基于改进CNN的海军军事文本分类模型

Naval Text Classification Model Based on Improved CNN

针对传统的文本分类方法在海军军事文本分类上准确度不高的问题,根据海军军事文本中重点信息的位置分布规律,改进了传统的一维卷积神经网络,并进一步设计海军军事文本分类模型。在一维卷积方面,提出变步长卷积方法,文本首尾位置采用低步长、中间位置采用高步长挖掘文本特征,提高文本首尾位置的重点特征的挖掘能力;在一维池化方面,提出带权池化方法,将文本位置信息转化为权重值参与池化运算,体现文本位置信息的重要程度。实验结果表明,与传统的支持向量机、K近邻算法、一维卷积神经网络以及长短期记忆网络模型相比,该文本分类模型的准确率、召回率、F1值均有所提高。

The accuracy of the traditional text classification method is not high enough for naval text classification task.According to the regularity of location distribution of key information in the naval military text,the traditional one-dimensional Convolutional Neural Network(CNN) is improved,and the naval text classification model is designed.In the aspect of one-dimensional convolution,a variable-step convolution method is proposed.The text features are mined by using low step-size at the beginning and end of the text,and high step-size in the middle,so as to improve the mining ability of key features at the beginning and end of the text.In the aspect of one-dimensional pooling,a weighted pooling method is proposed to convert text location information into weighted values to participate in the pooling operation,reflecting the importance of text location information.Experimental result shows that,compared with the traditional support vector machine,K-neighbour algorithm,the traditional one-dimensional CNN and the long/short-term memory network,our text classification model has higher accuracy rate,recall rate and F1 value.