MFDNN: Mixed Features Deep Neural Network Model for Prompt-independent Automated Essay Scoring
Chang Liu, Gejian Ding · 2021 4th International Conference on Algorithms, Computing and Artificial Intelligence · 2021
Most of the existing Automatic Essay Scoring (AES) models are prompt-dependent models that need the rated essays of specific prompt for training. However, there are few studies on prompt-independent AES. This paper studies how to fully use the effective prompt-dependent features to solve the prompt-independent AES problem. Different from the common method of only extracts multiple features, we consider reducing the interference between different features. We propose a new feature, called the deep dependent feature, which is extracted from the essay by a deep neural network. It is the representative feature that can distinguish the prompt label of the essay and has less overlap and conflict with other features. Firstly, we pre-scored the unrated target prompt data to generate pseudo data based on the manually extracted features. Then we build a new model, which is training on pseudo data to learn prompt-dependent information. Our model considers relevance feature, syntactic feature, and deep dependent feature. The performance of our model is evaluated on ASAP datasets, and the results show that our model outperforms the existing methods for prompt-independent AES.