Sequence-based High-order Dependency Parser For Non-projective Languages

As with the development of NLP(natural language process)tasks under computer science,getting a better understanding for syntax parsing has become a critical step in NLP systems.In recent years,the researchers have draw more attention to dependency parser,which is more concise than phrase structure parsers and is word-order free.Existing data-driven dependency parsers are divided into two classes,graph-based and transition based,while non-projective languages is poorly supported by current solutions.After exploring current parsing systems,this paper describes a novel high-order dependency parsing framework that targets non-projective languages.It imitated how a human parsed sentences in an intuitive way.At every step of the parse,it determined which word was the easiest to process among all the remaining words,identified its head word and then folded it under the head word.Gradually,the dependency tree is constructed.The established system consists of two parts: the sequence predictor and the head mapper,they cooperate with each other by the processing sequences.The sequence predictor generates an easy-first processing sequence for each sentence,which determines the parsing order.The head mapper is responsible for mapping current word to its best head according to a linear scorer for dependency arcs.This greedy framework achieved a competitive accuracy on WSJ evaluation set and got an end-to-end accuracy as 89.5%.Judging from the accuracies for non-projective arcs on 5treebanks,it showed additional advantage over graph-based and transition-based methods on the non-projective corpus.Further,this work is flexible enough to be augmented with other parsing techniques and the two key components can be optimised separately.