A Deep Bidirectional Highway Long Short-Term Memory Network Approach to Chinese Semantic Role Labeling

Abstract

Existing approaches to Chinese semantic role labeling (SRL) mainly adopt deep long short-term memory (LSTM) neural networks to address the long-term dependencies problem. However, deep LSTM networks cannot address the vanishing gradient problem properly. In addition, the complexity of the Chinese language, as a hieroglyphic language, decreases the performance of traditional SRL approaches to Chinese SRL. To address these problems, this paper proposes a new approach with a deep bidirectional highway LSTM network. The performance of the proposed approach is further improved by introducing the conditional random fields (CRFs) constraints and part-of-speech (POS) feature since POS tags are the classes of formal equivalents of words in linguistics. The experimental results on the commonly used Chinese Proposition Bank dataset show that the proposed approach outperforms existing approaches. With an easily acquired and reliable POS feature for practical applications, the proposed approach substantially improves Chinese SRL.