Implementasi Sistem Pencarian Semantic Metadata Buku Berbasis Sentence-BERT

Abstract

Book discovery in digital libraries is often hindered by keyword-based methods that fail to capture semantic nuances and user query context, leading to low relevance. This research aims to implement and evaluate a semantic search system for book metadata using the Sentence-BERT (SBERT) model, accelerated with Facebook AI Similarity Search (FAISS), and to compare its performance against a traditional TF-IDF-based baseline system. The system was developed using the "Goodreads Books 100k" dataset. Book metadata (title, author, and description) were converted into vector representations (embeddings) using a pre-trained SBERT model and indexed using FAISS for efficient similarity search. Relevance effectiveness was evaluated using an LLM-as-a-Judge approach with Gemini 2.5 Pro to assess the top 10 results for 20 test queries, using Mean Relevance Score and Precision@10 metrics. Efficiency was measured by average search latency. The results demonstrate the significant superiority of the SBERT system, achieving a Mean Relevance Score of 3.40 and a Precision@10 of 53%, far surpassing the TF-IDF system, which only scored 2.73 and 35%, respectively. In terms of efficiency, the SBERT+FAISS system was also 8.4 times faster (26.40 ms average latency) than TF-IDF (221.29 ms). This study proves that the implementation of SBERT with FAISS not only produces more contextually relevant searches but is also dramatically more efficient, making it a superior solution compared to traditional methods for book metadata search.