Dialogue robots, called Dialog Systems in the NLP field, are
essentially making machines understand human language through machine learning
and other AI technologies. Because it relates to in-depth human language, this direction
is also a very difficult and comprehensive direction in AI technology. Jiwei Li's
innovative work will occupy an important part in the evolution of the dialogue
system in coming years.
The undergraduate, who studied at the School of Life Sciences
of Peking University, completely "derailed" from the computer field
at the Ph.D. stage, and finally found his true love in the sea of computer
science, NLP. He became the first person in Stanford's history to spend 3 years in earning a doctorate in computer science.
In 2014, Jiwei Li decided to give up his study at the
Department of Biomedical Engineering at Cornell University and join Prof. Dan
Jurafsky’s research group, which is a language processing group at the School of Computer Science at Stanford University. Since then, he has worked on paragraph
analysis, machine translation, artificial dialogue generation, and has become
one of the first group of pioneers who successfully used neural networks to
improve dialogue systems.
Jiwei Li completed one task in 2015 to make machine answers
more meaningful rather than just returning universal responses such as, "Ha." In the multi-round dialogue, such responses often appear but
have no practical significance. His work, "A Diversity-Promoting
Objective Function for Neural Conversation Models," explores the
possibility of reducing these kinds of nonsense answers by using the mainstream
model Seq2Seq in the dialogue generation task to significantly improve the
quality of the generated dialogue system.
His paper, "Deep Reinforcement Learning for Dialogue
Generation," received by EMNLP in 2016, explores the above issues in
greater depth. Jiwei Li's method of training dialogue neural networks with
reinforcement learning shows that reinforcement of deep learning can help improve the number of dialogue cycles in the dialogue system and the diversity of
words in dialogue.
Now, Jiwei Li has set up his own company, "ShannonAI,"
in China with the hope that NLP technology can intelligently help upgrade traditional fields, such as finance. At the same time, he will also continue exploring the application of neural networks to NLP, including man-machine
dialogue systems based on deep learning, information extraction based on the
combination of common sense knowledge and corpus data, the study of
interpretable mechanism of deep learning model, natural language semantic unit
representation and language generation, and more.
