一种结合压缩激发块和CNN的文本分类模型

Text Classification Model with Squeeze-and-excitation Block and CNN

针对单一卷积神经网络进行文本分类,容易出现忽视局部与整体之间关联性的问题,本文构建了一种基于压缩激发块的卷积神经网络文本分类模型,提高文本分类的精确度.主要工作分为三部分:1)使用字符级词向量作为卷积神经网络的输入;2)引入压缩-激发块学习使用全局信息,有选择地强调有用的特征,来增加提取特征的多样性,弥补单一卷积神经网络多样性的不足;3)使用多头注意力机制进行权重更新计算,突出类别向量的重要程度.实验结果显示,本文提出的文本分类模型,在THUCNews数据集和搜狐数据集上,比单一的字符级卷积神经网络模型精确度分别提高了2.29%、4.75%.

Aiming at the problem of ignoring the relevance between the part and the whole when text classification is carried out by single convolutional neural network,this paper constructs a text classification model of convolutional neural network based on extruded excitation blocks to improve the accuracy of text classification.The main work is divided into three parts:1)To use char-level word vector as the input of convolutional neural network;2)Introduce Squeeze-and-Excitation block to learn to use global information,the useful features are selectively emphasized to increase the diversity of feature extraction and make up for the lack of diversity of single convolutional neural network;3)To use Multi-Head Attention mechanism weight updating calculation,highlight the importance of class vector.The experimental results shows that the text classification model proposed in this paper improves the precision of the THUCNews data set and sohu data set by 2.29%and 4.75%,respectively,compared with the single character-level convolutional neural network model.