Research On Automatic Short Answer Grading

The automatic grading task is to predict the score for the given student answers based on the reference answer.Automatic scoring is a research focus in smart education,which can effectively increase scoring consistency.This paper studies the methods of Chinese short answer automatic grading.The main works are as follows:(1)The scoring strategy based on attention mechanism is proposed.This paper uses the attention mechanism to describe the semantic matching between student answers and reference answers,and proposes an automatic grading model Att-Grader.The model encodes student answers and reference answers by Long Short Term Memory networks.Then,it uses the Co-Attention to capture the semantic information between student answers and reference answers,and finally predicts the score by using a convolutional neural network.Experiments on relevant datasets show that: compared with the Baseline model,the scoring model with the attention mechanism improved the accuracy metrics by 1.9%-14.7%.(2)This paper proposes a idea of building reference answer set to expand the reference answer to solve the question that the reference answer cannot completely cover the diversity of student answers.Firstly cluster the full student answers;secondly,calculate the representative answer of each cluster through similarity calculation;finally Combine representative answers into a reference answer set.Experiments on Chinese data sets show that the Att-Grader scoring model integrated with the reference answer set has improved the accuracy metrics by up to 3%.(3)This paper attempts to use data augmentation methods to overcome the problem of insufficient training data sets.Specially,the methods based on replacement,back translation and soft contextual data augmentation are used.Experiments on Chinese datasets show that: Att-Grader model with soft contextual data augmentation works best.This paper studies automatic scoring for Chinese short answer questions,its contribution is as follows:(1)the attention mechanism is used to describe the degree of matching between student answers and reference answers,and proposes an automatic scoring model Att-Grader;(2)a reference answer set is constructed to solve the problem that students have diverse answers;(3)data augmentation technologies are applied to the automatic scoring task to improve the accuracy of the automatic scoring.