SLDP: Sequence learning dependency parsing model using long short-term memory

Abstract

Recent work on neural network models shows success in dependency parsing. In this paper, we present a sequence learning dependency parsing (SLDP) model using long short-term memory for shift-reduce parser. A feed-forward neural network is used to build greedy model from rich local features. With the features extracted by the local model, we further train a long short-term memory (LSTM) model optimized for global parsing sequences. Our model has the capability of learning not only atomic feature combinations automatically but also the long distance dependent information for dependency parsing. Experiments on English Penn Treebank show that our SLDP model significantly outperforms the baseline, achieving 90.7% unlabeled attachment score and 89.0% labeled attachment score.