Humor Style Transfer for Virtual AI Teachers via Locally Deployed DeepSeek for Comedian Guided Script Rewriting in Virtual Educational Environments
Long Liu, Roliana Binti Ibrahim, Nor Azman Ismail, Weng Howe Chan · 2025
This study proposed a humor style transfer method for virtual AI teachers, and built the StyleDeepSeek-R1 system based on the locally deployed DeepSeek model. The system automatically captures the specified comedian corpus, extracts style vectors and combines them with structured prompts to rewrite the teaching script in a humorous style. The experiment randomly selected 500 TED texts from more than 4,000 TED talks and compared them with DeepSeek-R1 and DeepSeek-V3. The results showed that StyleDeepSeek-R1 performed best in humor structure density, semantic fidelity and style transfer strength. Unlike solutions that rely on cloud APIs, this system supports secure, low-cost local deployment and is suitable for virtual education application scenarios that have high requirements for data privacy and deployment controllability. It provides technical support for the emotional expression and multimodal presentation of virtual AI teachers in virtual education environments, and lays the foundation for future integration with multimodal virtual avatars. It has the potential to achieve coordinated presentation of voice, action and visual styles in immersive teaching environments.