Semantic search system for name of government work package using BERT

This study develops a semantic search system to improve query matching for government work packages using data from SIRUP (General Procurement Plan Information System) and KBLI 2020 (Indonesian Standard Industrial Classification). The existing SIRUP search system relies on keyword matching, which often fails to capture semantic relationships between queries and data. To address this, a two-stage approach was introduced: first, matching queries with main KBLI categories, and second, refining results by aligning queries with work package names under the identified category. The system leverages IndoBERT as its base model and incorporates fine-tuning to adapt to the procurement domain. While fine-tuning enhanced semantic relationships, it was insufficient to handle irrelevant data with high similarity scores. To overcome this, Triplet Loss was employed to optimize embeddings, improving the separation of relevant and irrelevant data. The evaluation involved 30 queries, with each query individually assessed for precision, recall, and F1 score. The system achieved an overall precision of 0.84, recall of 0.78, and F1 score of 0.80, demonstrating its ability to retrieve semantically relevant results effectively. This approach highlights the potential of semantic search to enhance query accuracy and relevance in the government procurement domain.