Simulating judicial trial logic: Dual residual cross-attention learning for predicting legal judgment in long documents

Legal Judgment Prediction (LJP) plays a vital role in judicial assistance systems, aiming to predict judgment outcomes automatically from the fact descriptions in legal cases. A key challenge in LJP lies in effectively capturing decisive information that influences legal judgments — such as criminal events, behaviors, and consequences — especially from lengthy legal documents. Existing approaches, which incorporate domain-specific knowledge, have attempted to improve the ability to capture decisive information but often fail to adequately address the complexities of longer texts and may introduce noise, leading to incorrect predictions. To overcome these limitations, we propose a novel method, JuriSim, for predicting Chinese criminal legal judgments in long documents by incorporating the knowledge of judicial trial logic. Specifically, JuriSim extracts legal events and generates rationales based on fact descriptions to capture the decisive information that influences judgment outcomes. Then, a dual residual cross-attention mechanism is introduced to interactively process facts, events, and rationales for predicting relevant legal statutes, charges, and term of penalties. This mechanism allows the model to reduce the loss of important information and the retention of incorrect information during the aggregation process. Furthermore, we present a constrained cross-entropy loss, utilizing the topological relationship between charges, terms, and applicable law statutes. Experiments conducted on the publicly available CAIL-Long criminal dataset demonstrate the efficiency of the JuriSim framework in predicting legal judgments, especially for cases involving long documents. JuriSim_Lawformer shows a relative improvement of 3.45% in Macro-F1 for charge prediction and 3.05% for term of penalty prediction, compared to Lawformer.