Hejun Huang, Lanming Ma, Runkun Long, L. Mao
To address the common problem of noisy label interference in text classification tasks, this paper proposes a selfcorrecting text classification method for large language models to overcome the shortcomings of semantic learning shift, unstable category discrimination, and accumulation of erroneous signals under erroneous supervision. Based on the contextual semantic modeling capabilities of large language models, this method introduces key mechanisms such as label credibility evaluation, self-correction gating adjustment, historical semantic memory updating, and consistency constraint optimization during the classification process. This enables the model to dynamically identify and correct unstable supervision information during the training phase, thereby reducing the continuous impact of erroneous labels on the learning of classification boundaries. Unlike traditional methods that directly treat the original labels as fixed supervision targets, this method focuses more on the consistency between text semantic expression and supervision signals. Gradually feeding back semantic judgment results to the target update process, it enhances the model’s stable modeling ability in complex, noisy environments. Furthermore, this paper completes a unified modeling of the entire method flow, from text input, semantic encoding, noise perception, target correction, to final classification output, forming a relatively complete collaborative relationship between the various functional modules. Comparative results show that the proposed method performs well in text classification tasks, maintaining strong category recognition ability and output reliability even in noisy label interference scenarios. This study provides an effective approach for robust text classification using large language models in lowquality labeled environments, and also offers a reference for research on intelligent semantic processing methods under complex real-world data conditions.