Label-Related Adaptive Graph Construction Based on Attention for Multi-label Text Classification

Multi-label text classification (MLTC) is a fundamental task in the natural language processing field, with the objective of assigning the most relevant labels to each document. Previous research has neglected the distinctions between similar labels and has not taken all of the relationships in documents into account. Furthermore, certain hidden distant correlations among components may be missing with those approaches. To address the challenges, we propose a Label-related Adaptive Graph Construction (LAGC) method based on the attention mechanism which incorporates both word and label information to capture label-related components from documents. Additionally, LAGC constructs local and global dynamic adaptive graph networks to effectively model the interactions among components. Simultaneously, a gated fusion module is implemented to integrate the outputs of these two graph networks, thereby obtaining the final representation of documents and subsequently constructing the multi-label text classifier. Extensive experimental results on two benchmark datasets illustrate that the LAGC method significantly outperforms all the baselines.