关键词: 刑期预测, 知识图谱, 大语言模型, 图模融合, 量刑规范化

Abstract: Sentencing prediction stands as a core task for legal artificial intelligence empowering criminal justice,playing a vital role in overcoming sentencing bias,enhancing judicial efficiency,and safeguarding fairness and justice.Addressing the bottleneck issues of low prediction accuracy and insufficient interpretability inherent in traditional machine learning models,this paper proposes an explainable sentencing prediction method driven by sentencing rule knowledge graph.The method innovatively designs a knowledge graph and large language model integration architecture.The technical roadmap is as follows.Firstly,a structured sentencing rule knowledge graph is constructed top-down using the BERT-BiLSTM-CRF model.Subsequently,Chain-of-Thought reasoning for sentencing is distilled from the Sentencing Guidelines,and structured prompting based on the graph’s data is employed to conduct supervised instruction fine-tuning on large language models(LLaMA-3-8B-Chinese-Chat,Qwen-2-7B,Baichuan2-7B-Chat,GLM-4-9B-Chat),guiding them to learn standardized sentencing reasoning logic.Finally,during the prediction phase,retrieval-augmented generation is implemented on the fine-tuned model via the graph’s entity recognition and retrieval mechanism,outputting sentencing predictions alongside explainable step-by-step analyses consistent with sentencing rules.Expe-rimental results demonstrate that:1)BERT-BiLSTM-CRF model achieves an F1 score of 0.953 8 on the entity-relation extraction task,outperforming conventional models;2)GLM-4-9B-Chat model achieves the best overall performance in both test-set generation quality and downstream tasks;3)The final sentencing prediction model achieves an F1 score of 0.627 6,significantly outperforming baseline models such as MTL-Fusion,Lawformer,and BERT.Moreover,generating explanatory text following the standardized logic of “determining the sentencing starting point-baseline sentence-adjusting the baseline sentence-declared sentence” significantly enhances the interpretability and user acceptance of results;4)Ablation studies and human evaluations jointly demonstrate the model’s significant superiority over baselines in sentencing accuracy,precision of legal provisions citation,logical cohe-rence and fluency of reasoning,as well as compliance with standardized sentencing steps.This research establishes a novel paradigm integrating knowledge-driven and data-driven approaches for legal AI.

Key words: Sentencing prediction, Knowledge graph, Large language model, KG-LLM integration, Sentencing standardization