Linguistically Informed Essay Assessment Framework to Analyze Writing Style Vocabulary Usage and Coherence

Automated essay scoring (AES) has become an essential tool in educational technology, yet many existing approaches rely on black-box models that lack interpretability and adaptability across diverse prompts and writing styles. Conventional transformer-based AES systems demonstrate strong accuracy, but often fail to provide pedagogically meaningful feedback or generalize effectively in low-resource settings, limiting their practical applicability. The proposed COSMET-Net (Contrastive and Explainable Semantic Meta-Evaluation Network) addresses these limitations by integrating contrastive learning, meta-learning, and explainable AI to produce an adaptive and interpretable evaluation of academic essays. Essays are processed through text cleaning, tokenization, and lemmatization, and embeddings are generated using pretrained transformers such as BERT and RoBERTa. Contrastive learning distinguishes high- and low-quality essays, while a Contrastive Linguistic Regularization (CLR) layer aligns embeddings with linguistic properties, enhancing interpretability. Meta-learning enables rapid adaptation to novel prompts with minimal additional data. The explainable output module, employing attention visualization and SHAP values, provides detailed feedback on grammar, coherence, vocabulary richness, and readability. The framework was implemented in Python with PyTorch and Hugging Face Transformers and evaluated on the IELTS Writing Scored Essays Dataset. COSMET-Net achieved an accuracy of 92%, a recall of 93%, and an F1-score of 92%, surpassing existing models such as hybrid RoBERTa + linguistic features (F1-score 84%) and discourse + lexical regression (F1-score 88%). These results demonstrate that COSMET-Net delivers highly accurate, flexible, and linguistically interpretable assessments, providing a scalable solution for automated and pedagogically meaningful essay evaluation.