Knowledge Graph-Guided and LLM-Based Semantic Communication for Challenging Edge Networks

Traditional wireless communication systems, based on Shannon’s information theory, are designed to ensure that every transmitted bit is received exactly as sent. This bit-level accuracy is efficient when channels are stable and bandwidth is enough, but it becomes inefficient in real environments where noise, fading, interference, or low bandwidth can distort signals. In many modern applications, such as Internet of Things (IoT) sensors, unmanned aerial vehicles (UAVs), or edge computing nodes, what truly matters is the meaning of the message. Semantic communication aligns perfectly with this shift in focus, which transitions from bit accuracy to message preservation. Instead of sending every word or symbol, the system seeks to transmit an encoded compact representation of the underlying meaning, which can then be reconstructed at the receiver using shared models of language and knowledge. Our work proposes an end-to-end semantic communication framework for text that integrates two complementary technologies. The first is the Knowledge Graph (KG), a structured network that represents entities and the relationships between them, capturing the essential semantic structure. The second component is a Large Language Model (LLM), trained to understand and generate natural language and capable of encoding and reconstructing the semantics of a message. KGs offer structured semantic grounding while LLMs handle contextual encoding and decoding, enabling efficient meaning transmission under bandwidth or noise limits. At the transmitter, a natural-language processing pipeline, based on spaCy (a widely used industrial NLP toolkit) and OpenIE (Open Information Extraction), analyzes input sentences to extract their main entities, relations, and summary statements. The result is a set of triples that form a Knowledge Graph representation. A sequence-to-sequence (seq2seq) encoder, such as T5 or BART, then performs semantic compression, transforming the text into a compact sequence of tokens and their contextual embeddings. This encoded representation is sent over a wireless channel affected by typical physical-layer damages such as fading (signal weakening due to movement or obstacles), multipath propagation (the signal taking multiple paths and arriving at different times), additive white Gaussian noise (AWGN), and interference from other devices. These defects can corrupt the transmitted sequence, challenging the receiver to recover meaning despite distortion. On the receiver side, a two-phase semantic decoder reconstructs the message. First, the same LLM used at the transmitter tries to rebuild the original message by predicting the most likely words from the received, possibly corrupted, tokens or embeddings. Then, a BERT-based masked-language model refines the output, validating and correcting uncertain words using contextual reasoning. We evaluate our system on a sentiment analysis dataset (SST-2) across different signal-to-noise ratio (SNR) levels, testing both the quality of semantic reconstruction and its robustness to transmission errors. We also perform ablation studies to assess the KG’s impact, transmission mode, and decoder configuration, and compare the results with a generic LLM. The results show that this KG-LLM hybrid framework reduces transmitted data size and processing time while maintaining high semantic fidelity under channel deterioration. This shows that transmitting meaning, rather than raw bits, can achieve a more efficient, resilient, and context-aware communication process.