Joint Coding-Shaping-Modulation for Digital Semantic Communications

Abstract

Existing digital semantic communication schemes rely on predefined conventional constellations. Such designs overlook the potential of optimizing constellation geometry for specific semantic tasks. Meanwhile, traditional constellation geometric shaping methods are inadequate for semantics-oriented communication tasks. To address this issue, this paper proposes a joint coding-shaping-modulation (JCSM) framework for digital semantic communications that optimizes constellation coordinates in an end-to-end manner. By integrating a learnable geometric constellation shaper with probabilistic symbol generation, JCSM enables the joint optimization of symbol distribution and signal space geometry to adapt to both task performance and channel conditions. We further develop a multi-stage training strategy to ensure stable convergence for JCSM. Experimental results demonstrate that JCSM consistently outperforms traditional digital schemes with standard constellation geometry across various signal-to-noise ratio (SNR) regimes. Notably, the learned constellations exhibit a task-aware layout that protects critical semantic features, confirming the effectiveness of semantic-driven signal space design.