Aristya Vika Wijaya, Shintami Chusnul Hidayati, Ratih Nur Esti Anggraini, Fatma Said Abousaleh, Yeni Anistyasari
Wikipedia offers a vast and valuable repository of knowledge; however, its unstructured and extensive content presents challenges for users seeking accurate and efficient information retrieval. While large language models (LLMs) have demonstrated strong performance in natural language understanding, their deployment often demands considerable computational resources. This study proposes an efficient and interpretable question answering (QA) pipeline by augmenting the pre-trained BERT model with Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) layers. In this architecture, CNN captures local semantic features, while LSTM models long-range contextual dependencies within the text. The system is evaluated on the Stanford Question Answering Dataset (SQuAD), using F1 Score and Exact Match (EM) as evaluation metrics. Experimental results show that the proposed model outperforms baseline configurations, achieving an F1 score of 0.827 and an EM score of 0.662, with a training time of 16,523 seconds. These findings highlight the model's potential in offering a practical balance between performance and interpretability for real-world QA applications. © 2025 IEEE.
Institut Teknologi Sepuluh Nopember, Department of Informatics, Surabaya, Indonesia; Suez University, Department of Artificial Intelligence, Suez, Egypt; Universitas Negeri Surabaya, Department of Informatics, Surabaya, Indonesia
