6G-Oriented Semantic Extraction in Semantic Communications: From DL to Large AI Models

The sixth-generation (6G) communication technology is driving a deep transformation from traditional syntactic to semantic-intelligent transmission. Semantic communication (SC), as a new paradigm that breaks through the Shannon limit, focuses on transmitting the meaning rather than precise symbols and is emerging as a core component of 6G. Semantic extraction, as a key step in SC, directly determines system performance. Artificial intelligence (AI), as the core technology supporting semantic extraction, has been deeply integrated into SC. Its powerful capabilities in semantic representation and reasoning provide a fundamental driving force for semantic extraction. This paper focuses on the semantic extraction module in SC, systematically summarizes deep learning-based and large AI models (LAMs)-based semantic extraction models. Meanwhile, semantic extraction in image, text, audio, and multimodal scenarios is discussed by analyzing model architectures, advantages and limitations. Then, we propose an advanced LAM-based SC (LAM-SC) scheme for image SC systems, which introduces a modality-specific cross-training strategy to achieve efficient semantic image transmission. Simulation shows that the proposed advanced LAM-SC achieves an average improvement of 5.6% in the structural similarity index measure over the traditional LAM-SC. Additionally, incorporating a semantic extraction module into traditional communication reduces model loss by 36.77%, improving reconstructed image quilty by 11.43%. These results demonstrate the significance of semantic extraction in SC. Finally, motivated by the LAM-SC model, we outline open issues to inspire future research on LAM-based SC.