<p>Phishing attacks pose significant threats to personal and financial security, as cybercriminals employ deceptive websites to steal sensitive information. Existing methods heavily depend on manually labeled datasets, which are expensive to construct, prone to labeling noise, and quickly become outdated due to the rapid emergence of new URLs. This limitation hinders model adaptability to evolving phishing strategies. To address these challenges, we propose a novel semi-supervised phishing detection framework that exploits contextual relationships among URLs through optimal transport theory, Sinkhorn optimization, and task-specific loss functions. By enhancing pseudo-labeling quality for unlabeled data, our method effectively mitigates the impact of noisy annotations and improves model robustness. Extensive experiments conducted on benchmark phishing datasets demonstrate the superior performance of our approach, achieving up to a 5% improvement in F1-score under 40% label noise compared with state-of-the-art baselines.</p>