Text Classification Study Based on Graph Convolutional Neural Networks

Because the research of machine learning and deep learning so far has presented good results in the field of text classification, but it ignores the influence of words and words, and the relationship to a certain extent between words and documents on text classification in an overall perspective. Existing text classification researches are usually based on graph convolutional neural networks, in this paper, by using Mish() activation function to deal with the problem of experimental data in the hard zero boundary and by modifying the momentum of one-step moving average in the optimizer, a certain degree of improvement is proposed to the short-term memory problem after the optimization of experimental parameters, which is called LTM-TEXT-GCN. By comparison with Text-GCN, the experiment results demonstrate the effect of LTM-TEXT-GCN, and the improvement was 0.14%, 0.64%, 0.8% and 0.13% on the four data sets, respectively.