Optimized BERT Model for Question Answering System on Mobile Platform

In Natural Language Processing (NLP), question-answering systems are a classic problem, yet they pose several open challenges. After the pandemic situation of COVID-19, online learning has become crucial, in which question-answering systems are also beneficial to students in searching for answers to subject-related questions. The paper aims to address the challenge of providing a lightweight and efficient question-answering system, especially for students with limited resources, particularly those using low-cost mobile devices. The strategy involves post-training quantization to reduce the size of the BERT model while maintaining high accuracy. Quantization is a technique that reduces the memory footprint of deep learning models, making them more suitable for deployment on mobile devices and low-cost hardware. The reduction in the BERT model size from 438 MB to 181 MB with negligible accuracy degradation is a significant achievement. The resulting mobile application works offline, making educational content accessible in areas with limited internet connectivity. The system achieved an F1 score of 0.87 with negligible degradation compared to the accuracy of the BERT model F1 score: 0.90. Key considerations include user experience, maintenance and memory optimization. This paper offers a practical and cost-effective solution to enhance educational access on mobiles while addressing resource constraints.