Semantic Noise Decomposition and Saliency-Weighted Mutual Information in Task-Oriented Communication

This work characterizes semantic noise from a task-oriented perspective in semantic communication systems. Semantic distortion is decomposed into source-side, channel-side, and receiver-side components, revealing that source and receiver semantic noise can dominate task performance even when channel-induced semantic errors are negligible. A dependency-aware, saliency-weighted semantic mutual information metric is introduced to quantify task-relevant meaning preservation and to establish a direct relationship between semantic fidelity and task distortion under a semantic rate–distortion framework. Triplet-based semantic examples and illustrative analysis validate how semantic noise and task saliency jointly influence task performance, providing an interpretable foundation for saliency-aware and robust task-oriented semantic communication design.