A Prompt-independent and Interpretable Automated Essay Scoring Method for Chinese Second Language Writing

Computational Linguistics A Prompt-independent and Interpretable Automated Essay Scoring Method for Chinese Second Language Writing Yupei Wang School of Science Beijing Jiaotong University yepiwong@gmail.com Renfen Hu�∗ Institute of Chinese Information Processing Beijing Normal University irishu@mail.bnu.edu.cn Abstract With the increasing popularity of learning Chinese as a second language (L2), the development of an automatic essay scoring (AES) method specially for Chinese L2 essays has become an important task. To build a robust model that could easily adapt to prompt changes, we propose 90 linguistic features with consideration of both language complexity and correctness, and introduce the Ordinal Logistic Regression model that explicitly combines these linguistic features and low- level textual representations. Our model obtains a high QWK of 0.714, a low RMSE of 1.516 and a considerable Pearson correlation of 0.734. With a simple linear model, we further analyze the contribution of the linguistic features to score prediction, revealing the model’s interpretability and its potential to give writing feedback to users. This work provides insights and establishes a solid baseline for Chinese L2 AES studies. 1 Introduction Automatic Essay Scoring(AES) is one of the most important Natural Language Processing (NLP) ap- plications in the field of education (Page, 1966; Ke and Ng, 2019), and has been widely used in stan- dardized language tests (Burstein and Chodorow, 1999; Attali and Burstein, 2006). However, existing works mainly focus on the scoring of English essays (Yannakoudakis et al., 2011; Taghipour and Ng, 2016; Alikaniotis et al., 2016) or Chinese essays by native speakers and minority learners (Chang and Lee, 2009; Peng et al., 2010; Song et al., 2020). Although Chinese second language (L2) acquisition has enjoyed an increasing boom in recent decades, the AES system designed for Chinese L2 writing has received much less attention. Meanwhile, existing AES methods face two important challenges. Firstly, the scoring models are mostly built in a prompt-dependent style, i.e. training and testing for each specific prompt. It requires to collect prompt-specific data, yielding great costs in dataset construction (Attali and Burstein, 2006). Besides, the built models are of weak generalization capabilities and cannot be used to score essays of other prompts. Secondly, although neural network methods have achieved great success in NLP tasks, the gains in neural AES systems are far from being satisfactory. For example, Mayfield and Black (2020) find that fine-tuning BERT produces similar performance to classical models at significant additional cost. Apart from the costs, the deep neural models are also weak in interpretability of the results. However, it is a very important property for AES users who expect to get feedback on the writing (Woods et al., 2017; Ke et al., 2018; Ke and Ng, 2019), not just a score. To solve the above problems, this paper proposes a prompt-independent and interpretable AES method for Chinese L2 writing. Specifically, we build prompt-independent models that could make full use of L2 writing data, and make predictions without the prompt limitations. For interpretability considerations, we extract 90 linguistic indices on accounting of the usage of characters, …

Content truncated to 500 words