Modeling label correlations implicitly through latent label encodings for multi-label text classification

Multi-label text classification (MLTC) aims to assign a set of labels to each given document. Unlike single-label text classification methods that often focus on document representation learning, MLTC faces a key challenge of modeling label correlations due to complex label dependencies. Previous state-of-the-art works model label correlations explicitly. It lacks flexibility and is prone to introduce inductive bias that may not always hold, such as label-correlation simplification, sequencing label sets, and label-correlation overload. To address this issue, this paper uses latent label representations to model label correlations implicitly. Specifically, the proposed method concatenates a set of latent labels (instead of actual labels) to the text tokens, inputs them to BERT, then maps the contextual encodings of these latent labels to actual labels cooperatively. The correlations between labels, and between labels and the text are modeled indirectly through these latent-label encodings and their correlations. Such latent and distributed correlation modeling can impose less a priori limits and provide more flexibility. The method is conceptually simple but quite effective. It improves the state-of-the-art results on two widely used benchmark datasets by a large margin. Further experiments demonstrate that its effectiveness lies in label-correlation utilization rather than document representation. Feature study reveals the importance of using latent label embeddings. It also reveals that contrary to the other token embeddings, the embeddings of these latent labels are sensitive to tasks; sometimes pretraining them can lead to significant performance loss rather than promotion. This result suggests that they are more related to task information (i.e., the actual labels) than the other tokens.