Affected by inter-annotator cognitive differences, fatigue effects, and data poisoning, training data inevitably contains a certain proportion of noise, which severely impairs model performance. Traditional manual verification is costly and inefficient, while existing automatic detection methods generally suffer from limited precision, poor interpretability, and insufficient robustness. This paper proposes a noise label detection and correction method based on Bayesian weighted consensus inference. First, an ensemble of multiple lightweight heterogeneous models is constructed, and model prior knowledge and dataset noise are obtained on a clean validation set. Second, the model ensemble predicts noisy samples to extract two-dimensional consensus evidence. Then, prior knowledge and consensus evidence are fused, and the posterior probability of label noise is calculated via Bayesian inference to generate correction suggestions. Finally, high-confidence noisy labels are precisely screened based on the posterior probability threshold. Experimental results on three datasets show that the proposed method achieves a precision of 96.50%, a recall of 98.61%, an F1-score of 97.54%, and a correction accuracy of 95.53%, with improvements of 5–20% over mainstream methods. With a computational cost comparable to that of basic ensemble methods, the proposed approach achieves a favorable balance among precision, robustness, and interpretability. It thus offers a promising and cost-effective solution for automated quality control of large-scale annotated datasets, especially in text classification tasks.
Yang et al. (Tue,) studied this question.