融合双向门控循环单元和卷积神经网络的歧视性言论鉴别

Discriminatory Comment Identification Based on Bidirectional Gated Recurrent Unit and Convolutional Neural Network

针对歧视性言论的鉴别,提出了一种融合双向门控循环单元(BIGRU,bidirectional gated recurrent unit)和多元卷积神经网络(MCNN,multi-convolution neural network)的BGM-CNN模型。模型先采用BIGRU结构进行时序特征提取,再经过一维多元卷积神经网络进行降维池化,最后结合多组特征输出进行分类。实验结果表明,BGM-CNN模型比现有的单一模型和CNN-LSTM(long short-term memory)等模型分类效果更好,该模型在五分类验证数据集上分类的F1值为0.6733,在两个歧视性言论二分类数据集上的F1值分别为0.8373和0.8156。

For the identification of discriminatory comment,a BGM-CNN model combining bidirectional gated recurrent unit(BIGRU)and multi-convolution neural network(MCNN)is proposed.The model first uses BIGRU structure to extract time series features.Then one-dimensional multi-convolution neural network is used for dimension reduction pooling,and finally classification is made according to the combination of multiple sets of feature outputs.The experimental results show that the BGM-CNN model has better classification performance than the existing single model and CNN-LSTM(long short-term memory)model.The classification F1-score value of the model on the five-category verification data set is 0.6733.For the other two discriminatory comment two-category data sets,F1-score values are as high as 0.8373 and 0.8156,respectively.