关键词: 实体关系联合抽取, 全局指针网络, 提示学习, 预训练语言模型, 中文古籍

Abstract: The entity-relation joint extraction methods based on the "pre-training + fine-tuning" paradigm rely on large-scale annotated data. In the small sample scenario of Chinese ancient text where data annotation is difficult and costly, the fine-tuning efficiency is low and the extraction performance is poor; entity nesting and relation overlapping problems are common in Chinese ancient text, which limit the effect of entity-relation joint extraction; the pipeline extraction methods have error propagation problems, which affect the extraction accuracy. In response to the above problems, an entity-relation joint extraction method from Chinese ancient text based on learning and global pointer network was proposed. First, the learning method of span extraction reading comprehension was used to inject domain knowledge into the pre-trained language model (PLM) to unify the optimization goals of pre-training and fine-tuning, and encode the input sentences. Then the global pointer networks were used to predict and jointly decode the boundaries of the subject and object and the boundaries of the subject and object of different relationships, align them into entity-relation triples, and complete the construction of the PTBG ( Tuned BERT with Global Pointer) model. The model solved the problem of entity nesting and relation overlaping, and avoided the error propagation problem of pipeline decoding. Finally, on this basis, the impact of different templates on extraction performance was analyzed. Experiments were conducted on the "Historical Records" dataset. Compared with the Onerel model before and after injected domain knowledge, the F1 value increased by 4.71 and 1.91 percentage points. Experimental results show that the proposed method can better jointly extract entity-relation in ancient Chinese text, and provides new research ideas and methods for low-resource, small-sample deep learning scenarios.

Key words: entity-relation joint extraction, global pointer network, learning, Pre-trained Language Model（PLM）, Chinese ancient text