This project focuses on the application of semantic communication in 6G networks, as the evolution of network technology propels communication into the 6G era, introducing new application scenarios like holographical communication, the Internet of Things (IoT), etc. These emerging technological landscapes impose advanced requirements necessitating corresponding advancements in communication technologies. Specifically, our focus lies in the application of semantic communication within the 6G framework. Shannon and Weaver's theory about information transmission comprises three levels: bit-level information transmission, semantic-level information transmission, and application-level information transmission. Our emphasis in this project is about semantic information transmission, delving into the extraction of semantic information embedded in speech, language, images, audio files, etc. This realization is made feasible through recent technological advancements in deep learning and large-scale language models (LLM), bringing forth advantages such as heightened information density, optimization of communication channel capacity, and acceleration of communication transmission speed. Our methodology relies on machine learning and deep learning techniques, particularly autoencoder-based architectures, to model semantic information processing. The project explores two primary directions: 1. Dynamic neural networks – investigating model compression and adaptive inference techniques to reduce computational overhead during semantic encoding and decoding, thereby improving transmission efficiency. （Most experiments were finished before 2025/05/01, and now one paper has been submitted to IEEE Transactions on Communication for revision.) 2. Multimodal semantic communication – exploring the joint encoding of multiple data modalities to enhance the expressive capacity and flexibility of semantic transmission.