基于时序卷积网络的简答题评阅方法

Short answer questions grading method based on temporal convolutional network

提出一种基于时序卷积网络(temporal convolutional network,TCN)并结合预训练语言模型BERT的简答题评阅模型(SA-TCN).该模型使用BERT对评分答案和参考答案进行编码,在建立二者之间内在联系的同时提取深层次文本语义特征.为减少信息丢失并获取深层全局特征,基于TCN捕获多尺度语义信息.在公开数据集ASAP的set5上进行实验,结果表明,该模型的精度和二次加权Kappa分别达到86.87%和87.46%,优于Bi-LSTM、TextCNN和RNN等其他模型.

An grading model for short answer questions(SA-TCN)is proposed based on temporal convolutional network(TCN)combined with the pre-training language model BERT.The model uses BERT to encode the scoring answers and reference answers to establish the internal connection between them and extract the deep semantic features from the texts.In order to reduce information loss and capture deep global features,TCN is introduced to capture multi-scale semantic information.The comparative experiment is conducted on the set5 of public data sets ASAP.The results show that the accuracy and quadratic weighted Kappa of the model reaches 86.87%and 87.46%,respectively,which are better than other models such as Bi-LSTM,TextCNN and RNN.