Relation Extraction Method Based on Deep Learning for Theft Police Records
Weicheng Huang, Xiangwu Ding · 2023
Theft police records contain information about the relations between different entities in the theft case, which can provide auxiliary support for the police to build a knowledge graph in the criminal justice field. However, theft police records are unstructured textual data, which makes it difficult to extract the relation between different entities directly. To improve the performance of relation extraction, we propose a novel joint entity-relation extraction model to extract entity-relation information from police records of theft cases. Considering the specificity of Chinese, we used a LERT pre-training model to generate word vectors with more semantic information from three linguistic features, then the generated word vectors were normalized using weight normalization. After that, the word vector is input to the head entity recognition layer to label all the head entities in the record using a binary entity labeling method based on Bi-SRU neural network. Finally, after merging the head entity information into the word vector, the same approach is used in the tail entity recognition layer to tag all the tail entities contained in the record, thus effectively extracting the relation between different entities in the police records of theft cases. Experiment results show that the model we proposed improves the recall by 4.39% and the F1Score by 1.55% compared to the baseline model.