Research On Automatic Scoring Methods For Cross-prompt

With the rapid development of educational technology,traditional teaching methods have been gradually replaced by online learning,online network courses,online teaching and online assessment.With the change of teaching methods,the new examination forms also change.Due to the subjectivity of manual scoring,different manual raters have great differences in the scoring of texts,and manual scoring costs a lot of time.Automatic text scoring system comes into being and has many advantages in reducing manual scoring activities,ensuring consistent scoring standards and the objectivity of scoring.And with the rise of online education in recent years,more and more attention has been paid to the automatic text scoring method,has made good progress.The existing automatic text scoring method scores the text in the context of the same prompt(the training set and test set are from the same prompt).When these scoring methods are applied to new scenarios(where the training and test sets come from different prompts),the performance of the scoring methods deteriorates.This is because there are differences between texts of different prompts.It is very expensive to train a scoring method specifically for texts of different prompts.Therefore,a new automatic scoring method is needed to ensure the reliability of scoring results when scoring texts with inconsistent themes.Therefore,this paper proposes an automatic cross-prompt text scoring method(KDALTS)based on adversarial learning and knowledge distillation to solve the problem of inconsistent distribution of source prompt and target prompt and lack of label data during training.This approach is able to learn prompt-irrelevant embedded feature representations from multiple source prompt texts and generalize them to unseen target prompt texts by using adversative learning process.In order to further improve the generalization of KDALTS scoring method,knowledge distillation is used as a regularization means,so that the learned source prompt tag information can be added in the process of confrontation training.At the same time,on the basis of adversarial learning,we use Maximum Mean Discrepancy(MMD)regularization to measure the distribution difference between the source prompt and the target prompt in the common feature space,and align the distribution between the source prompt and the target prompt in the adversarial training process.In this paper,experiments are carried out on single prompt and multiple prompts respectively to make full use of text knowledge of multiple prompts to achieve better scoring effect.On the ASAP-SAS public dataset,this paper compares KDALTS with the baseline model,and proves the effectiveness of KDALTS method,and has good generalization.