基于抽象语法树的编程题静态阅评算法

Static Evaluation Algorithm of Programming Problems Based on AST

针对传统计算机程序评阅过程中出现的人工评阅劳动量大、运行评阅安全性低等问题,论文提出了一种基于抽象语法树的编程题静态阅评的算法。首先,将答案源代码按语句分割并生成对应语句的抽象语法树,引入基于双向编码器的预训练模型实现节点向量化;然后,通过特定的卷积神经网络和双向树卷积神经网络提取词汇特征和结构特征并实现特征融合;最后,将融合特征输入深度神经网络进行多分类阅评。实验结果表明:该算法较传统静态分析算法,评阅准确率提升了5.1%,是一种可行的编程题评阅方法。

Aiming at the problems of heavy manual reviewing labor and low running review security in the process of tradition-al computer program review,this paper proposes an algorithm for static review of programming questions based on AST.First,it seg-ments the answer source code by sentence and generates the AST of the corresponding sentence,and introduces the pre-training model based on the bidirectional encoder to realize node vectorization.Then,it extracts the vocabulary through the specific convolu-tional neural network and bidirectional tree convolutional neural network feature and structural features are combined to achieve fea-ture integration.Finally,the fusion features are input into the deep neural network for multi-category review.Experimental results show that compared with traditional static analysis algorithms,this algorithm has a 5.1%improvement in review accuracy,and it is a feasible method for reviewing programming questions.