SMANet: Sequence-enhanced multi-head attention network for robust neural semantic learning in noisy computational environments

Traditional communication systems often fail to efficiently transmit meaningful information in noisy and dynamic environments, prompting the adoption of neural network architectures in semantic communication to prioritize semantic content over raw data. Existing neural models face persistent challenges in mitigating high noise interference, capturing long-range dependencies in sequences, and preserving semantic fidelity under varying conditions. This paper proposes SMANet, sequence-enhanced multi-head attention network for robust neural semantic learning in noisy computational environments. SMANet integrates multi-head attention mechanisms with a Dilated Normalization Block (DNB)—a specialized neural module for extracting local temporal features and global semantic representations—to enhance sequence processing capabilities, alleviate gradient vanishing/explosion issues during training, and improve network stability. At the transmitter, a neural semantic encoder employs dilated convolutions and normalization for robust feature extraction, paired with a channel encoder to achieve noise resilience; at the receiver, neural decoders precisely reconstruct semantics, facilitating applications in machine learning-driven cognitive systems. Experimental evaluations on AWGN and Rayleigh fading channels demonstrate SMANet’s superior performance, outperforming DeepSC by 23 % in BLEU scores, achieving a sentence similarity of 0.91 at SNR=18 dB, and maintaining 85 % semantic fidelity at SNR $<$ 6 dB, highlighting its potential for neurocomputing in resource-constrained networks.