| Scientific argumentation ability is an important evaluation content of international education evaluation projects,and China’s compulsory education quality monitoring released in 2017 also clearly pointed out the need to strengthen the cultivation of students’ scientific argumentation ability.Evaluating open answers to scientific questions can help improve students’ scientific reasoning abilities,but manual real-time evaluation is difficult to achieve.Therefore,using computer technology to provide automatic real-time scoring for students’ scientific reasoning texts has great research value and significance.In previous studies,traditional machine learning methods were mainly used to grade scientific argumentation texts,including SVR,SVM,decision trees,etc.Although good grading results were achieved,there were also certain drawbacks,such as a tendency to give higher scores to longer answers and the inability to recognize polysemy problems.Some researchers have begun to introduce deep learning technology into the field of automatic scoring and have achieved better human-machine scoring consistency than traditional machine learning.However,there is relatively little research on the use of deep learning in the field of automatic grading of scientific argumentation texts.Therefore,this study proposes a scientific argumentation scoring method based on deep learning and verifies the scoring effectiveness of the model on both English and Chinese datasets.The main research work of this article is as follows:(1)This study proposes a hybrid model based on the development of previous technologies,which combines Bert,Bi LSTM,and CNN.Bert can effectively solve the problem of polysemy in text;Bi LSTM encodes text data in both forward and reverse directions,and can extract deep Semantic information;CNN can read local information of text through high-dimensional representation of text data.The Bert Bi LSTM CNN hybrid model combines the advantages of the three models to improve scoring accuracy.(2)Compare the effectiveness of a single deep learning model and a pairwise combination of deep learning models with the model proposed in this study.Due to the fact that the scoring problem can be viewed as a classification problem,this study evaluates the effectiveness of the hybrid model and baseline model from the perspectives of human-machine scoring consistency and classification effectiveness.The results indicate that the hybrid model based on deep learning proposed in this study outperforms the baseline model in terms of human-machine consistency and classification accuracy.(3)Due to the lack of publicly available scientific evidence in Chinese,this study manually translated the publicly available English dataset into a Chinese dataset to verify the effectiveness of the model.The results indicate that the model performs slightly better on English datasets than on Chinese datasets. |
