Gated convolutional neural network for fine-grained automatic essay scoring

• Main Author: Lee, Xing Zhao
• Other Authors: Hui Siu Cheung
• Format: Final Year Project
• Language: English
• Published: 2017
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/72840
• Institution: Nanyang Technological University

This report proposes a novel method for incorporating character level information into neural models for Automatic Essay Scoring (AES). Intuitively, character level information is highly important to the AES task but is unfortunately neglected in many newly proposed neural models. Notably, the usage of character level information is also an issue of correctness aside from performance. Firstly, character level information is capable of modelling fine-grained features that are predictive of writing ability such as spelling or grammatical errors. Secondly, it also able to mitigate and handle out-of-vocabulary (OOV) words. Given that vocabularies are often constructed based on frequent words, assigning unknown tokens (UNK-ing) can be unfair to students who use rare and difficult words. Additionally, spelling mistakes tend to be UNKed which might cause an essay to become semantically incomprehensible. However, using both characters and words concurrently may incur a risk of overfitting. As such, a novel neural mechanism that incorporates character level information and models the relationship between words and characters using novel neural gating mechanism has been designed. To this end, two variations of word-char gating for adaptively controlling information fusion between words and characters are proposed. Extensive experiments show that the proposed neural word-char gating achieves state-of-the-art performance on the ASAP Kaggle dataset for AES while containing significantly lesser parameters than many other published works and neural network architectures. Ablation studies confirm the effectiveness of the proposed neural gating mechanism over numerous neural and feature engineering baselines. A web system prototype is also built to demonstrate how the system works.