A Shift-Reduce Dependency Parser Based on Reinforcement Learning

Abstract

Dependency parsing is a task of extracting the relationships between words in a sentence. Researchers have achieved great success in transition-based and graph-based methods recently. However, there are problems of error propagation and high time complexity. This paper proposes dependency parser based on reinforcement learning to improve transition-based parser. We regard the actions in transition-based method as RL agent’s actions. Meanwhile, we introduce a BACK action, so that the agent can get back to the previous state after entering the error state. We use the Universal Dependencies dataset to make experiments on different language. Experiments show that this method effectively solves the problem of error propagation of transition-based method and improve the accuracy significantly.