Wrapper methods that exploit pseudo-labelling techniques are among the most widely used semi-supervised learning methods. However, the effective selection of high-quality pseudo-labelled data for model training remains a significant challenge. In this paper, a novel self-organizing fuzzy belief ensemble classifier is proposed for semi-supervised learning from partially labelled data, addressing this challenge through the pioneering use of paraconsistent logic. The proposed ensemble classifier consists of multiple self-organizing fuzzy belief inference systems trained as binary base classifiers, each distinguishing one specific class from all others. Predictions generated by the base classifiers are integrated by a paraconsistent logic-based decision-maker, which quantitatively analyzes the levels of uncertainty and contradiction in the predictions to select high-quality unlabelled samples as pseudo-labelled data. This mechanism enhances the effectiveness of the self-training process and improves the overall classification accuracy. Extensive experiments on 14 benchmark datasets demonstrate that the proposed method consistently outperforms a range of state-of-the-art semi-supervised learning approaches with higher classification accuracy, showcasing its strong potential for real-world applications where labelled data is scarce or expensive to obtain.