【摘要】 实验报告自动批改是智慧教育领域的重要任务。遵循OBE理念，文章复现教师批改思路，将评分项转化为问题，统筹文字、表格、图片等多模态作答信息进行评分，以贴近实际教学和课程建设需求。多模态信息的理解和评分阶段，在深度学习的基础上，引入大语言模型，以实现表格题的内容提取与转化，解决定位和逻辑判别难点。对于文字类内容，采用BERT进行理解；对于图片类内容，则使用BERT和ResNet-18组合构建的自主训练模型，将图片匹配权值缩放后用于图像特征评价图形题。该方案采用小样本数据进行训练，适配不同学科实验，克服依赖大量数据训练导致的泛化性和迁移性不足等痛点。通过两门课程的批改试验，报告评分平均准确率达到92.20%，填补了非定制化实验报告自动批改的空白。更多还原

【Abstract】 Automatic correction of experimental reports is an important task in the field of intelligent education.Following the OBE concept,the paper reproduces the teacher’s correction idea,transforms the scoring items into questions,and coordinates the multimodal response information such as text,tables,and pictures to score,so as to be close to the actual teaching and curriculum construction needs.In the understanding and scoring stage of multimodal information,on the basis of Deep Learning,LLM is introduced to realize the content extraction and transformation of table questions,and solve the difficulties of positioning and logical discrimination.For text content,BERT is used to understand.For the image content,the self-training model constructed by the combination of BERT and ResNet-18 is used to scale the image matching weights for the image feature evaluation in graphic questions.The scheme uses small sample data for training,adapts to different subject experiments,and overcomes the pain points such as insufficient generalization and migration caused by relying on a large amount of data training.Through the correction test of two courses,the average accuracy of the report score reaches 92.20%,bridging the gap of automatic correction of non-customized experimental reports.更多还原