基于Capsule Network的数学简答题自动反馈

AUTOMATIC FEEDBACK OF MATHEMATICAL SHORT-ANSWER QUESTIONS BASED ON CAPSULE NETWORK

自动反馈是指通过参考答案、教师意见等自动给出学生答案的修改意见。数学简答题答案具有类型有限和易混淆的特点,对此,提出一种自动反馈的方法。使用基于静态路由的Capsule network预测答案类型,根据答案类型通过相似度计算确定反馈内容。构建一个数学简答题自动反馈数据集并据此进行实验。实验结果表明,基于静态路由的Capsule network在确定答案类型任务上表现更好,相对于CNN有3百分点至8.3百分点的提升,相对于基于动态路由的Capsule network有0.3百分点至2.5百分点的提升,总体上能达到85.3%以上的准确率。

Automatic feedback refers to the modification of students answers through reference answers,teachers opinions,etc.The students answers to the short answer questions of mathematics have the characteristics of limited types and easy to be confused.In view of the above characteristics,a method to determine the feedback content is proposed.The Capsule network based on static routing was used to predict the answer type,and then the feedback content was determined through similarity calculation according to the answer type.An automatic feedback data set of mathematical short answer questions was constructed and the experiment was carried out accordingly.The experimental results show that the Capsule network based on static routing performs better in determining the answer type tasks,with an improvement of 3 percentage points to 8.3 percentage points compared with CNN,and 0.3 percentage points to 2.5 percentage points compared with the Capsule network based on dynamic rounting,with an overall accuracy of more than 83.3%.