[2017-06-26] Dr. Ming-Wei Chang, Microsoft,"Learning Dynamic Neural Semantic Parsers with Implicit Supervision Signals"


Title: Learning Dynamic Neural Semantic Parsers with Implicit Supervision Signals
Date: 2017-06-26 10:00am-11:00am
Location: R210, CSIE
Speaker: Dr. Ming-Wei Chang,Microsoft
Hosted by: Prof. SD. Lin


Semantic parsing, the task of translating natural language to executable code, is important because it allows users to access and interact with external databases or physical environments in an intuitive way. However, such systems are difficult to train, as it is expensive to ask annotators to produce code directly.

In this talk, I will present recent advances in learning semantic parsing systems via neural networks. Specifically, I will talk about how to cast the learning problem as a search problem, and how to learn a model to explore search space efficiently by using implicit supervision signals. Our algorithm uses a dynamic neural network design to compose a specific neural network architecture for each example. The proposed framework shows strong performance on two applications: knowledge base question answering and sequential question answering on semi-structured tables.



Ming-Wei Chang is a researcher at Microsoft Research, Redmond. He is
interested in making computers perform complicated tasks by learning from various types of supervision signals, which ranges from explicit supervision signals such as complete annotations to implicit supervision signals such as natural language responses. During his Ph.D., he proposed several algorithms to reduce the cost of supervising structured learning models. He and his colleagues also first proposed to reduce the cost of training semantic parsers by learning from the world’s responses, which is now a widely-used training protocol. Recently, he focuses on building models to ground entities and answer natural language questions and received an Outstanding Paper Award at ACL 2015 for his work on combining entity linking and knowledge base question answering models.

最後修改時間:2017-06-23 PM 1:35

cron web_use_log