Research And Application Of Multi-label Text Classification Algorithm

In recent years,with the development of artificial intelligence technology,the deep learning theory based on artificial neural network has made revolutionary progress in the field of natural language processing,and has already derived many applications and research directions.Multi-label text classification(MLC)tasks,as one of the sub-tasks of-natural language processing,have broad practical application prospects,such as text classification,label recommendation,information retrieval,and so on.In the research and application process,multi-label classification tasks have many commonalities and fundamental differences compared with traditional multi-classification tasks.Based on the analysis and research of previous research work,this study summarizes the research results of predecessors based on the relationship between text information,leading labels information and predicted label information.Based on the above theory,Three problems caused by the disappearance of the gradient in the seq2seq implementation with RNN or LSTM as the basic encoder and decoder are discussed.The original text information is lost,the leading label information is lost,and the decoding error is accumulated.Based on the analysis of related research work,we propose an improved multi-label text classification model with end-to-end structure to alleviate the phenomenon of false prediction,label repetition and error accumulation.The model uses multi-step and multi-classification task to complete multi-label prediction.It uses the leading label and the original text as input,and the next to-be-predicted label as output.The co-attention mechanism is operated between the original text information and the leading label information.The attention of leading labels to the orisginal text information help to alleviate the loss of the original text information,and optimizes the training difficulty of the cyclic neural network encoder due to the gradient disappearance problem.The attention of the original text information to leading labels help to filter out the error accumulation caused by the error prediction and correct the error memory of the external memory unit.The model combines the original text information with the leading label information in a differential and concatenation manner,which highlights the auxiliary role of the model structure for feature extraction.At the same time,the model replaces the model structure such as LSTM with multi-layer fully-connected classifier for label prediction,thus avoiding the influence of large feature dimension on LSTM performance.Through experimental verification,the performance of our model on the multi-label text classification task shows the current optimal level,which fully proves the superiority of our theoretical method and model.