Research on Information Extraction in Judicial Field

摘要

In the era of information explosion, how to effectively identify the named entities and the relationships between them from the massive judicial data is a key step to realize automatic trial, and it is also the focus of this paper. The traditional way is to use the pipeline method to identify named entities first, and then extract relationships, which will lead to the loss of related information between tasks and ignore the overlapping relationships between entities. For this reason, this paper designs a joint extraction algorithm for entities and relations in the judicial field. It uses the BiLSTM network model based on the BERT pre-training language algorithm and combines it with the attention mechanism, realizes joint learning through parameter sharing, and makes full use of the connections between tasks to optimize the results. In order to verify the effectiveness of this method, this paper constructs a small judicial data set to verify its generalization ability. Experimental results show that the model proposed in this paper can extract entities and relationships without manually defining complex features, and the average F value can reach 0.7.