Diese Präsentation wurde erfolgreich gemeldet.
Wir verwenden Ihre LinkedIn Profilangaben und Informationen zu Ihren Aktivitäten, um Anzeigen zu personalisieren und Ihnen relevantere Inhalte anzuzeigen. Sie können Ihre Anzeigeneinstellungen jederzeit ändern.

論文紹介: Mojitalk: generating emotional responses at scale

2.266 Aufrufe

Veröffentlicht am

ACL2018 読み会の発表資料です

Veröffentlicht in: Ingenieurwesen
  • Als Erste(r) kommentieren

論文紹介: Mojitalk: generating emotional responses at scale

  1. 1. MOJITALK: Generating Emotional Responses at Scale ACL2018
  2. 2. : (B4, ) : , , : NLP 1 , CV : , 3D , NMT, HCI
  3. 3. 🐠 MOJITALK: Generating Emotional Responses at Scale : ▶︎ ▶︎ ( ) ▶︎ #kawaii_paper_name
  4. 4. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 4.Dataset 5.Model 6.Experimental Results 7.Conclusion & Future Work
  5. 5. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 🐶 2.Conclusion 3.Related Work 4.Dataset 5.Model 6.Experimental Results 7.Conclusion & Future Work
  6. 6. … Abstract 🐥
  7. 7. Introduction 🐶 : ▶︎ (emotional) : ▶︎ … [Pang et al., 2002; Maas et al., 2011; Socher et al., 2013]
  8. 8. Introduction 🐶 + ▶︎ [Go et al., 2016], [Li et al., 2017b] ▶︎ [Pang et al., 2002; Maas et al., 2011; Socher et al., 2013] Twitter
  9. 9. Introduction 🐶 Original tweet( ) Response( )
  10. 10. Introduction 🐶 ▶︎ IR [Huang et al., 2017] ▶︎ CVAE VAE [Zhao et al., 2017] ▶︎ sentence-to-emoji classifier[Feibo et al., 2017]
  11. 11. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 🐱 3.Related Work 4.Dataset 5.Model 6.Experimental Results 7.Conclusion & Future Work
  12. 12. Conclusion 🐱
  13. 13. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 🐭 4.Dataset 5.Model 6.Experimental Results 7.Conclusion & Future Work
  14. 14. Related Work 🐭 Twitter Emoji2vec [Eisner et al., 2016] ▶︎
  15. 15. Related Work 🐭 Twitter Emoji2vec [Eisner et al., 2016] ▶︎
  16. 16. Related Work 🐭 Twitter DeepMoji [Felbo et al., 2017] ▶︎ bi-LSTM https://deepmoji.mit.edu/
  17. 17. Related Work 🐭 ConditionalVAE [Sohn et al., 2015] ▶︎ ( )
  18. 18. Related Work 🐭 [Li et al., 2016][Li et al., 2017] ▶︎ 
 Action , State , Policy[pi, qi] pRL(pi+1 |pi, qi)
  19. 19. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 4.Dataset 🐹 5.Model 6.Experimental Results 7.Conclusion & Future Work
  20. 20. Dataset 🐹 ▶︎ SNS ▶︎ 64 DeepMoji
  21. 21. Dataset 🐹 ▶︎ 2017 8 12~14 ▶︎ ▶︎ URL, , ▶︎
  22. 22. Dataset 🐹 ▶︎ ▶︎
  23. 23. Dataset 🐹 ▶︎ ▶︎ yess yes
  24. 24. Dataset 🐹
  25. 25. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 4.Dataset 5.Model 🐰 6.Experimental Results 7.Conclusion & Future Work
  26. 26. Model 🐰 seq2seq + Attention [Luong et al., 2015] ▶︎ ▶︎ ▶︎ embedding->Dense
  27. 27. Model 🐰 CVAE [Sohn et al., 2015] p(x|c) = ∫ p(x|z, c)p(z|c)dz [vo; ve]
  28. 28. Model 🐰 CVAE p(x|c) = ∫ p(x|z, c)p(z|c)dz ▶︎ cf)VAE BoW loss
  29. 29. Model 🐰 CVAE recog/prior network Reparameterization trick[Kingma and Welling, 2013] recog net z decoder Prior net
  30. 30. Model 🐰 CVAE CVAE Attention ( ) ▶︎ VAE Decoder [Bowman et al., 2015] KL loss construction loss ▶︎ KL annealing early stopping [Bowman et al., 2015] bag-of-words loss [Zhao et al., 2017]
  31. 31. Model 🐰 Reinforced CVAE ▶︎ CVAE policy gradient( ) policy gradient bi-GRU [Felbo et al., 2017] ▶︎ policy training x c CVAE x’ x’ R x emoji clf
  32. 32. Model 🐰 Reinforced CVAE REINFORCE ▶︎ ▶︎ Reinforce hybrid objective α ∈ [0,1]
  33. 33. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 4.Dataset 5.Model 6.Experimental Results 🦊 7.Conclusion & Future Work
  34. 34. Experimental Results 🦊 CVAE seq2seq ▶︎ Perplexity ( ) : www.slideshare.net/hoxo_m/perplexity ▶︎ top5 ▶︎ SotA
  35. 35. Experimental Results 🦊 ▶︎ seq2seq ▶︎ CVAE ▶︎ 1-gram, 2-gram, 3-gram ▶︎ Reinforced CVAE (?)
  36. 36. Experimental Results 🦊 ▶︎ top5 CVAE Reinforced CVAE
  37. 37. Experimental Results 🦊 ▶︎ Amazon Mechanical Turk 100 5 ▶︎
  38. 38. Experimental Results 🦊 ▶︎ Turing Test ▶︎ 18% 27% ▶︎ (inter-rater reliability)
  39. 39. Experimental Results 🦊 ▶︎ seq2seq I’m ( ) ▶︎ CVAE ▶︎ Reinforced CVAE
  40. 40. MOJITALK: Generating Emotional Responses at Scale 1.Introduction 2.Conclusion 3.Related Work 4.Dataset 5.Model 6.Experimental Results 7.Conclusion & Future Work 🐮
  41. 41. Conclusion & Future Work 🐮
  42. 42. Thank you for listening 👻

×