With rapid development of deep learning, the messages hidden through linguistic steganography technology has become increasingly difficult to explore, bringing new challenges to linguistic steganalysis. The existing neural network-based linguistic steganalysis methods fail to fully integrate global and local dependencies when extracting complex features from the text, leading to suboptimal performance in detecting hidden messages generated by sophisticated linguistic steganography. In this paper, a linguistic steganalysis method utilizes multi-granularity semantic extraction using dual-mode fusion, which enhances detection capabilities by capturing local and global dependencies between words. The experimental demonstrates that, the proposed method exhibits superior multi-granularity feature extraction and inference capabilities, achieving improved detection performance.