【摘要】 针对现有模型未充分利用社交媒体中文本历史动态信息进行自杀风险检测的问题，引入CapsNet模型。在CapsNet模型中，层与层之间传递的是有向神经元组，能够更好地感知长文本中的空间信息，发现社交媒体用户的情感趋势，为自杀风险检测提供依据。文中对CapsNet模型进行改进，首先改变尺度空间，增加网络宽度，充分提取隐藏在句子中的特征信息；其次，使用指数函数对Squash函数进行优化，放大胶囊输出，充分利用胶囊提取用户历史动态中的特征信息；最后，在动态路由中采用优化算法对耦合系数进行初始化，去除噪声胶囊的干扰。使用预训练的SBERT模型对社交媒体文本数据进行特征提取，得到改进CapsNet文本自杀风险检测模型二分类的准确率达到95.93%,F1分数达到95.86%，优于自杀风险检测的其他模型。更多还原

【Abstract】 In allusion to the problem that the existing models do not make full use of the historical dynamic information of text in social media for suicide risk detection, the CapsNet model is introduced. In the CapsNet model, groups of directed neurons are transmitted between layers, which can better perceive spatial information in long texts, find emotional trends of social media users, and provide a basis for suicide risk detection. The CapsNet model is improved. The scale space is changed and the network width is increased to fully extract the feature information hidden in the sentence. The exponential function is used to optimize the Squash function, so as to enlarge the capsule output, and make full use of the capsule to extract the feature information in user′s historical dynamics. In dynamic routing, an optimization algorithm is used to initialize the coupling coefficient to remove the interference of noisy capsules. The pre-trained SBERT model is used to extract features of social media text data. The binary classification accuracy of the improved CapsNet text suicide risk detection model can reach 95.93%, and the F1 score can reach 95.86%, which is better than other models of suicide risk detection.更多还原