Automatic ICD coding is an extreme multi-label classification task that requires assigning standardized diagnostic codes to clinical notes. Recent pretrained language model (PLM) based methods achieve strong performance, but they often overfit to repetitive semi-structured patterns in discharge summaries rather than learning clinically meaningful semantics. Moreover, existing approaches insufficiently utilize the rich semantic information contained in ICD code descriptions. In this study, we propose SARL-ICD, a structure-aware framework for automatic ICD coding centered around two key ideas: contrastive representation learning and label-text semantic interaction. First, we introduce JMoCo, a Jaccard-guided MoCo-style pretraining strategy that constructs positive pairs using ICD-code similarity while restricting pretraining to clinically informative narrative sections. This design encourages the model to learn patient-specific clinical semantics. Second, we propose a lightweight label-text interaction module that injects semantic information from LLM-augmented ICD code descriptions into document representations. We further employ an adaptive fusion module to combine complementary strengths of different ICD coding architectures. Extensive experiments on MIMIC-III-full and MIMIC-IV demonstrate that our method achieves state-of-the-art performance across multiple evaluation metrics. Additional analyses further show that JMoCo improves representation similarity alignment and reduces structural shortcut learning in semi-structured clinical documents.