中国电化教育

随着教育数字化转型以及义务教育均衡发展的需求,大语言模型作为新兴的通用认知工具,正逐步应用于“整本书阅读”等语文教学场景。大语言模型能否理解经典名著,是决定其能否有效赋能语文整本书阅读的关键因素。基于此,文章构建了一个包含14部名著、500道选择题的测评数据集,覆盖语言特色、逻辑推理等10类问题,系统评估Deep Seek-V3、Qwen2.5-7B-Instruct等10个主流大模型的阅读理解能力。结果表明:(1)大语言模型总体正确率达82.5%,可以理解中学阶段的名著,其中国内模型表现显著优于国外模型,不同参数规模的大模型的表现不存在显著差异;(2)基于10个大语言模型的得分计算每道题的难度并绘制不同难度的题目频率分布,该分布为幂律分布而非常见的正态分布;(3)大语言模型的小说类名著与文言类名著阅读理解表现较弱,在生成创造任务中优势显著而在细节提取等精准定位任务中表现欠佳等。文章针对大语言模型在整本书阅读领域的应用提出面向教学层、内容层与技术层三个方面的建议,为推动整本书阅读教学提供智能化解决方案。

With the advancement of educational digital transformation and the demand for balanced development in compulsory education, large language models(LLMs), as emerging general-purpose cognitive tools, are gradually being applied to Chinese teaching scenarios such as “wholebook reading.” Whether LLMs are capable of understanding classical masterpieces is a key factor in determining their effectiveness in empowering whole-book reading instruction. Based on this, the study constructs an evaluation dataset consisting of 14 masterpieces and 500 multiple-choice questions, covering 10 categories of questions including linguistic features and logical reasoning, to systematically assess the reading comprehension abilities of 10 mainstream LLMs, such as DeepSeek-V3 and Qwen2. 5-7B-Instruct. The results show that:(1) The overall accuracy rate of large language models reaches 82. 5%, demonstrating their ability to understand classic works at the secondary school level. Domestic models perform significantly better than foreign models, and there are no significant differences in performance among large models of different parameter scales;(2) By calculating the difficulty of each question based on the scores of the 10 large language models and plotting the frequency distribution of questions of different difficulties, this distribution follows a power-law distribution rather than the commonly observed normal distribution;(3) Large language models exhibit weaker performance in reading comprehension of novel-style classics and classical Chinese classics, but demonstrate significant advantages in generative creative tasks while performing poorly in precise localization tasks such as detail extraction. The article proposes recommendations targeting the teaching, content, and technological aspects for the application of large language models in the domain of whole-book reading, providing intelligent solutions to advance the instruction of whole-book reading.