Full-stack web development for auto-assessment platform

• Main Author: Cholakov Kristiyan Kamenov
• Other Authors: Loke Yuan Ren
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Computer and Information Science; Auto-assessment; Handwritting recognition; Full-stack platform
• Online Access: https://hdl.handle.net/10356/181183
• Institution: Nanyang Technological University

This project highlights the limitations of the existing digital evaluation system and identifies the lack of a complete self-sustained platform for conducting comprehensive handwritten examinations on a digital platform that does not involve the usage of external software or hardware tools. The paper presents a novel end-to-end architecture that stands out from the competitors with its ability to integrate every stage of the examination, while previous approaches provide partial solutions that aim to speed only parts of the whole process. The project represents the first platform with its own exam management console, handwritten input capturing, image-to-text recognition for pages with handwritten content, and LLM-based grading assistant. The individual components of the proposed solution integrate the current state-of-the-art models for each stage of digital conversion and automated assessment. These include PosFormer for HMER, TrOCR for HTW, DETR for layout analysis, and Qwen2.5-Math for autonomous grading of exams in the STEM domain. Furthermore, the project included the development of the first dataset for page layout analysis between the two classes of handwritten mathematical expressions and handwritten text. The custom dataset contains 10,000 training data points and 2,000 evaluation ones, holding the ground truths for the bounding boxes of the regions and the recognizable content in them. Then, the new dataset was used for training the first document layout analysis model to successfully distinguish between handwritten text and formulas on a blank A4-size digital page. The model, based on the state-of-the-art architecture of DETR, achieves F1 scores of 0.921 and 0.916 for handwritten text and formulas, respectively. The whole recognition pipeline with TrOCR and PosFormer used for the further HMER and HTW achieves a combined F1 score of 82.19%, demonstrating its enhanced effectiveness in accurately recognizing handwritten content. The modality and scalability of the proposed platform make it adaptable to changes like new state-of-the-art models, which can be easily swapped by just maintaining the RESTful API’s structural functionality and routes. The project represents an agile and modular future-proof system, independent of external tools and services, which sets the basis for a seamless transition from traditional pen-and-paper exams to digital handwritten platforms.