On the Use of Bert for Automated Essay Scoring: Joint Learning of Multi-Scale Essay Representation

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 3416 - 3425 July 10-15, 2022 ©2022 Association for Computational Linguistics On the Use of BERT for Automated Essay Scoring: Joint Learning of Multi-Scale Essay Representation Yongjie Wang1 Chuan Wang1 Ruobing Li1 Hui Lin1,2 1LAIX Inc. 2Shanghai Key Laboratory of Artificial Intelligence in Learning and Cognitive Science {yongjie.wang, chuan.wang, ruobing.li, hui.lin}@liulishuo.com Abstract In recent years, pre-trained models have be- come dominant in most natural language pro- cessing (NLP) tasks. However, in the area of Automated Essay Scoring (AES), pre-trained models such as BERT have not been properly used to outperform other deep learning mod- els such as LSTM. In this paper, we introduce a novel multi-scale essay representation for BERT that can be jointly learned. We also employ multiple losses and transfer learning from out-of-domain essays to further improve the performance. Experiment results show that our approach derives much benefit from joint learning of multi-scale essay representation and obtains almost the state-of-the-art result among all deep learning models in the ASAP1 task. Our multi-scale essay representation also generalizes well to CommonLit Readability Prize (CRP2) data set, which suggests that the novel text representation proposed in this paper may be a new and effective choice for long-text tasks. 1 Introduction AES is a valuable task, which can promote the development of automated assessment and help teachers reduce the heavy burden of assessment. With the rise of online education in recent years, more and more researchers begin to pay attention to this field. AES systems typically consist of two modules, which are essay representation and essay scoring modules. The essay representation module extracts features to represent an essay and the essay scoring module rates the essay with the extracted features. When a teacher rates an essay, the scores are often affected by multiple signals from different granularity levels, such as token level, sentence level, paragraph level and etc. For example, the 1https://www.kaggle.com/c/asap-aes 2https://www.kaggle.com/c/ commonlitreadabilityprize/data features may include the numbers of words, the essay structure, the master degree of vocabulary and syntactic complexity, etc. These features come from different scales of the essay. This inspires us to extract multi-scale features from the essays which represent multi-level characteristics of the essays. Most of the deep neural networks AES systems use LSTM or CNN. Some researchers (Uto et al., 2020; Rodriguez et al., 2019; Mayfield and Black, 2020) attempt to use BERT (Devlin et al., 2019) in their AES systems but fail to outperform other deep neural networks methods (Dong et al., 2017; Tay et al., 2018). We believe previous approaches using BERT for AES suffer from at least three lim- itations. First, the pre-trained models are usually trained on sentence-level, but fail to learn enough knowledge of essays. Second, the AES training data is usually quite limited for direct fine-tuning of the pre-trained models in order to learn better representation of essays. Last but not least, mean squared error (MSE) is commonly used in …

Content truncated to 500 words