Chinese Semantic Role Labeling Based on BILSTM-CRF Extended Model

Abstract

Semantic role labeling (SRL) is a technique to analyze the structure of predicates and thesis elements in a sentence as a unit. It plays an important role in Chinese information recognition processing. Among the models of SRL studied in recent years, most of them are based on bidirectional long and short term memory loop network and conditional random field. In this paper, we first narrate the SRL model based on BILSTM-CRF, based on which the second model narrates the SRL model integrating Bert and BILSTM-CRF models due to the ability of pre-training and fine-tuning of Bert model. However, since the word vectors in Chinese text are obtained based on word stitching in the context window, making the words between them influence each other, the word vectors depend on this joint relationship. Therefore, for this, Gate filtering mechanism is integrated to adjust it, and in the third model, Gate mechanism is added to filter and denoise the word vectors based on BILSTM-CRF to further improve the recognition ability of SRL.