Research On Multi-Label Text Classification Method Based On Deep Learning And Topic Models

With the continuous development of information technology,the amount of textual information has dramatically increased,and the content of text has become diversified.Traditional single-label text classification is no longer able to cope with the diversity of text content.In recent years,multi-label text classification has become one of the hot research topics in natural language processing tasks.Compared with traditional single-label text classification,multi-label text classification tasks pose the following challenges: Firstly,there is correlation between labels and words,and different labels focus on different words.Therefore,how to extract text features specific to certain labels is a major challenge in multi-label text classification tasks.On the other hand,labels are not independent of each other,and there exist dependency relationships such as label hierarchy and co-occurrence.The output space grows exponentially as the number of labels increases.Therefore,how to mine the dependency relationships between labels is another major challenge in multi-label text classification tasks.This dissertation aims to address the problems and challenges existing in multi-label text classification tasks by conducting the following main works:(1)To address the problem of correlation between labels and words in multi-label text classification,this dissertation proposes a method based on topic model and label-word interaction.The method utilizes the Labeled-LDA topic model to generate the probability distribution of labels for words in the training set,which measures the contribution of each word to different labels.Combining the text features extracted by a bidirectional long short-term memory network,the method designs a label-word interaction mechanism to compute the text representation of a specific label,guiding the word vectors to map towards different label directions.The experimental results indicate that our proposed method outperforms other benchmark methods on two public datasets.(2)To address the problem of correlation between labels in multi-label text classification,this dissertation proposes a method based on graph convolutional neural network(GCN).The method utilizes label co-occurrence to construct a label graph to represent the complex relationships between labels.By applying GCN to learn the dependency propagation between labels,the proposed method effectively captures high-order dependencies between labels.The experimental results indicate that our proposed method outperforms benchmark methods in terms of evaluation metrics Hamming-Loss and Micro-F1.