Abstract: The causes of criminal psychology were analyzed to assess criminal risk and develop psychological rehabilitation programs. Limitations were exhibited by traditional models in comprehending criminal psychology texts, revealing correlations between prompt vectors and downstream tasks, and dynamically identifying the importance of key terms. To address these limitations, a CP-BERT-based attribution classification model for criminal psychology was proposed, enabling four-category classification of root causes. First, an 8.03-million-character criminal psychology corpus was constructed for domain-adaptive pre-training, and semantic modeling capability was enhanced； Second, a hybrid prompt method was employed in the prompt learning component, in which prompt templates were divided into prefix and suffix segments: highly relevant prior information was injected into the prefix during initialization, while the suffix was initialized using a normal distribution to supplement uncovered information, and optimal proportions were determined experimentally； Third, an improved TF-Chi association probability feature method was proposed, in which numerical features were dynamically mapped to information relevance probability values, and feature fusion was achieved by constructing a weight matrix. In addition, an annotated dataset consisting of 3,484 growth and criminal histories was constructed for model training and evaluation. Experimental results demonstrated that an accuracy of 91.25% and an F1 score of 91.56% were achieved, significantly outperforming multiple baseline models across all metrics, and effectiveness in criminal psychological attribution tasks was validated.

Key words: Criminal psychology, Domain-adaptive pretraining, Prompt learning, Feature fusion, Deep learning

摘要： 分析犯罪心理的成因对于评估犯罪风险及制定心理矫治方案至关重要。传统模型在理解犯罪心理文本、揭示提示向量与下游任务的关联性，以及动态识别特征词重要性等方面仍存在局限。为此，提出了一种基于CP-BERT（Criminal Psychology-Bidirectional Encoder Representations from Transformers）的犯罪心理归因分类模型，以实现对犯罪心理成因的四分类。首先构建803万字符的犯罪心理语料库，进行领域自适应预训练，增强其语义建模能力；其次，设计了混合提示学习方法（Hybrid Prompt），将提示模板划分为前缀与后缀两部分：前缀在初始化阶段注入与归因任务高度相关的先验信息，后缀则采用正态分布随机初始化，用以补充前缀未覆盖的信息，并通过实验确定最优比例；第三，提出改进的TF-Chi（Term Frequency-Chi Square）关联概率特征方法，将数值特征动态映射为信息关联性概率值，并通过构建权重矩阵实现特征融合。本文同时构建了一个包含3,484条成长史和犯罪史的标注样本数据集进行模型训练与评估。实验结果表明，该模型的准确率和F1分别达到91.25%与91.56%，在所有指标上均显著优于多个基线模型，验证了其在犯罪心理归因任务中的有效性。

关键词: 犯罪心理, 领域自适应预训练, 提示学习, 特征融合, 深度学习