Cross-Topic Text Auto-Scoring Method Based on Generative Adversarial Networks

Automatic text scoring has important advantages in reducing manual work and ensuring the consistency and objectivity of scoring, and has received great attention in the current online education. However, most of the existing automatic text scoring methods are based on specific themes and task environments, which are difficult to migrate to different theme scenarios, so they are difficult to be applied to small sample scoring tasks with various themes such as online classes. Therefore, in order to solve the problem of inconsistent distribution of source topic and target topic and the lack of label data in the training process, this paper proposes a cross topic text automatic scoring method based on adversarial learning and knowledge distillation (KDALTS). This method learns the invariable features between the source topic and the target topic using the adversarial learning framework, and uses knowledge distillation as a regularization method, so that the tag information of the learned source topic can be added in the adversarial training process, further improving the generalization of the scoring method. Experimental evaluation on the open ASAP-SAS dataset, which is divided into dataset 1 and dataset 2 based on two scoring criteria. The experimental results show that the average quadratic weighted kappa value (QWK) of the KDALTS scoring method is 70.82% for the single source cross topic scoring task on dataset 1, and 72.70% for the multi-source cross topic scoring task. The average quadratic weighted kappa (QWK) of the single source cross topic scoring task on dataset 2 reached 73.92%, and the average quadratic weighted kappa (QWK) of the multi-source cross topic scoring task reached 72.67%. All the results are better than the baseline method, which fully shows the effectiveness and generalization of the method.