QwenMoE-SC: A Mixture-of-Expert Semantic Communication Model with GNN-Based Unequal Error Protection, NEFTune Technique and Direct Preference Optimization

We propose QwenMoE-SC, a semantic communication framework that integrates a Mixture-of-Experts (MoE) Large Language Model with three complementary modules: (1) a Graph Neural Network (GNN)-based Unequal Error Protection (UEP) plug-in that assigns semantic importance scores via syntactic dependency graph message passing for adaptive bit allocation, without modifying the pre-trained LLM; (2) NEFTune noise injection during fine-tuning for channel robustness; and (3) a Communication-aware Direct Preference Optimization (C-DPO) strategy that favors semantically faithful yet token-efficient transmissions. Comprehensive ablation studies on AWGN and Rayleigh fading channels show that each component contributes distinct gains, and their combination consistently outperforms traditional separation-based methods and neural baselines in sentence similarity, BLEU score, and semantic-level BER, with the largest improvements at low-to-mid SNR regimes. QwenMoE-SC can also serve as a semantic interface layer within expert and decision-support systems, enabling robust, bandwidth-efficient communication between data sources, inference engines, and human users.