A simulated shallow dependency parser based on weighted hierarchical structure learning

Abstract

In the past years much research has been done on data-driven dependency parsing and performance has increased steadily. Dependency grammar has an important inherent characteristic, that is, the nodes closer to root usually make more contribution to audiences than the others. However, that is ignored in previous research in which every node in a dependency structure is considered to play the same role. In this paper a parser based on weighted hierarchical structure learning is proposed to simulate shallow dependency parsing, which has the preference for nodes closer to root during learning. The experimental results show that the accuracies of nodes closer to root are improved at the cost of a little decrease of accuracies of nodes far from root. © 2008 Springer-Verlag Berlin Heidelberg.