Abstract:
AutoGradePro is an automated grading platform developed to reduce educator’s
workload in short-answer assessment evaluation in educational institutions. This
full-stack web application addresses the limitations in manual grading, particularly
excessive time consumption and scalability issues, by leveraging large language
models and grading algorithms. The platform is built using Django for the
backend, Next.js for the frontend, PostgreSQL database, and Ollama LLM for semantic
evaluation. It supports four grading types: exact matching for single-word
answers, semantic similarity assessment for short-phrase answers using Ollama’s
qwen2.5:1.5b model, comparison algorithms for ordered and unordered list-type
answers, and predefined range validation for numerical answers. The system processes
multiple document formats (TXT, PDF, DOCX, XLSX) through a fivestage
pipeline consisting of file upload, format detection, text extraction, answer
parsing, and automated grading with report generation. The system can process
multiple student answer files simultaneously, which decreases the manual evaluation
time and increases efficiency when grading large batches of assessments.
Performance evaluation shows high accuracy across objective types: 100% for
single-word answers, 92.9% for list-type answers, 84.6% for numerical responses,
and 73.3% for short-phrase responses. When grading 20 student files containing
400 questions, average backend processing time was 307.82ms for single-word
questions, 601.75ms for list-type questions, and 840.88ms for numerical questions.
Compared to manual grading, the system provides a substantial reduction in
evaluation time for objective questions and delivers an approximately fourfold
increase in processing speed for short-phrase responses. Key features include automated
marking scheme parsing, analytics dashboards and spreadsheet results
export capabilities. AutoGradePro can reduce repetitive grading tasks while allowing
educators to focus on high-value instructional activities.
