Tourism education, particularly the teaching of tourist attraction interpretation, often lacks an on-site interpretation environment, opportunities for repeated practice, and a real-time assessment mechanism. This makes the teaching of tourist attraction interpretation complex and inefficient. Traditional teaching models struggle to meet the demands of modern education for personalised learning and real-time feedback, especially in tourist attraction interpretation, which requires professional language use and the integration of multimedia resources. Educators and students in this field therefore urgently need an innovative approach to improve learning outcomes and ensure teaching effectiveness.

Against this background, this study proposes TourExAIGC, a closed-loop, integrated automated assessment system based on multimodal AI-generated content (“AIGC”) technology. The system combines large language models (“LLMs”), LoRA-based fine-tuning, retrieval-augmented generation (“RAG”), chain-of-thought (“CoT”) prompting, image and video generation tools, and multimodal analysis methods to create a comprehensive closed-loop assessment framework for evaluating and improving tourism students’ interpretation skills.

The TourExAIGC system is designed to provide a comprehensive and efficient solution for the teaching and assessment of tourist attraction interpretation. Students are provided with automatically generated videos of tourist attractions for repeated learning, alongside real-time, multi-faceted assessment reports, thereby meeting the need for a more robust and dynamic educational approach. The assessment process covers multiple dimensions, including coverage of knowledge points, logical coherence, accuracy of content expression, creativity, interactivity, feedback-based improvement, and scoring consistency. By incorporating an iterative feedback loop, the system enables students to progressively refine their skills, aligning with educational goals of developing competence and confidence in professional practice.

Figure 1. Pipeline of the Teaching and Assessment Solution for Tourist Attraction Interpretation (The red box indicates the operational steps of the system, while the green box indicates the AI models corresponding to each step)

First, students use the TourExAIGC system and input a description of a tourist attraction in Step 1, Script Drafting. The system then generates an initial video script through ChatGLM’s GLM-4-9B-Chat (128K) large language model. Second, in Step 2, Script & Image Gen, the initial video script is processed using the RAG capability of GLM-4-9B-Chat. By retrieving image descriptions from MLKB, a dataset of Macao tourist attraction information, the system generates the final descriptive script for video generation. It then calls Kling’s image-to-image API to generate the corresponding images for the video. Third, in Step 3, Video Gen, the final video-generation script and the corresponding images are used as prompts. Through video CoT technology and Kling’s image-to-video API, the system decomposes the complex task of generating a tourist attraction video into a sequence of smaller sub-video generation tasks. Each sub-video clip is generated based on the video script and corresponding image output from Step 2. The sub-video clips are then stitched together to form the final, complete video for the tourist attraction. In Step 4, Text Transcription, students use the TourExAIGC system to interpret the tourist attraction video generated in Step 3. Tencent’s Speech-to-Text API, 16kzh-PY, is used to obtain the textual transcription of the student’s interpretation, and the timestamps from the speech-to-text output are aligned with the video timeline. In Step 5, Scoring & Suggestions, the time-aligned transcription of the student’s interpretation is used as input and provided to the GLM-4-9B-Chat model fine-tuned on the TourGuide scoring and suggestion dataset. The system then generates preliminary scores and suggestions in real time. Finally, students provide feedback on the preliminary scores and suggestions generated in Step 5 and engage in repeated practice. The fine-tuned GLM-4-9B-Chat model then produces the final scoring and suggestion report in Step 6.

The system design of this study holds significant value for educators and educational software developers, as it addresses the critical tasks of evaluating educational trajectories and improving student outcomes within the context of tourism education.