Research On The Scoring Method Of Open-ended Question Answer Based On Adversarial Text

The open-ended question answer scoring task is the special task in the field of automatic text scoring.Compared with general automatic text scoring tasks,openended questions often do not have standard answers.All answers that meet the requirements of complete answers,appropriate questions,and fluency in the language can be considered as excellent answers.In order to score the answers to open-ended question more objectively and efficiently,it is necessary to create a corresponding model to automatically score the answers to open-ended question.The existing automatic text scoring models mostly rely on feature knowledge,standard answers and a lot of manual labeling.It is difficult to accurately score the answers to open-ended question without standard answers.Therefore,this paper used deep learning model based on the idea of adversarial text to research the open-ended question answer scoring task and improve the performance of the model with less annotated text.In this paper,a classification model with text pairs as input is used as the basic model of the open-ended question answer scoring task.It can automatically score the input text pairs,and finally output the classification results.There are three categories of poor,medium and good.This paper proposed the ATQA model,which uses a gradient-based white-box adversarial training method to add disturbances to the word embedding space of the input text to improve the overall performance of the model and the class balance loss function to solves the problem of category imbalance.Compared with other adversarial training methods on the Chinese data set,the accuracy is improved by 1.04%.Because the cost of obtaining annotated text is high,this paper proposed the Auto Gen method to better expand the data.The specific method is to use the autoencoder to map the gradient-based continuous disturbance to the discrete text space,and to increase the accuracy of the Ro BERTa model by 1.82% by means of data enhancement.In addition,the text generated by Auto Gen method is a sentence-level disturbance to the original text.While maintaining the original meaning of the sentence,various operations such as addition,modification and replacement are performed on the sentence.Compared with other adversarial text generation,the transformation in this method is more abundant.This paper refined the relevant dataset with real application scenarios.And used the proposed method to perform experimental verification on the data set.The experimental results prove that the method proposed in this paper can improve the performance of the model,effectively complete the open-ended question answer scoring tasks,and apply in actual scenarios.