Level-specific feedback generation for scene descriptions via fine-tuning multimodal large language models

Texto completo

Scene description tasks effectively enhance students' English writing skills in contextual settings, facilitating the establishment of authentic situational connections. However, evaluating descriptive quality and providing accurate, level-appropriate feedback present significant challenges. Although Multimodal Large Language Models (MLLMs) have demonstrated strong capabilities in vision-language tasks, their generated feedback for scene description tasks often remains generic. It fails to account for students' educational stages. To address this limitation, we construct a novel level-specific feedback dataset for scene description tasks. This dataset is constructed using GPT-4o with Retrieval-Augmented Generation (RAG), guided by the Hong Kong primary and secondary school English word lists, which categorize vocabulary into four educational stages (key stages 1–4). We fine-tuned a designed MLLM on this dataset and evaluated its performance against open-source and closed-source baselines. Experimental results demonstrate that the proposed fine-tuned MLLM significantly enhances educational stage relevance in feedback generation while reducing hallucinated content. These findings substantiate the efficacy of fine-tuned MLLM in providing level-specific feedback for scene description tasks, advancing the potential for more adaptive AI-assisted writing support in educational contexts.