Yansong Feng
Fri 31 Aug 2018, 11:00 - 12:30
Informatics Forum (IF-4.31/4.33)

If you have a question about this talk, please contact: Diana Dalla Costa (ddallac)

Abstract:

Recent success of many neural network models in natural language processing (NLP) tasks is often bound by the number and quality of annotated data, but there is often a shortage of such training data. In this talk, I will discuss how we can  combine a neural network model with human knowledge, such as regular expressions or hand-crafted rules, to improve supervised NLP solutions when the training data is not perfect.  Firstly, we develop methods to exploit the rich expressiveness of regular expressions at different levels within a neural network model, showing that the combination can significantly enhance the learning effectiveness when a small number of training examples are available. In the second attempt, we encode human-crafted constraints as a semantic loss to teach a neural network model to behave as we expect, where the learned model can, to some extent, effectively work on imperfect training data that can be easily obtained. 

Bio:

Dr. Yansong Feng is an associate professor in the Institute of Computer Science and Technology at Peking University. Before that, he worked with Prof. Mirella Lapata and obtained his PhD from ICCS (now ILCC) at the University of Edinburgh. His current research interests include using probabilistic methods to distill knowledge from large volumes of natural language texts, and supporting intelligent human-computer interfaces, such as question answering and dialogue. Yansong received the IBM Faculty Award in 2014 and 2015,  and the IBM Global Shared University Research Award in 2016.