| We have proposed a novel approach to automated essay scoring based on recurrent and convolutional neural networks. Unlike existing systems, our approach does not rely on manually-engineered features and learns features from data. The experiments show that our approach outperforms state-of-the-art automated essay scoring systems and can be used for modeling various essay scoring traits, such as argument strength and essay organization. We have also proposed a novel method based on density estimators to identify and penalize fake (computer-generated) essays. Unlike existing methods, our module does not need fake essays in the training data. This module maps the essays into an N-dimensional feature space and uses a simple decision rule to detect fake essays. We have shown that our module is able to identify fake essays generated based on N-gram language models and context-free grammars. |
