Abstract: Auxiliary secret classification is a special text classification task that divides undecided encrypted text into different levels of confidentiality.. In order to solve the problems of the traditional method, such as weak feature representation and extraction ability and low interpretability of the classification process, keypoints feature representation method based on graph structure was proposed. On that basis, an auxiliary secret classification model based on keypoints extraction was further proposed, so as to enhance the ability of secret point features in describing the confidential matters, thus the performance of the auxiliary classification model is enhanced. Specifically, this paper deeply analyze the characteristics of classification rules, constructs the keypoints template with reference to text representation method of the graphic structure, extracts the keypoints and calculates the confidence level of the keypoints of the text to be classified, and obtains the secret level prediction results and the classification basis items through the filtered effective keypoints. The experimental result on the ACD indicates that the accuracy and recall rate of this model are 10% and 7% higher than those of BERT and TextCNN, which verifies the effectiveness of keypoints feature representation method based on the graph structure.

Key words: graph structure keypoints, confidence of the keypoints, auxiliary classification, classification rules, extracting keypoints

摘要： 辅助定密是将待定密文本按照密级进行划分的特殊文本分类任务.针对传统辅助定密方法存在的特征表示和提取能力不强、定密过程可解释性弱等问题，提出基于图结构的密点特征表示方法，并进一步提出基于密点抽取的辅助定密模型，以增强密点特征描述涉密事项的能力，提升辅助定密模型性能.深入分析定密规则特征，借鉴图结构文本表示方法构建密点模板，对待定密文本进行密点抽取和密点置信度计算，通过筛选出的有效密点得出密级预测结果和定密依据条目.在针对辅助定密任务构建的数据集(ACD)上的实验结果表明，基于图结构密点抽取的辅助定密模型在准确率和召回率等指标上，相较于BERT，TextCNN等模型分别提升10%和7%以上，验证了图结构密点特征表示方法的有效性.

关键词: 图结构密点, 密点置信度, 辅助定密, 定密规则, 密点抽取