基于非迭代训练层次循环神经网络的快速文本分类算法

FAST DOCUMENT CLASSIFICATION ALGORITHM BASED ON NON-ITERATIVE TRAINING HIERARCHICAL RECURRENT NEURAL NETWORK

基于神经网络的文本分类算法需要较长的训练时间,难以满足在线文本分类的需求。针对这种情况,提出基于非迭代训练层次循环神经网络的快速文本分类算法。为循环神经网络设计了对抗训练模型,缓解层次注意力网络的过拟合问题。给出一种循环神经网络的非迭代训练算法,对激活函数进行线性逼近,快速地学习网络连接的权重。实验结果表明,在英文和中文文本的情况下,采用该算法均获得了理想的分类准确率,并且大幅度地减少了训练时间。

Neural network-based document classification algorithms need a lot of training time,and they are difficult to meet the demand of online documents classification problems.In view of this,we propose a fast document classification algorithm based on the non-iterative training hierarchical recurrent neural network.We designed a adversarial training model for the recurrent neural network,and it reduced the overfitting problem of the hierarchical attention network.What is more,we presented a non-iterative training algorithm for the recurrent neural network.It linearly approximated the active functions,and learned the weights of network connections fast.The experimental results show that in both English and Chinese documents conditions,using our algorithm can obtain an ideal classification accuracy,and the training time reduces observably.