Chinese Argument Identification Based on Bert

Abstract

Traditional semantic role labeling is mostly based on the results of syntactic analysis. On the basis of syntactic analysis, argument identification and argument classification are carried out in two steps. Due to the problem of error cascade, the effect of argument identification directly determines the quality of semantic role labeling. However, converting the semantic role labeling into a sequence labeling task tries to ignore syntactic information, which increases the difficulty of the model and relies on limited labeling data. Therefore, this paper focuses on improving the accuracy of Chinese argument identification, that is, identify all candidate arguments given a predicate. Specifically, based on the results of Chinese chunk dependency parsing, an argument identification model is built based on the pretrained language model BERT (bidirectional encoder representations from transformers). The AUC of the model reaches 97.18% and the accuracy reaches 98.10%, which provides a reliable data for subsequent argument classification task. At the same time, large-scale pretrained language model can overcome the problem of sequence dependency, and perform well in argument identification of complex sentences.