Deep Variational Inference with Discrete Latent Structure for Graph-based Parsing
Graph-based Parsing is an interpretable and simple parsing technique used for predicting a semantic graph from a sentence, where labeled nodes and edges are predicted from their corresponding sentence tokens independently. However, the such correspondence might not be available due to the lack of annotation as in the Abstract Meaning Representation. We show how the lack of correspondence can be modeled as discrete latent variables and train graph-based parser end2end without domain knowledge.
Chunchuan Lyu (吕纯川) specialized in natural language processing, deep learning, and Bayesian reasoning. He spent his undergraduate at Xi'an Jiaotong-Liverpool University, studying machine learning. Then, he got interested in NLP, and got both his master and doctoral degree from the University of Edinburgh. Last year, he did a postdoc at Instituto Superior Tecnico at Lisboa, focusing on structured prediction. Chunchuan has published papers on ACL,EMNLP,ICDM,TMLR, etc.