We introduce a regression-based framework to explore the dependence that global features have on score predictions from pretrained transformer-based language models used for Automated Essay Scoring (AES). We demonstrate that neural networks use approximations of rubric-relevant global features to determine a score prediction. By considering linear models on the hidden states, we can approximate global features and measure their importance to score predictions. This study uses DeBERTa models trained on overall scores and trait-level scores to demonstrate this framework with a specific focus on convention errors, which are errors in the use of language, encompassing spelling, grammar, and punctuation errors. This introduces a new form of explainability and provides evidence of validity for Language Model based AES.

Citation Key: ormerod2022a