Semisupervised text classification leverages a small amount of labeled data alongside a large volume of unlabeled data to train text classification models. However, existing pseudo-labeling methods suffer from error accumulation due to underfitting of decision boundaries and pseudo-label bias. To address these issues, we propose a crossannotation model based on multigranularity semantic margin loss. Our approach begins by capturing global and local semantic features through a multigranularity semantic extraction model. These features are then fused, taking into account the semantic similarities between different categories, to form new semantic features, thereby mitigating the impact of error accumulation. Additionally, we introduce a categoryrelated margin, where pseudo-labeled samples obtained through adaptive local threshold learning are used to compute positive and negative sample losses with the margin. This forces samples of each category to form highdensity distributions in the embedding space, alleviating underfitting of decision boundaries. Experimental results demonstrate the superior performance of our model.