Generating Informative and Diverse Conversational Responses via Adversarial Information Maximization (NIPS2018) Yizhe Zhang, Michel Galley, Jianfeng Gao, Zhe Gan, Xiujun Li, Chris Brockett, Bill Dolan Microsoft Research, Redmond, WA, USA 紹介者:品川 政太朗(NAIST/RIKEN) 2018/11/11 2018ⒸSeitaro Shinagawa AHC-lab NAIST ※Figures without citation are quoted from the authors’ paper 1/25
2018/11/11 2018ⒸSeitaro Shinagawa AHC-lab NAIST 自己紹介 Favorite model(?): Interest: Interaction between human and machine Research Topic: Dialog based Image generation 1989 Born in Sapporo 2009-2015 Tohoku Univ. 2015- NAIST(Ph.D student) 2/25
2018/11/11 2018ⒸSeitaro Shinagawa AHC-lab NAIST 著者らの注目点 diverseかつinformativeな応答を生成できるようにしたい diverse informative • I don’t know. • I haven’t clue. • I haven’t the foggiest etc... I don’t know. • I like music. • I like jazz. etc... I like music. 発話:”What is your hobby?”に対しての応答例 7/25 Informativeな文とは?:何らかの情報が得られる文 音楽、ジャズが好きだと いう情報が入っている
2018/11/11 2018ⒸSeitaro Shinagawa AHC-lab NAIST 余談:InfoGAN [Chen+, NIPS2016] z real/fake Gen Dis real fake c • c: discrete latent code • z: vector derived from random noise • c’: predicted latent code learning to make c and G(z,c) highly correlated c’ Maximize mutual information ; , The point for disentanglement 13/25
2018/11/11 2018ⒸSeitaro Shinagawa AHC-lab NAIST How to maximize I(c;G(z,c)) using Q distro. z real/fake Gen Dis real fake c c’ Lemma 5.1 random variable X,Y, function f ~, ~|, ′~| (′, ) = ~, ~| , ; , = − , = ~ , ′~ log ′ + = ~ ,
′ ∥ ′ + ′~ log ′ + ≥ ~ , ′~ log ′ + = ~ , ′~ log ′ + = ~ , ~ , log + loss between c and c’ 14/25