Constructing Chinese Historical Literature Knowledge Graph Based on BERT

Qingyan Guo, Yang Sun, Guanzhong Liu, Zijun Wang, Zijing Ji, Yuxin Shen, Xin Wang

Abstract: Knowledge graph construction (KGC) aims to organize knowledge into a semantic network which can reveal relations between entities. Its basis is named entity recognition (NER) and relation extraction (RE) tasks. In recent years, KGC methods for Chinese have made great progress. However, most existing methods concentrate on modern Chinese and ignore the classical Chinese due to its complexity, making research in this field relatively lacking. In this paper, we construct a high-quality classical Chinese labeled dataset for NER and RE tasks. More specifically, we conduct a series of experiments to select an optimal NER model to strengthen the whole pipeline model for NER and RE tasks, augmenting our dataset iteratively and automatically. Additionally, we propose an improved RE model to better combine semantic entity information extracted by the NER model. Moreover, we construct a knowledge graph (KG) based on Chinese historical literature and design a visualization system with intuitive display and query functions.