With the rapid development of information technology, artificial intelligence has entered every aspect of life, and the intelligence in the judicial field is an inevitable trend for future development. Therefore, this paper utilizes deep learning text classification technology to achieve crime prediction and legal recommendation tasks. The theoretical part elaborates on the structural principles and attention mechanism of the BiGRU model, and provides the detailed mathematical definition of the model in this paper. The experimental part is divided into four parts: data preprocessing, experimental model design, indicator evaluation, and result analysis. A BiGRU + Attention model based on BERT word embedding is proposed through the GRU and attention mechanism. Experiments are designed on the CAIL2018 small dataset and compared them with baseline methods. In the crime prediction task and legal recommendation task, its comprehensive Funion value (calculated from the F1 micro-mean and F1 macro-mean, with the formula Funion= (Fmi+Fma)/2) reached 85.05% and 80.14%, respectively. The results indicate that this method can effectively improve the semantic accuracy of criminal fact encoding. This research provides strong support for the intelligent development of the judicial field, improves the efficiency of judicial work, and helps judges make more accurate judgments and decisions.