Multi-Label Modeling By Sequence Generation For Text Classification With Dependence And Sparsity

In the Internet,the accuracy of text labels affects the speed of retrieval,and the accuracy of labels depends on the performance of the classification algorithm.In the actual scene,an article belongs to different labels,so the accuracy of multi-label text classification determines the experience of Internet users.However,the existing algorithms have limitations in dealing with multi-label classification problems.Therefore,designing efficient and accurate multi-label classification algorithms has great practical significance.Compared with the traditional single-label text classification task,there are two difficulties in the classification of multi-label texts: one is that the text features are sparse,it will reduce the accuracy of the existing classification model;the other is that the label is dependence.Sexuality,previous algorithms tend to ignore these features,resulting in weak generalization of the model.To this end,research from the above two aspects:For the sparsity of text features,the word2 vec model is used for modeling.Through unsupervised learning,the context is used to predict the current words,so that the hidden layer neurons can learn the semantic information of each word,thereby solving the sparse problem of the original text features.Due to the computational complexity of the original model,the algorithm uses a hierarchical softmax to accelerate.From the theoretical theory of the model,the problem is transformed into a stochastic gradient descent method by the maximum likelihood estimation method,and each parameter gradient is given.Finally,through the comparison experiment,the proposed model is superior to other algorithms in the main evaluation indicators,thus verifying the effectiveness of the algorithm.For the dependency of the class label,the classic machine translation model seq2 seq is used for modeling.The Seq2 seq decoding process uses a long-short-term memory network(LSTM)to generate a label sequence,and generates a next label tag based on the predicted label,effectively solving the dependency problem between the labels.The following three improvements are made to the original seq2 seq model: First,the attention mechanism is adopted,which effectively considers that different feature words in the text have different importance to different class predictions.Second,a tag vector is used to mark the decoding.The class label that has appeared in the stage avoids repeated prediction of the label.The third is to use the convolutional neural network(CNN)to globally encode the entire text into the decoding of seq2 seq.In theory,the gradient of the parameters in each sub-module is solved for the entire model sub-module.In the specific numerical verification analysis,the learning optimization adaptive Adam optimization algorithm is used for iterative solution.By comparing the experimental settings and using multiple standard datasets for simulation experiments,the proposed CNN-seq2 seq multi-label classification model is superior to other traditional algorithms in the main evaluation indicators.And through the visual analysis of the prediction results,the proposed CNN-seq2 seq model can predict the correlation label,which further proves that the model can solve the label-dependent problem.