COMPARATIVE PERFORMANCE OF TRANSFORMER AND LSTM MODELS FOR INDONESIAN INFORMATION RETRIEVAL WITH INDOBERT

Neural network-based Information Retrieval (IR), particularly with Transformer models, has gained prominence in information search technology. However, the application of this technology in Indonesian, a low-resource language, remains limited. This study aims to compare the performance of the LSTM model and IndoBERT for IR tasks in Indonesian. The dataset consists of 5,000 query–document pairs collected via scraping from three Indonesian news portals: CNN Indonesia, Kompas, and Detik. Evaluation was performed using MAP, MRR, Precision@5, and Recall@5 metrics. The results show that IndoBERT outperforms LSTM in all metrics with a MAP of 0.82 and MRR of 0.84, while LSTM only reached a MAP of 0.63 and MRR of 0.65. These findings confirm that Transformer models like IndoBERT are more effective at capturing semantic relevance between queries and documents, even with limited datasets.