基于多粒度标签扰动的文本分类研究OACHSSCDCSTPCD

[Purpose/Significance]Supervised learning algorithms based on deep learning are currently the main re-search methods for text classification.However,the training of supervised deep learning algorithms heavily relies on the ac-curacy of the sample labels.Due to the annotator's experience and subjectivity,sample labels inevitably contain noise.La-bel perturbation is an effective way to deal with noisy labels.However,noisy label learning algorithms based on label per-turbation lack effective utilization of multiple granularity information at present,which limits the performance of the algo-rithms.[Method/Process]In order to address the problem,the paper proposed a multi-granularity label perturbation al-gorithm(MGLP),which combined sample-level granularity and category-level granularity perturbation methods.The MGLP algorithm used the idea of meta-learning to learn the fusion weights of different granularity perturbation methods,which could adaptively adjust the fusion weights according to different data characteristics.[Result/Conclusion]The paper conducts experiment on three text classification datasets,including tweet sentiment classification,movie review sentiment classification,and citation intent classification.The results show that the proposed MGLP algorithm effectively improves the performance of deep learning models in text classification tasks and has broad application prospects in information organiza-tion and information analysis.