFUͰͦΕͧΕQFSQMFYJUZΛൺֱ w5XJUUFSͷํ͕ߴ͘ͳΔͷ͸ϊΠζͷͨΊͱߟ͑ΒΕΔ 19

  20. ݁Ռ wϥϯμϜʹਓͷ4QFBLFS&NCFEEJOHΛ
 4QFBLFS.PEFMʹೖྗ 20

  21. ݁Ռ w4QFBLFS"EESFTTFF.PEFM
 ͷධՁ wฦ౴ऀʹහײͰ͋Δ͜ͱ͕
 ୯ޠ͔Β෼͔Δ wlIJNz͔ΒੑผΛਖ਼͘͠
 ೝ͍ࣝͯ͠Δ͜ͱ͕෼͔Δ 21

  22. ਓखධՁ wΫϥ΢υιʔγϯάΛ࢖ͬͯग़ྗΛධՁ w4QFBLFS*%ຖʹग़ྗ͕Ұ؏͍ͯ͠Δ͔Λ࣮ݧ wϕʔεϥΠϯͱ1FSTPOB.PEFMͷग़ྗΛൺֱͯ͠
 ʮҰ؏͍ͯ͠Δʯ͸  ʮ΍΍Ұ؏͍ͯ͠Δʯ͸  
 ಉఔ౓Ͱ͋Δ৔߹͸ͷείΞΛ෇͚Δ

    wਓͷධՁऀͷείΞΛฏۉ͠ɺ౳෼ͷ࠶ׂ౰Λߦ͏ 22

  23. ਓखධՁ݁Ռ wಉఔ౓ͷ৔߹Λແࢹ͢Δͱɺͷࣄྫʹ͓͍ͯ
 1FSTPOB.PEFM͕ʮҰ؏͍ͯ͠Δʯʮ΍΍Ұ؏͍ͯ͠Δʯ ͱ൑ఆ͞Εͨ wʮ΍΍Ұ؏͍ͯ͠ΔʯΛແࢹ͢Δͱɺ1FSTPOB.PEFM͕
 ࣄྫͷͰ༏ҐͱͳΓɺϕʔεϥΠϯ͸ʹཹ·Δ 23

  24. ࣮ࡍͷग़ྗࣄྫ 24

  25. ࣮ࡍͷग़ྗࣄྫ 25

  26. ݁࿦ w1FSTPOBCBTFEͷԠ౴ੜ੒ϞσϧΛఏࣔ w#-&6 QFSQMFYJUZ Ұ؏ੑͷਓखධՁʹ͓͍ͯ
 ܶతͰ͸ͳ͍΋ͷͷϕʔεϥΠϯΛ্ճΔ݁Ռ wฦ౴ऀฉ͖खͷਓΛೖྗ͢Δ͜ͱʹϝϦοτ͕͋Δ͜ͱ ͕4QFBLFS"EESFTTFFϞσϧͷ݁ՌͰࣔ͞Εͨ 26