基于循环结构的卷积神经网络文本分类方法

Text classification method based on the cycle structured convolutional neural network

现有卷积神经网络在文本分类性能上受到词向量窗口长度的影响,在研究卷积神经网络分类方法的基础上,提出一种基于循环结构的神经网络文本分类方法,该方法对文本进行单次正向及反向扫描,能够在学习单词表示时尽可能地捕获上下文信息,整体算法时间复杂度为O(n),是线性复杂度;该方法构建文本语义模型可以捕获长距离的依赖关系,使得词向量窗口长度对文本分类性能没有影响,对上下文更有效地建模。实验结果表明,该方法构建文本语义模型的准确率达到96.86%,召回率达到96.15%,F1值达到96.5%,性能优于传统文本分类算法和卷积神经网络方法。

The existing convolutional neural network is influenced by the length of the word vector window in the text classification performance.On the basis of studying the convolutional neural network classification method,a text classification method based on the cycle structured convolutional neural network is proposed in this paper.The method only needs a single forward and reverse scan of the text to get as much as possible context representation.In this paper,the time complexity of the whole algorithm is O(n),which is linear complexity.In addition,the method can capture the long distance dependency by constructing the text semantic model.The word vector window length has no effect on the text classification performance,which can get more efficient modeling of the context.The experimental results show that the accuracy rate of the text model is 96.86%,the recall rate is 96.15%,the F1 value is 96.5%,and the performance is superior to the traditional text classification algorithm and the convolution neural network method.