A Separate-and-Learn Approach to EM Learning of PCFGs

Wepropose a new approach to EM learning of PCFGs. We completely separate the process of EM learning from that of parsing, and for the former, we introduce a new EM algorithm called the graphical EM algorithm that runs on a new data structure called support graphs extracted from WFSTs (well formed substring tables) of various parsers. Learning experiments with PCFGs using two Japanese corpora indicate that our approach can signi cantly outperform the existing approaches using the Inside-Outside algorithm (Baker, 1979) and Stolcke's EM algorithm (Stolcke, 1995).