<p>Facial expression recognition (FER) in natural settings is often hampered by label noise due to subjective annotations and ambiguous expressions. While previous studies have attempted to address this through uncertainty estimation and relabeling, they have overlooked the accuracy of uncertainty learning and the adaptability of label correction. We propose Adaptive Label Modification based on Uncertainty Learning (ALM-UL), a novel approach that dynamically adjusts noisy labels based on learned uncertainty without manual sample selection. ALM-UL consists of two key components: (1) an uncertainty learning module that obtains precise uncertainty value by focusing on relatively challenging samples, and (2) an adaptive label modification module that revises noisy labels using the learned uncertainty. This approach allows ALM-UL to concentrate on critical facial samples and implement a parameter-free relabeling mechanism, effectively mitigating the impact of uncertain samples. Our method is easily implementable with minimal additional parameters. Experiments on both synthetic and real-world datasets demonstrate that ALM-UL significantly outperforms state-of-the-art algorithms, achieving an average improvement of 2% in recognition accuracy.</p>