AASE: AI-Driven Automated Answer Script Evaluation

Abstract In today’s educational systems, evaluating student answerscripts are challenging due to the varied grading criteria, different types of questions, and different ways of attempting the same question. Traditional manual grading can often be inconsistent, inefficient, and sometimes prone to bias, making it difficult to ensure fairness in assessments. These challenges are further complicated when dealing with different types of answers, such as written responses in English, mathematical solutions, etc. Each of these requires distinct approaches, increasing the workload and chances of errors in manual grading. To address these challenges, we propose an Automated Answer Script Evaluation (AASE) system. This solution leverages advanced Natural Language Processing (NLP) techniques along with mathematical parsing algorithms to automate the grading process comprehensively. The proposed AASE system is trained on a diverse dataset containing various grading criteria and incorporates multiple techniques for evaluating both English-based and mathematical answers. For English responses, the system employs different approaches like keyword matching and the Word Movers Distance (WMD) algorithm, a BERT model with additional layer and BERTbased model with dropout layers for sequence classification. These models ensure accurate and unbiased evaluations of student answers in English. Additionally, the AASE system integrates optical character recognition (OCR) technology to recognize handwritten mathematical expressions, which are then converted into LaTeX format using an encoder-decoder architecture. The converted expressions undergo evaluation providing flexibility in assessing both direct answers and detailed step-wise solutions. The AASE system is trained and tested on real-world datasets, and has achieved an accuracy of 80.45% for Subjective and 76% for Mathematical answers evaluation.