关键词: 文本分类, 特征偏移, 卷积神经网络

Abstract: Aiming at the  problems that there were feature deviation for  the minority class in massive open online course (MOOC) reviews, we proposed a deep intelligent teaching evaluation method based on compensation for feature deviation. Firstly, this method used the Glove pre-training model to obtain the distributed word vectors of MOOC reviews. Secondly, the shallow convolutional neural networks were used to learn the semantics of teaching evaluation  through multiple convolution kernels. The number of different types of reviews was introduced to design influence  factors,  which was  normalized and applied to  the cross-entropy loss function. Finally,  the data set of undergraduate teaching reviews based on Coursera was compared with other loss functions on F₁,gmean,balance,gmeasure and other evaluation indicators. The experimental results show that the  loss function based on normalized  feature deviation compensation has  a performance improvement of up to 15.40% on gmeasure than the base loss function, and  the classification model using this loss function also shows strong stability.

Key words: text classification, feature deviation, convolutional neural network