L2-SVM下的短文本情感分类动态CNN模型

摘要

为了解决情感分类文本稀疏、传统方法过分依赖情感词典和人工设定特征工程等问题,提出一种基于L2-SVM和动态卷积神经网络的LDCNN模型.该模型采用不同于经典CNN模型的L2-SVM目标函数,解决了参数优化过程梯度弥散现象.通过真实网络评论数据集与经典方法的定量对比,实现了LDCNN模型准确率的大幅提升,并通过调整惩罚系数获得了最佳模型性能.%An LDCNN model based on L2-SVM and dynamic convolution neural network is proposed to solve the problem of sparse emotion classification text,over-depending on emotional dictionary in traditional method and human setting feature engineering.The model adopted L2-SVM objective function rather than the classical CNN model to solve the gradient dispersion phenomenon in parameter optimization process.After comparing the real network review data set and the classical method quantitatively,there is a significant improvement in the accuracy of LDC model and the best model performance is obtained by adjusting the penalty coefficient.