Scholarlink

<p>For large student enrollments that are characteristic in large education institutes, the outcome publication process of examinations is typically slow, as grading all answer sheets manually is time-consuming. Automatic grading of descriptive answers can greatly improve outcome publication speed as well as scalability. This paper proposes the use of a fair and uniform AI-based grading system that integrates Optical Character Recognition (OCR) and Large Language Models (LLMs). The proposed system involves the use of EasyOCR to read and extract written student answers, thereafter natural language processing techniques are applied to prepare the data to be fed to an LLM. Grading is accomplished through the use of a fine-tuned Robustly Optimized BERT Pretraining Approach (RoBERTa) model, while constructive feedback to inform students on how their answers can be improved is possible through the use of a fine-tuned Text-to-Text Transfer Transformer (T5) model. The RoBERTa-based grading mechanism has a Weighted Root Mean Squared Error (wRMSE) which is on-par with much larger, computationally intensive LLMs like GPT-4o and GPT-3.5-turbo. The feedback module based on a finetuned T5 model allows for rich, in-context explanations, thus enhancing interpretability for students. Through integration with OCR and LLM methodologies, the proposed system significantly enhances the scalability as well as efficiency in grading. All these steps reduce human bias, promote more standardized testing, and are reasonable moves towards fair, AI-based educational assessment of descriptive answers.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Automated Grading of Descriptive Answers Using AI and Large Language Models

  • 期刊论文

  • 发表日期：

    2026年7月30日

查看全文
  • Aswin S. Kumar,
  • Niranjan Krishnan,
  • Vighnesh Mudaliar,
  • S. RanjithKumar

摘要

For large student enrollments that are characteristic in large education institutes, the outcome publication process of examinations is typically slow, as grading all answer sheets manually is time-consuming. Automatic grading of descriptive answers can greatly improve outcome publication speed as well as scalability. This paper proposes the use of a fair and uniform AI-based grading system that integrates Optical Character Recognition (OCR) and Large Language Models (LLMs). The proposed system involves the use of EasyOCR to read and extract written student answers, thereafter natural language processing techniques are applied to prepare the data to be fed to an LLM. Grading is accomplished through the use of a fine-tuned Robustly Optimized BERT Pretraining Approach (RoBERTa) model, while constructive feedback to inform students on how their answers can be improved is possible through the use of a fine-tuned Text-to-Text Transfer Transformer (T5) model. The RoBERTa-based grading mechanism has a Weighted Root Mean Squared Error (wRMSE) which is on-par with much larger, computationally intensive LLMs like GPT-4o and GPT-3.5-turbo. The feedback module based on a finetuned T5 model allows for rich, in-context explanations, thus enhancing interpretability for students. Through integration with OCR and LLM methodologies, the proposed system significantly enhances the scalability as well as efficiency in grading. All these steps reduce human bias, promote more standardized testing, and are reasonable moves towards fair, AI-based educational assessment of descriptive answers.