Attention-Driven Semantic Transmission Scheme for AI-Native Wireless Communications

초록

This letter proposes an attention-driven semantic transmission scheme for AI-native wireless communications. Leveraging self-attention scores from a pretrained bidirectional encoder representations from Transformers (BERT) model, the framework prioritizes semantically important tokens in both initial transmission and retransmission stages, without task-specific supervision. Experiments on sentences from public web sources confirm consistent improvements over conventional baselines in semantic fidelity, measured by cosine similarity and BERTScore. This model-agnostic approach provides a practical solution for robust and bandwidth-efficient communication, supporting future AI-native systems that prioritize meaning preservation over exact symbol reconstruction.

키워드