All case studies
ResearchEdTech / Assessment

AI-Driven Exam Paper Quality Assessment

NLP that scores question-paper quality against Bloom's taxonomy—98.04% accuracy, Springer ICTIS 2024.

CHARUSAT Research · 2023 – 2024February 20245 min read

Problem & industry context

Institutions struggle to ensure exams balance cognitive levels, difficulty, and syllabus coverage before administration. Manual review is slow and subjective. EdTech research increasingly uses NLP to classify items, detect bias patterns, and align assessments with learning outcomes—if models are explainable to faculty.

Insight

BERT and Bi-LSTM ensembles capture semantic structure better than bag-of-words alone for pedagogical labels. Quality is not one score—it is multi-criteria (difficulty spread, taxonomy coverage, redundancy). Publishing demands rigorous datasets (2,809 labeled instances here) and honest limitation discussion.

What I built

Developed an AI system to categorize and score pre-exam question papers using BERT, LSTM, and Bi-LSTM, reaching 98.04% accuracy. Co-authored publication at Springer ICTIS 2024. Open-sourced implementation for reproducibility and classroom adoption discussions.

Technical approach

Stack and tooling for this work: Python, BERT, LSTM, Bi-LSTM, NLP, GloVe. Topics covered: NLP, BERT, Education, Research.

Topics

NLPBERTEducationResearch