基于序列卷积网络的Twitter情感分类

为解决传统用于文本分类等时序性问题的循环神经网络无法留住长远记忆及模型框架复杂的问题,提出一种基于序列卷积神经网络的分类模型。利用卷积的思想处理时序性问题,将因果卷积和扩张卷积结合作为卷积层来保证网络具有足够大的感受野,应用残差模块和批处理加深神经网络并消除层数增加误差增大和模型难训练的问题,用卷积层代替全连接层以改善网络特征选取的局限性。实验结果表明,序列卷积分类模型用于Twitter情感分类任务中可获得更优的分类效果,验证了卷积网络的思想可以处理时序性问题。

Twitter sentiment classification based on sequential convolution network

To solve the problems that the recurrent neural network used for text classification and other timing problems cannot retain long-term memory and the design of model framework is complex,a classification model based on sequential convolution neural network was proposed.The scheduling problem was coped with convolutions,causal convolution and expansion convolution were combined as a convolution layer to ensure the network gaining big enough receptive field,application of residual module and batch process deepened the neural network and the problems that errors increase with the layers increase and that it is hard to train the model were solved.Convolution layer was used instead of full connection layer to improve the network feature selection.Experimental results show that the sequential convolution classification model can achieve better classification effects in Twitter sentiment classification tasks,which verifies that the convolution network idea can be used to deal with sequential problems.