Abstract:

Automated Essay Scoring (AES) is an important task in educational technology, which assists us to evaluate written text with consistency and fairness. This paper presents MonarchAES as a hybrid model with RoBERTa semantic representation, Graph Attention Networks (GAT) structural relationship and dual optimisation using Genetic Algorithm (GA) and Butterfly Monarch Optimization (BMO) hyperparameter tuning and feature selection respectively. The model was trained and validated with the ASAP 2.0 dataset. Class imbalance was addressed with weighted random oversampling. The hybrid method shows potential to leverage language learning, graph reasoning and evolutionary optimization for the improvement of fairness, coherence and accuracy in essay scoring.