The abundance of heavy data on social media enables users to share opinions freely, leading to the rapid spread of misleading content. However, existing fake news detection methods exaggerate the influence of public opinions, making it challenging to combat misinformation since its early spreading state. To tackle this issue, we propose a novel fake news detection framework through news semantic environment perception (NSEP) to identify fake news content. The NSEP framework consists of three major steps. First, NSEP divides the news semantic environment with time-constrained intervals into macro and micro semantic environments using an in-depth distinguisher module. Second, graph convolutional networks are applied to perceive the semantic inconsistencies between intrinsic news content and extrinsic post tokens in the macro semantic environment. Third, a micro semantic detection module guided by multihead attention and sparse attention is utilized to capture the semantic contradictions between news content and posts in the micro semantic environment, providing explicit evidence for determining the authenticity of fake news candidates. Empirical experiments conducted on real-world Chinese and English datasets show that the NSEP framework on Chinese datasets achieved as high as 86.8% accuracy, performing at most 14.1% higher accuracy than that of other state-of-the-art baseline methods and confirming that detecting news content through both micro and macro semantic environments is an effective methodology for alleviating early propagation of fake news. The findings also comprehensively indicate that both news items and posts are critical for the early debunking of fake news and in theories concerning information science.