Automated essay scoring is one of the most exciting tasks in natural language processing, reducing massive workloads and speeding up the learning process and its effectiveness. Many researchers have made momentous efforts in this matter. However, as far as we know, most AES works have concentrated on the AES technique; no relevant paper has been seen on finding creative essays while performing automated scoring. One of the reasons is that creativity is difficult to judge. This paper explores this concern: we assume that a creative essay is more challenging to write than a common essay; if we mask part of a creative essay, then it is difficult to predict or train the masked part of the essay, while a common essay is relatively easy to predict or train. We build a generative adversarial network framework to predict (train) the hidden parts. By calculating the distance between the generated essay and the original essay, the proposed method gives a judgment of whether or not an essay is creative. We developed a small-scale dataset based on the ASAP dataset for creative essay training. The experimental results show that the proposed method is feasible for finding creative essays among the datasets.

自动论文评分是自然语言处理中最令人兴奋的任务之一，可以减少大量工作量并加快学习过程及其有效性。许多研究人员在这方面做出了巨大的努力。但是，据我们所知，大多数 AES 工作都集中在 AES 技术上；没有看到在执行自动评分的同时寻找创意论文的相关论文。原因之一是创造力难以判断。本文探讨了这个问题：我们假设一篇有创意的文章比一篇普通的文章更具挑战性；如果我们屏蔽了一篇创意文章的一部分，那么很难预测或训练文章中被屏蔽的部分，而普通的文章则相对容易预测或训练。我们构建了一个生成对抗网络框架来预测（训练）隐藏部分。通过计算生成的文章和原始文章之间的距离，该方法给出了一篇文章是否具有创造性的判断。我们开发了一个基于 ASAP 数据集的小规模数据集，用于创意论文训练。实验结果表明，所提出的方法对于在数据集中寻找创意文章是可行的。