Abstract: This article constructs a judicial judgment prediction model that optimizes similar case retrieval and recommendation, improves the quality and efficiency of trials, and supports the design of an intelligent court system. It combines the characteristics of the legal field for knowledge representation and reasoning, incorporates specialized legal knowledge, and extracts objective factual information, geographical details, and the latent text structure from legal judgments. A mapping between legal topics and the latent text structure is established. Based on these features, a judicial decision prediction model is constructed to conduct similar case retrieval and recommendation, further supporting the functional design of an intelligent court system. Experimental results show that the judicial decision prediction model achieves an AUC value of 0.876 on the test set, a 9.9% improvement over the baseline model. The proposed model enhances the interpretability of judgments, and significant regional variations are observed among the "bride price refund" cases. The experimental results confirm the predictive accuracy and interpretability of the proposed legal judgment prediction model. This research provides a new tool for similar case retrieval and recommendation.