Multi-module Fusion Relevance Attention Network for Multi-label Text Classification

To solve the multi-label text classification (MLTC) task, we propose a multi-module fusion relevance attention network (MFRAN) to explore the semantic correlation between text and category labels. Firstly, the MFRAN model uses a text feature extraction module to capture text information with a strong correlation with category labels and uses multi-head self-attention to obtain the attention score of the corresponding text. Then the learned word-level text semantic information is transmitted to the label attention layer of the category label feature extraction module through multi-dimensional dilated convolution. At the same time, the attention score of category labels is obtained by the bidirectional long short-term memory and label attention layer. The adaptive attention fusion module is used to fuse the text attention score with the attention score of the category label and select the text representation with large output information. We performed a large number of comparative experiments and ablation experiments on the RCV1-V2 and AAPD datasets. The experimental results have proved the MFRAN model is similar to or even exceeds the baseline model when dealing with MLTC tasks.