Automated Evaluation of English Essays According to Predefined Checking Criteria

This work focuses on the development of an automated essay scoring system based on modern machine learning models, designed to assess texts according to multiple predefined criteria. The relevance of this topic lies in the growing demand for objective and efficient evaluation of written responses, particularly in high-stakes language proficiency exams, where essays are a mandatory component. The study provides a comprehensive analysis of current approaches to automated essay scoring, examining various neural network architectures, including pretrained transformer models from the BERT family and generative language models. The findings indicate that BERT-based regression models achieve the highest accuracy while maintaining low computational costs. Although generative models require significantly more resources and time for inference, they offer unique capabilities for generating detailed and contextually relevant feedback, making them a promising tool for explainable AI in education. The experimental part of the research involved training and testing models on a dataset of student essays written in response to Task 37 of the Unified State Exam (USE) in English. Special attention was given to the challenges of class imbalance and high topical variability within the dataset, both of which were addressed in the data preparation and model training processes. The developed system demonstrated strong predictive performance and practical viability for educational applications. The results of the study suggest that multi-criteria automated essay evaluation can be effectively implemented using compact, fast-performing models that deliver scoring quality comparable to that of human raters. The proposed system may be deployed as a standalone web service or integrated into existing educational platforms, offering substantial benefits such as reduced teacher workload, faster feedback delivery, and improved assessment objectivity.