Design of MLP-GRU Composite Model Short Answer Grading System Based on Deep Learning
Automatic short answer grading(ASAG)plays a crucial role in alleviating the burden on educators and advancing educational digitization.To address the limitations of existing ASAG models in considering multiple aspects of short answer responses,this paper leverages deep learning techniques,using the BERT model as a foundation,to integrate a Multilayer Perceptron(MLP)and Gated Recurrent Unit(GRU),establishing an MLP-GRU based ASAG model.This model synchronously considers the entailment and similarity between standard answers and student responses,enabling a comprehensive evaluation of student answers from various angles.Initially,the BERT model is employed for feature extraction from student responses,providing rich semantic information for subsequent grading processes.Subsequently,the Perceptron and GRU models are introduced,each combined with the BERT output,to compute textual entailment and textual similarity scores concerning standard answers.This approach ensures the assessment takes into account the fluency and logical coherence of responses.Finally,the textual entailment and textual similarity scores are merged to produce the ultimate grading result.The results show that the root-mean-square error of the model is 0.266,the Kappa coefficient is 0.969,and the F1-score is 0.778,indicating that the model score is highly consistent with the manual score.
natural language processingautomatic short answer gradingBERT modeleducational assessment
NETL
NSTL
万方数据
