GLASS: Knowledge Graph-Guided for LLM-Assisted Semantic Communication

Conventional wireless communication systems are designed for exact bit-level delivery, but this objective has become inefficient in noisy, bandwidth-constrained environments and misaligned with modern applications that primarily require preservation of semantic meaning. While several semantic communication solutions have been proposed, for example, using deep learning-based joint source-channel coding, recent solutions mostly rely on Large Language Models (LLMs) for implicit semantic representations, but offering limited interpretability, controllability, and structured reasoning. In this paper, we present GLASS, an end-to-end semantic communication framework for text that shifts the objective from bit accuracy to meaning fidelity. Our key contribution is a hybrid transmission architecture that integrates Knowledge Graphs (KGs) with LLMs. The system extracts entities and relations to build a structured semantic representation, applies LLM-based semantic compression, and transmits compact meaning-aware representations over impaired wireless channels. We evaluate the framework under varying channel conditions using a standard text classification benchmark and demonstrate that GLASS significantly reduces transmitted data while preserving high semantic accuracy under channel degradation, showing that structured, meaning-centric communication is a robust and efficient alternative to traditional bit-level wireless systems.