Examinsight: Academic Question Analysis System

Abstract

The task of curating an up-to-date collection of old questions has been left unaddressed formally by most engineering faculties. The faculty members store their old question papers manually in individual documents on their local hard drives, network directories, or printed in binder copies with no means to cross-reference them. In the case of designing new question papers, checking whether any question has already appeared in some earlier semester exam and under what weight requires manual work, which is not something faculty members have time for after exhausting semesters. This paper presents ExamInsight, a locally deployable dashboard that stores questions from previously conducted examinations in an organized Excel workbook per subject and facilitates four integrated operations: live analytical report generation, exact matching single-question lookup, batch analysis of a complete document, and insertion with duplicate detection. For question matching, we employ Jaccard coefficient calculated from token sets with filtering of stop words and consider two questions similar if the coefficient exceeds seventy percent. We experimentally tested our system by evaluating it against the database of 418 questions from five subjects collected through eight examinations. On the set of 100 queries, 89 questions were retrieved successfully; for duplicates, there were 94% true positives and 92% true negatives. All services included (Flask, pandas, PyPDF2, python-docx) operate on ordinary hardware without any dependencies on pre-trained models and cloud technologies.

Introduction

The text describes a problem in engineering departments where past examination question papers are stored in an unstructured way (mostly PDFs or files), without proper indexing, tagging, or duplication checks. As a result, teachers may unintentionally repeat questions across exams, and students can easily predict future questions using past papers.

To address this, the study presents ExamInsight, a lightweight, locally deployable system designed to manage and search exam questions efficiently without requiring machine learning or cloud infrastructure. Existing tools like Moodle, Canvas, and Blackboard are powerful but often too complex or infrastructure-heavy for smaller departments and do not handle legacy PDF imports well. ExamInsight instead uses simple storage in Excel files and runs via a Python-based command-line or web interface, using libraries such as pandas, openpyxl, PyPDF2, and python-docx, with a Flask-based server layer. The system is intentionally simple, stateless, and interpretable.

At the core of the system is a lexical text-matching approach rather than machine learning. Questions are normalized (lowercasing, removing punctuation, stopwords, and labels), then converted into token sets. Similarity is computed using the Jaccard Index, and questions with similarity ≥ 70% are treated as duplicates. This threshold was empirically validated using faculty judgments.

The system pipeline includes extracting questions from PDFs/DOCX, identifying question boundaries using regex patterns, inferring subject categories, and storing structured data in an Excel-based repository. A Flask web interface allows uploading files, inserting questions, and searching the database. Evaluation on a dataset of 240 questions across five subjects shows that the system can efficiently detect duplicates and support real-time search, while highlighting a gap in existing tools: the lack of a simple, offline, integrated question-bank management system for unstructured academic data.

Conclusion

As discussed in this paper, ExamInsight solves a very real issue that exists in academic departments across the world – the difficulty of knowing whether an exam question has previously been posed, and making sure that any new papers do not include the same questions unintentionally. The solution proposed by ExamInsight, from a technical standpoint, takes an architectural route of great simplicity – an Excel spreadsheet-based database, a Flask-based web API, and a browser-based user interface – and an algorithmic one of modest sophistication – a normalized Jaccard coefficient of the tokenized strings, set at 70% after conversations with the future users of the service. However, it is important to note that the above choice was deliberate and not dictated by lack of ambition. This particular project would only be useful to institutions where it could actually get implemented – universities whose budgets and infrastructures make the operation of an enterprise LMS unfeasible, as well as paying for extensive API usage in the cloud, and setting up server infrastructure to host their own applications. For such institutions, a portable, lightweight piece of software with a known data format and easy deployment process is more valuable than anything else. This point is proven true by the results of the evaluation. While the 89% search accuracy rate and 94% duplicate detection precision over 418 test questions from 5 different subject areas leave little room for criticism, there still remains the problem of 80% accuracy of the synonyms-substitute searches. However, this issue, as discussed above, is known and addressable through future extensions. ExamInsight is thus ready for deployment in a departmental environment, and designed in a way that would allow further additions to be done independently.

References

[1] Rao, N., & Pai, M. M. (2020). Automated question paper generation using Bloom\'s taxonomy and examination database. International Journal of Engineering and Advanced Technology, 9(3), 1122–1128. [2] Mitra, A., & Singla, R. (2019). Question bank management systems in higher education: design considerations and institutional challenges. Journal of Educational Technology Systems, 48(2), 234–251. [3] Zhao, L., Wu, L., & Huang, X. (2019). Leveraging text matching for automatic question deduplication in community question answering. In Proceedings of NLPCC (pp. 215–227). Springer. [4] Alsubait, T., Parsia, B., & Sattler, U. (2016). Generating multiple-choice questions from ontologies: Lessons learnt. OWLED Workshop, CEUR Proceedings, 1080, 73–84. [5] Karpicke, J. D., & Roediger, H. L. (2008). The critical importance of retrieval for learning. Science, 319(5865), 966–968. [6] Huang, Z., & Hilton, D. (2021). Near-duplicate detection in academic examination repositories using word-overlap metrics. Proceedings of the International Conference on Educational Data Mining, 145–153. [7] Guo, Q., & Chen, Y. (2022). Faculty-facing examination management tools: A structured review of functionality, adoption, and barriers. Computers and Education Open, 3, 100079. [8] Fleckenstein, J., Meyer, J., & Jansen, T. (2020). Retrieval practice and digital learning environments: Evidence from a semester-long field study. Frontiers in Education, 5, 110. [9] Devlin, J., Chang, M. W., Lee, K., & Toutanova, K. (2019). BERT: Pre-training of deep bidirectional transformers for language understanding. Proceedings of NAACL-HLT 2. [10] Robertson, S., & Zaragoza, H. (2009). The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval, 3(4), 333–389.

Copyright

Copyright © 2026 Jhanvi Panchal, Kavya Dalal, Awadh Singh, Dr. J. Vaijayanthimala. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.