A Relation Extraction Model Based on BERT Model in the Financial Regulation Field

Abstract

Relation extraction is a natural language processing (NLP) task to extract semantic relations between entities from unstructured texts, and is an important research content of constructing knowledge graph. Based on the documents issued by the China Banking and Insurance Regulatory Commission, we construct a knowledge graph in the field of financial regulation through relation extraction, which helps bank staff to realize intelligent query of regulatory documents and locate business-related regulatory requirements quickly and accurately. This paper proposes the Financial Regulation BERT (FR-BERT) model for relation extraction in the financial regulation field, aiming at the dataset's characteristics of rich semantic information of target entities, clear keywords and long texts. FR-BERT uses the BERT model to obtain sentence vector information containing each word vector, then obtains the target entity vectors and keyword vector by locating the sentence vector information, and obtains the text vector by sending the whole sentence vector information into Bi-directional Long Short-Term Memory (BiLSTM). Finally, the model uses the above vectors for classification information. FR-BERT integrates target entity vectors, keyword vector, and text vector output by BiLSTM to achieve relation extraction between entities. Compared with other relation extraction models selected in this paper, the experimental results show that FR-BERT performs better such as Macro-F1 on the relation extraction task in the financial regulation field, which verifies the effectiveness of the method.