MEPT-LLM: a multimodal generative AI model for identifying and understanding cultural-emotional barriers in the language classroom.

With the rapid transformation of the educational environment driven by artificial intelligence (AI), cross-cultural communication in teaching Chinese as a foreign language has seen significant advancements. However, traditional models still face limitations in emotion recognition, content generation, and personalized recommendations, which restrict their effectiveness in multimodal educational settings, particularly when addressing emotional barriers in language classrooms. To overcome these challenges, this paper proposes the MEPT-LLM model, which integrates three core modules: multimodal emotion perception, psychological state analysis, and personalized content generation. Experimental results demonstrate that MEPT-LLM outperforms traditional models on the MELD and MEMD datasets, with improvements in emotion perception, content generation quality, and personalized recommendations. Specifically, the model achieves a performance increase of 12%-15%. Ablation experiments further highlight the crucial role of each module, with the collaborative effect of the modules being key to the model's success. The MEPT-LLM model provides strong support for the development of intelligent learning systems in cross-cultural education, particularly in educational psychology, by optimizing learning motivation through emotional regulation and personalized feedback.