Adaptive Goal-Oriented Semantic Communication with Pretrained Semantic Encoding and Effectiveness-Aware Equalization for Beyond-5G/6G Networks

Abstract

Semantic communication seeks to transmit task-relevant meaning rather than reproduce every source bit, but learned semantic representations can deteriorate sharply when they are perturbed by channel noise. This paper presents an adaptive, goal-oriented semantic communication framework that combines a pretrained Sentence-BERT (SBERT) encoder with a noise-conditioned neural equalizer. Text is mapped to a 384-dimensional semantic vector, perturbed by additive white Gaussian noise, and processed by the equalizer only when the estimated noise standard deviation exceeds an activation threshold of τ=0.05. A cosine-similarity rule then determines whether the recovered representation is accepted or sent for refinement or retransmission. In the reported embedding-domain experiments, adaptive equalization increased mean semantic similarity from 0.2475 to 0.6303 at σ=0.20 and from 0.0740 to 0.3545 at σ=0.70. The corresponding absolute gains were 0.3828 and 0.2805, respectively. At σ≤0.05, the equalizer was bypassed and produced no measurable change, avoiding unnecessary correction under mild perturbations. These results establish a proof of concept for noise-triggered recovery of pretrained semantic embeddings.