Legal texts usually contain complex entity relationships, and traditional manual analysis methods are not only inefficient but also easily affected by human factors. In this study, a new entity relationship extraction model for legal texts based on graph convolutional networks and BERT, named ON-BERT, is proposed. The model captures hierarchical semantic features in the text through the hierarchical structure parsing module and extracts global semantic information by combining with the BERT pre-trained language model. The experiments are conducted on 15,000 criminal judgments published in China Judgment Website, and 12,163 valid case texts are obtained after data processing. The experimental results show that the ON-BERT model outperforms the traditional model in terms of precision, recall and F1 value. In the test, the F1 value of ON-BERT is 83.56%, which is improved by 3.92% compared to the BERT model, and in terms of accuracy, ON-BERT also significantly outperforms the other models, reaching 82.55%. In addition, ON-BERT also shows significant improvement in training efficiency and inference speed, and its training time is shortened by about 4 times compared to the baseline model. The effectiveness and efficiency of this model provides a new technical path for legal text analysis.
