【摘要】 作文自动评分（Automated Essay Scoring,AES）作为自然语言处理技术在教育领域的典型应用，在国内外有着广泛的研究。其中，跨主题AES又是目前AES新的研究热点。当前的跨主题AES在源作文集与目标作文集共享特征的提取方面存在不足。针对该问题，论文提出一种基于二阶段架构的跨主题作文自动评分方法，一阶段通过混合人工方法和深度学习方法提取共享特征并对目标作文进行预打分标注，二阶段将标注数据作为伪数据训练一个基于注意力机制的深度神经网络实现跨主题评分任务。与基线模型的实验结果表明，研究设计的模型在跨主题的场景表现要优于其他AES模型。更多还原

【Abstract】 As a typical application of natural language processing technique in education,Automated Essay Scoring(AES)has been extensively studied both at home and abroad. Among them,cross-topic AES is a new research hotspot of AES. The current cross-topic AES is deficient in the extraction of shared features between source and target collections. To solve this problem,this paper proposes a cross-prompt essay automatic scoring method based on two-stage architecture. In the first stage,it extracts shared features and pre-marks the target essays through a mixture of manual methods and deep learning methods. In the second stage,it trains a deep neural network based on attention mechanism as pseudo data to achieve cross-topic scoring tasks. The experimental results with the baseline model show that the model designed in this study performs better than other AES models in cross-topic scenarios.更多还原