Machine Learning Seminar by Prof. Kenny Q. Zhu – Recent Advances in Dialogues and Text Generation at ADAPT-Lab of SJTU

Kenny Q. Zhu is a Professor of computer science at Shanghai Jiao Tong University. He graduated with B.Eng (Hons) in Electrical Engineering in 1999 and PhD in Computer Science in 2005 from National University of Singapore. He was a postdoctoral researcher and lecturer from 2007 to 2009 at Princeton University. Prior to that, he was a software design engineer at Microsoft, Redmond, WA. From Feb 2010 to Aug 2010, he was a visiting professor at Microsoft Research Asia in Beijing. Kenny’s current research interests are natural language processing and knowledge engineering. He has published extensively in AI and NLP at top venues such as ACL, EMNLP, AAAI, IJCAI, SIGMOD, KDD, CIKM, ICDE. He has served on the PC of AAAI, IJCAI, WWW, CIKM, ECML, COLING, SAC, WAIM, APLAS and NDBC, etc. His research has been supported by NSF China, MOE China, Microsoft, Google, Oracle, Morgan Stanley and AstraZeneca. Kenny is the winner of the 2013 Google Faculty Research Award and 2014 DASFAA Best Paper Award.​

In this talk, I will present several recent research projects in dialog generation and text generation. First, I will attack the problem of automatically selecting a response in an open-domain dialogue, from a list of candidate responses. This is a form of dialogue system which can be modeled as a classification problem. Our main approach is to segment the dialogue context into sub-dialogues through the use of a special discourse parsers, and hence gain better and more efficient representation of the context. Then, I will present a novel problem of clarification question generation for product descriptions. When writing a brief textual description for a product, one may omit certain important details such as the dimensions and certain key features. Our techniques can be used in a writing assistant, that can automatically ask questions about these missing details given a product description. Finally, I will talk about a novel method to do abstractive summarization of text documents. Our method is driven by the intuition that human summarizes a text by first selecting a number of key sentences and then rephrase these sentences. We focus on a new effective way of doing the sentence selection that can achieve state-of-the-art results on the popular CNN-Daily and DUC benchmarks.