CoocNet: a novel approach to multi-label text classification with improved label co-occurrence modeling

Multi-label text classification (MLTC) aims to assign one or more labels to each document. Previous studies mainly use the label co-occurrence matrix obtained from the training set to establish the correlation between labels, but this approach ignores the noise in label co-occurrence, and applies the ungeneralizable label co-occurrence relationship to model testing and validation. In addition, labelling co-occurrence relationship globally lacks attention to a specific document, which results in the loss of the local label co-occurrence relationship. To address this issue, we introduced a new multi-label text classification model in this study, presenting CoocNet, which adopts a two-step label detection to effectively tackle the challenge of modeling label co-occurrence relations. The model first captures the global co-occurrence relationships of labels using the label co-occurrence matrix and denoises the label noise through the label denoising attention mechanism, and then uses a contrast learning strategy to capture the local label co-occurrence relationships among specific different documents. In particular, we unify the co-occurrence labeling into an auxiliary training task that runs parallel to the multi-label classification task. The new task supervises the learning of sentence representations for documents by leveraging the modeled label co-occurrence relationships, enhancing the model's generalization ability. Another novelty is that the auxiliary task is only active during model training, thereby preventing label co-occurrence relationships from interfering with the model's predictions outside the training phase. The experimental results on three benchmark datasets (Reuters-21578, AAPD, and RCV1) demonstrate that our model outperforms the existing state-of-the-art methods.