Abstract: Extreme multi-label text classification(XMTC) is a challenging task that involves finding the most relevant labels from a large and complex label set for a given text sample.Currently,deep learning methods based on the Transformer model have achieved great success in XMTC.However,existing methods have not fully utilized the advantages of the Transformer model,ignoring the subtle local semantic information of texts at different granularities,and failing to establish and utilize the potential associations between labels and texts robustly.To address this issue,this paper proposes SemFA model—an extreme multi-label text classification model based on semantic features and association-attentionthat leverages semantic features and association attention for XMTC.In SemFA,the top-level outputs of multiple encoders are firstly concatenated as global features.Then,a con-volutional neural network is used to extract local features from shallow vectors of multiple encoders.By combining the rich global information and subtle local information at different granularities,more accurate and comprehensive semantic features are obtained.Finally,the potential association is established between label features and text features using an association-attention mechanism,and an association loss is introduced to continuously optimize the model.Experimental results on the Eurlex-4K and Wiki10-31K public datasets show that SemFA outperforms most existing XMTC models,effectively integrating semantic features and association attention to improve overall classification performance.

Key words: Natural Language Processing, Extreme multi-label text classification, Semantic features, Pre-trained models, Attention mechanisms