Correlation Learning with Enhanced Attention Mechanism of BERT for Multi-Label Text Classification

Yusei Kato, Sheng Xu, Fumiyo Fukumoto, Yoshimi Suzuki

Abstract: Semantic features encoded in the labels have a strong influence on multi-label text classification (MLTC) performance. This paper follows the assumption and proposes an MLTC approach with correlation learning. Unlike the existing work focusing on text representation learning, the method jointly learns word-word, label-label, and word-label correlations to classify text. We propose a correlation learning that enhances the attention mechanism of BERT. The experimental results of three benchmark datasets demonstrate that our approach is comparable to the state-of-the-art (SOTA) label-encoded method. Our source codes are available at https://github.com/ffukumoto/CLEA.git.