Crowdsourcing leverages collective intelligence for large-scale annotations, but varying worker expertise and task difficulty introduce noise, making truth inference crucial for data quality. Recent graph-based methods effectively capture worker–task–answer relations, but they overlook the long-tail phenomenon: most workers provide only a limited number of answers while a minority contribute many, most tasks receive only a few annotations whereas a small fraction gather a large number. These sparse nodes lack sufficient context, leading to unstable representations and degraded inference accuracy. To address this, we propose CAGCTI, a Community-Aware Graph Contrastive learning framework for long-tail crowdsourcing Truth Inference. CAGCTI detects communities on worker–task bipartite graphs via an improved Bi-Louvain algorithm to capture locally consistent annotation patterns, and guide the construction of community-aware augmented views to enrich context for long-tail nodes. A structure–semantic collaborative contrastive learning framework is further designed, where community-aware contrast reinforces intra-community coherence and a semantic consistency constraint preserves the separability of task nodes with different answer classes. Through these, CAGCTI mitigates long-tail bias, suppresses noise, and enhances representation quality. Experiments on 8 real-world datasets demonstrate that CAGCTI achieves superior robustness, accuracy, and generalization under noisy, long-tail conditions.