A gated multi-scale attention framework for automated short answer grading in Turkish

This study proposes a new deep learning architecture for automated short-answer grading (ASAG) in Turkish. The model integrates transformer-based encoders (BERTurk, RoBERTa-Turkish (RoBERTa), ELECTRA-Turkish (ELECTRA)) with a Gated Multi-Scale Head Attention (GMHA) layer. Additionally, a BiLSTM network and parallel CNN layers are employed to capture both global dependencies and local linguistic patterns.The model was evaluated on a newly created Turkish dataset containing 2200 student responses from 220 learners, each graded by four experts. Among the pretrained Turkish encoders, the BERT-based model with GMHA achieved the best performance (QWK = 0.839, Pearson = 0.856, RMSE = 0.918). The model was also compared with traditional machine learning approaches and large language models (LLMs) using the Leave-One-Question-Out (LOQO) validation protocol. Under this setting, it achieved QWK = 0.337, Pearson = 0.468, and RMSE = 1.213, outperforming all baselines.Ablation studies confirmed the effectiveness of the GMHA and parallel CNN components. These results highlight both the theoretical contribution of multi-scale gated attention and the practical potential of developing reliable automated grading systems for morphologically rich, low-resource languages like Turkish.