The continuous improvement of judicial construction has led to the emergence of a large amount of judicial data on the Internet, and how to make full use of judicial data to promote judicial openness, fairness and efficiency has become an important issue in the construction of judicial informatization. In the article, the word vector generation technique is used to obtain the annotation sequence of legal text, and then the BiLSTM model is combined with the CRF model to realize the recognition of legal text entities, and the Adam algorithm is used to optimize the training of the model, so as to improve the recognition effect of the model on legal text entities. The GCN model in the graph representation learning algorithm is introduced, and the legal text entity recognition results are used as inputs for the construction of sequential and semantic relationships, and the GCN-BiLSTM model for legal text entity relationship extraction is constructed by combining the graph representation attention network and the BiLSTM model. Based on the self-constructed legal text dataset, the validation analysis of the above model is carried out through simulation experiments.The accuracy of the BILSTM-CRF model in legal text entity recognition is 85.67%, which is 7.35% higher than that of the single LSTM-CRF model. The GCN-BiLSTM model improves its accuracy by 2.14 percentage points compared with the CasRel model in extracting the entity relationships of legal texts with multi-entity overlapping. Combined with the legal text entity relationship extraction results, the knowledge map of legal cases can be constructed to provide accurate knowledge relationship support for sorting out the veins of legal cases.