Research On Multi-label Text Classification Method Based On Deep Learnin

Most objects in the real world have multiple semantics,leading to the inapplicability of the traditional single-label classification method.However,the purpose of multi-label classification method is to obtain a model that can assign associated label sets to unknown examples through learning,so it can model multi-semantic objects well.With the improvement of computing capacity,tabbed classification method based on the technology of deep learning more and more widely applied to various fields,which is based on convolutional neural networks,recurrent neural network and attention mechanism has achieved impressive performance,but these methods still exist shortcomings in feature extraction proportion or excavation section.Therefore,this thesis proposes two improved multi-label text classification methods based on deep learning and a fusion model multi-label classification method based on ensemble learning technology.Specific work includes:1.Aiming at the deficiency that recurrent neural network methods using attention mechanism only consider the contribution of different parts of text to different labels,a multi-label classification method based on improved recurrent neural network model DI-RNN was proposed.In this method,the contribution of different words is considered from the direction of text to label,and the weight of label to text is calculated from the direction of label to text,and the two are combined to make use of the relationship between text and label in both positive and negative directions.In addition,GRU network is used to replace LSTM network in order to make the method have higher training efficiency on large data sets.Experimental results on four benchmark data sets and five representative methods show that the overall performance of the proposed method is better than that of other representative multi-label classification methods based on recurrent neural networks.2.Aiming at the problem that the advanced multi-label classification methods based on convolutional neural network are redundant in the convolution part and easy to lose information in the pooling part,a multi-label text classification method based on the improved convolutional neural model AE-CNN was proposed.This method proposed the pooling method of correlation mean,which not only retained the location information of feature,but also took into account the internal relationship between feature values in different segments and reduced feature losses.In addition,this method integrates the advantages of other convolution models to further improve the classification effect.Finally,in the experiment of 5 benchmark data sets,the classification method has significantly improved compared with the comparison method in most indicators.3.In order to improve the accuracy of multi-label classification method,a multi-label classification method based on ensemble learning is designed to integrate multiple independent models.This method combines the complementary information in different independent models by integrating model output,label voting,training subsequent models or synchronous dependence,so as to obtain a fusion model with better performance.In the comparison of the four benchmark data sets,the fused model method performed significantly better than the independent model method,achieving a 3.8% improvement in Macro-averaging F1 indicators.