Semantic Relaying of Sentences

Semantic communications in a relay network is considered. In particular, we focus on efficient semantic text communication. We propose semantic sentence forwarding (SSF), which utilizes the attention mechanism to transmit a single embedding capturing the entire message semantics. At the relay, SSF decodes the source signal autoregressively, one token at a time, incorporating previously generated tokens and the context embedding from the source node. In addition to traditional BLEU and SBERT-based semantic similarity metrics, we introduce a GPT-enabled semantic similarity score for semantic evaluation of the reconstructed messages. Extensive experiments demonstrate that SSF outperforms both semantic-agnostic baselines, including classical transmission and autoencoder-based designs. SSF also outperforms existing token-level semantic methods SLF and SPF, demonstrating the effectiveness of sentence-level semantic text communication.