Abstract:

In traditional convolution neural network text classification model, too many parameters in the fully connected layer can cause the over-fitting problem. Therefore, this paper introduces the idea of fully convolution in image segmentation into character-level text classification task for the first time, which not only avoids the over-fitting problem, but also reduces the number of parameters by replacing fully connected layers with convolutional layers, thus speeding up the convergence speed of the model. In addition, the processing of words and phrases in the text classification is insufficient to obtain text information. This paper uses a character-level fully convolutional neural network to obtain text information sufficiently. It adds the local response normalization layer to improve the performance of the model. The experimental part uses multiple indicators to evaluate the model, which fully proves the validation of the model. Compared with other models, the proposed model performs better in two classification and multi classification tasks.

Key words: text classification, fully convolutional neural network, character level, Local Response Normalization（LRN）, feature extraction

摘要：

传统卷积神经网络文本分类模型全连接层参数过多易引发过拟合问题，为此，将图像分割中的全卷积思想首次引入字符级文本分类任务中，不仅避免了过拟合问题，而且通过卷积层替换全连接层减少了参数数量，从而加快了模型收敛速度。文本分类问题中单词、短语等层面的处理方式存在获取文本信息不充分的问题。使用字符级全卷积神经网络进行文本分类，充分获取文本信息，并在卷积池化层后添加局部响应归一化层（LRN），提高了模型的总体性能。通过使用多指标在测试数据集中进行模型评估，充分验证了该模型的有效性，与其他模型相比，提出的模型在二分类与多分类任务中具有更好的分类性能。

关键词: 文本分类, 全卷积神经网络, 字符级, 局部响应归一化层（LRN）, 特征提取