Abstract: With the increasing research and application of deep learning, researchers have proposed a method by applying bidirectional long shortterm memory (BiLSTM) to semantic role labeling. Among them, many methods rely on a local window in word representation. In this case, the word representation may depend on the joint effect of some fixed word embeddings and it may preserve useless information in context. Aiming at this problem, we make improvements based on BiLSTM and introduce a Gated mechanism (GM) to filter information that will be fed to the next layer. In order to gain much more semantic information, we extend the depth of the BiLSTM. Due to the strong dependencies across output labels, we model tagging decisions using a conditional random fields (CRF) and transition matrices. Experiments on CPB benchmark dataset show that the F1 score of the Chinese semantic role labeling system finally improves 1.71%.

Key words:  Chinese semantic role labeling, Gated mechanism, BiLSTMCRF, transition matrix