Beyond Bit Fidelity: Dual-Mode Domain-Aware Knowledge-Enhanced Semantic Text Communication Framework

초록

Semantic communication has emerged as a transformative paradigm for sixth-generation (6G) wireless networks, prioritizing meaning preservation over bit-level accuracy. However, existing frameworks face critical limitations including insufficient terminology preservation, prohibitive API costs for LLM-based reconstruction, and inability to adapt to varying channel conditions. This paper presents a dual-mode domain-aware semantic communication framework integrating knowledge-enhanced compression, content-adaptive unequal error protection (UEP), and reinforcement learning (RL)-driven intelligent reconstruction. Mode 1 (cloud/API-assisted) achieves 98.2% semantic similarity and 96.8% BLEU using GPT-4-turbo for quality-critical applications, while Mode 2 (edge/local) achieves 87.8% semantic similarity and 82.3% BLEU with zero external dependency using a local Mini-LLM (355 M parameters) for latency-critical scenarios. Key contributions include: (1) entropy-adaptive VAE compression with dynamic dimensionality and domain-specific knowledge base enhancement for improved terminology preservation; (2) gradient-based dimension-level UEP with multi-level modulation allocating stronger protection to semantically critical embedding dimensions; and (3) PPO-based multi-path reconstruction intelligently selecting among DVAE-enhanced, pattern-based, and ensemble strategies based on channel conditions and semantic quality indicators. Extensive evaluation on Europarl, News, and Technical corpora across SNR 0-20 dB demonstrates statistically significant improvements over DeepSC, R-DeepSC, and KG-LLM-SC while enabling flexible deployment from cloud-assisted to real-time edge computing.

키워드