An improved model of MST for Chinese dependency parsing

In this paper, a Chinese dependency parsing method is proposed based on improved Maximum Spanning Tree (MST) Parser. Within this method, dependency direction discrimination model and head POS recognition model are used to modify the weights of directed edges in the MST model, and then the Eisner algorithm is used to search and generate the dependency trees. In this paper, the problems of dependency direction discrimination and head POS recognition are converted into sequence labeling; and the modeling is done by condition random fields. We tested our method on CoNLL 2009 Share Task, and the Unlabeled Attachment Score reached 86.27%.