ARGUE-TIRA: attention-based trait interaction reasoning for holistic and analytic automated essay scoring

Automated Essay Scoring (AES) has emerged as an important application of natural language processing providing scalable and consistent evaluation of written responses. These assessments can be conducted using holistic or analytic evaluation approaches. Recent studies have heavily focused on utilizing deep learning models and particularly transformer-based models in holistic essay scoring, while accurately modeling the multi-dimensional analytical scoring remains a challenging problem. Despite their strong performance in holistic essay scoring, most existing approaches work as black-box predictors and do not explicitly model the reasoning process underlying analytic rubric-based evaluation. In this paper, we propose ARGUE-TIRA, a multi-task framework for joint holistic and analytic essay scoring. The proposed framework first predicts multiple analytic writing traits and subsequently applies a Trait Interaction Reasoning Attention (TIRA) to explicitly model dependencies among rubric dimensions before generating the final holistic score. Unlike conventional multi-task AES approaches that independently predict analytic traits, the proposed reasoning mechanism captures inter-trait relationships through structured self-attention reasoning, improving both scoring consistency and interpretability. Experiments conducted on the ASAP and ASAP++ datasets demonstrate that ARGUE-TIRA achieves competitive holistic performance and strong analytic trait-scoring performance. The framework obtains an average holistic QWK of 0.762 on ASAP and the highest average analytic QWK of 0.7424 on ASAP++. Furthermore, it achieves the best performance in seven of the eight evaluated analytic traits, including content, organization, word choice, sentence fluency, conventions, prompt adherence, and language quality.