Research on Domain-Specific Knowledge Graph Based on the RoBERTa-wwm-ext Pretraining Model.

Abstract

The purpose of this study is to solve the effective way of domain-specific knowledge graph construction from information to knowledge. We propose the deep learning algorithm to extract entities and relationship from open-source intelligence by the RoBERTa-wwm-ext pretraining model and a knowledge fusion framework based on the longest common attribute entity alignment technology and bring in different text similarity algorithms and classification algorithms for verification. The experimental research showed that the named entity recognition model using the RoBERTa-wwm-ext pretrained model achieves the best results in terms of recall rate and F1 value, first, and the F value of RoBERTa-wwm-ext + BiLSTM + CRF reached up to 83.07%. Second, the RoBERTa-wwm-ext relationship extraction model has achieved the best results; compared with the relation extraction model based on recurrent neural network, it is improved by about 20%∼30%. Finally, the entity alignment algorithm based on the attribute similarity of the longest common subsequence has achieved the best results on the whole. The findings of this study provide an effective way to complete knowledge graph construction in domain-specific texts. The research serves as a first step for future research, for example, domain-specific intelligent Q&A.