Year: 2025
Type: conference-paper
Abstract: We present our system for TAQEEM 2025 Task A on Arabic automatic essay scoring.Building on a pretrained Arabic encoder, our work focuses on two key design axes: (i) replacing the standard linear head with a lightweight multi-layer perceptron (MLP) and (ii) optimizing with distribution-aware objectives.We introduce a Weighted Mean-Squared Error loss, which assigns higher weights to less frequent scores to counteract the imbalanced, bellshaped score distribution of the training data.On the official development folds, our system outperforms the baseline on Quadratic Weighted Kappa.Our findings underscore the importance of tailoring objective functions to specific data characteristics for achieving stateof-the-art results in AES.
Authors: Trong-Tai Dam Vu, Thìn Đặng Văn
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037593