综合多影响因素的试题难度自动预测模型构建研究

A Study on the Construction of an Automatic Model for Predicting the Difficulty of Item by Integrating Multiple Influencing Factors

合理设计难度适宜的试题资源是实现学习者知识状态精准评估的基础和前提,也是实现个性化学习的关键因素。为有效避免当下试题库建设中试题难度预测方法存在的预测题型单一、忽视知识认知目标等问题,继而提出基于高级机器学习技术,综合考虑试题文本、试题结构、知识深度和认知目标等试题难度多影响因素,构建包括试题文本表征、试题多知识点抽取、认知目标下试题难度预测并适应选择、填空、简答等多种题型的试题难度自动预测模型。基于智慧教育云平台数据分析后发现,综合多影响因素的试题难度自动预测模型更具实用性和可靠性。基于构建模型研发了高中选科智能推荐系统,选科推荐准确率达到了91%,实证了模型在题库构建、学习者知识状态精准评估和个性化学习服务等方面具有较高的应用价值。

Reasonable design of test question resources with appropriate difficulty is the basis and prerequisite for achieving accurate assessment of learners’ knowledge status and a key factor for realizing personalized learning. In order to effectively avoid the problems of single type of predicted questions and neglecting knowledge cognitive goals in the construction of current test bank, the article proposes to build an automatic model of test difficulty prediction based on advanced machine learning technology, taking into account multiple influencing factors of test difficulty such as test text, test structure, knowledge depth and cognitive goals, including test text characterization, multiple knowledge points extraction of test questions, test difficulty prediction under cognitive goals and adapting to multiple choice, fill-in-the-blank and short answer questions. The model of automatic difficulty prediction of test questions includes text characterization of test questions, extraction of test questions with multiple knowledge points, difficulty prediction of test questions under cognitive objectives and adaptation to multiple choice, fill-in-the-blank, short answer and other question types. Based on the data from the intelligent education cloud platform, it was found that the automatic test difficulty prediction model with multiple influencing factors was more practical and reliable. Based on the constructed model, we developed an intelligent recommendation system for high school subject selection, and the accuracy rate of subject selection recommendation reached 91%, which proved the high application value of the model in question database construction, accurate assessment of learners’ knowledge status and personalized learning services.