Research On Multi-Label Text Classification Based On Deep Learning

As a basic task in the field of natural language processing,multi-label text classification is widely used in sentiment analysis,question answering systems and recommendation systems.This thesis mainly studies multi-label text classification based on deep learning.Through in-depth analysis of the difficulties of multi-label text classification and developed by the current research status at home and abroad,two multilabel text classification models are proposed in this thesis:1.The first model in this thesis proposes a method of fusing label attention mechanism and self-attention mechanism to represent text features.At the same time,the Correlation network is added to the model is prediction layer to obtain the correlation between lables.Label information plays an important role in text classification.If the model can know the label is information in the feature extraction stage,then it can dig out the key information about the lable from the text based on the information provided by the label,so modle can obtain a more accurate feature representation of the label.At the same time,considering about some lables can be classified only by mining the local features of the text,while some lables need to be classified by mining the global features of the text.Therefore,this thesis also uses a self-attention mechanism to extract text features.Then according to the characteristics of the two feature representations,this model extracts the important information of the two representations to obtain the final text feature representation.Finally,this model adds the correlation network to the prediction layer,through which a more accurate label prediction vector with label correlation can be obtained.The experimental results show that the model achieves better results than the benchmark model.2.The second model in this thesis uses Seq2 Seq to do multi-label text classification.The model is composed of an encoder and a decoder.The encoder encodes the word vector to the hidden vector and the decoder generates the label sequence in turn.When model predicts lables,this model focuses on different parts of the text through the attention mechanism,and obtains the text feature representation of the labels to be predicted.Considering the importance of fully understanding text features for classification,this model proposes to fuse the text feature representation based on the attention with the pre-trained text vector to obtain a more comprehensive and accurate feature representation.The fused feature representation vector will be used for decoder.At the same time,this model uses the Mogrifier LSTM as decoder.In order to obtain the correlation between labels,at the time of model decoding,model will predict the current label based on the previously predicted label.Of course,this approach may cause the problem of exposure bias.If the previous label prediction is wrong,it may reduce the accuracy of the subsequent label prediction.Therefore,this model adopts a global label embedding method to alleviate this problem.The experimental results show that the model achieves better results than the benchmark model.