Integrating Semantic-Space Finetuning and Self-Training for Semi-Supervised Multi-label Text Classification

To meet the challenge of lack of labeled data in document classification tasks, semi-supervised learning has been studied, in which unlabeled samples are also utilized for training. Self-training is one of the iconic strategies for semi-supervised learning, in which a classifier trains itself by its own predictions. However, self-training has been mostly applied to multi-class classification, and rarely applied to the multi-label scenario. In this paper, we propose a self-training-based approach for semi-supervised multi-label document classification, in which semantic-space finetuning is introduced and integrated into the self-training process. Newly discovered credible predictions are used not only for classifier finetuning, but also for semantic-space finetuning, which further benefit label propagation for exploring more credible predictions. Experimental results confirm the effectiveness of the proposed approach and show a satisfactory improvement over the baseline methods.