Skip to Main Content
The high-order graph-based dependency parsing model achieves state-of-the-art accuracy by incorporating rich feature representations. However, its parsing efficiency and accuracy degrades dramatically when the input sentence gets longer. This paper presents a novel two-stage method to improve high-order graph-based parsing, which uses punctuation, such as commas and semicolons, to segment the input sentence into fragments, and then applies a two-level parsing. Experimental results on the Chinese data set of the CoNLL 2009 shared task show that our two-stage method significantly outperforms both the conventional one-stage method and previously-proposed three-stage method in terms of both parsing efficiency and accuracy.