Implementasi Named Entity Recognition (NER) berbasis Indobert untuk Question Answering System pada Cerita Rakyat Daerah Karo

Abstract

Folktales are a cultural heritage that must be preserved, yet they are becoming less appealing to the younger generation due to the lack of engaging media. Therefore, this study aims to develop a Question Answering (QA) system based on Named Entity Recognition (NER) using IndoBERT to enhance users' understanding of Karo folktales. The system is designed to recognize entities such as character names, places, and important objects, utilizing a Transformer-based approach. This study employs 15 Karo folktales in the knowledge base, consisting of 216 question-answer pairs. The dataset is divided into 80% for training data and 20% for validation data, while 150 user-generated questions are used as testing data. The IndoBERT-NER model achieves an accuracy of 94%, with precision of 94%, recall of 94%, and an F1-score of 93%. Testing was also conducted on a Telegram bot integrated with the QA system using IndoBERT-NER. The results indicate that the system can provide accurate answers with an average response time of 0.8 seconds. However, some classification errors were found, particularly in recognizing highly similar entities, such as between character names and place names in different stories. These findings suggest that the system is effective in facilitating question-answering for Karo folktales. Further improvements can be made through dataset expansion, deep learning model optimization, and hybrid method integration to enhance accuracy and system efficiency.