LD-Parser: Leaf Detection Based Dependency Parsing Using BiLSTM and Attention Mechanism

Abstract

Dependency parsing is one of the basic research of natural language processing. In recent years, transition-based and graph-based methods have been used widely, but there are still some problems such as feature limitation and high time complexity. In this paper, we proposed a new method named Leaf Detection based Dependency Parsing (LD-Parser), which is a bottom-up framework to detect leaf nodes of the dependency parsing tree. We use LSTM to construct a classifier to generate labels for each word and then remove the leaf nodes that are adjacent to their corresponding parents. Besides, an attention mechanism is introduced to sum the children of nodes as an extra feature according to the attention weight. We make experiments on Universal Dependencies in several languages. Experiments show that the LD-Parser with Attention performs better than transition-based and graph-based methods in dependency parsing tasks for short sentences.