Pengembangan Fitur Rekomendasi Rumusan Masalah Penelitian Menggunakan Pencarian Semantik

Abstract

Pemerintah menetapkan rumusan masalah strategis pada delapan bidang prioritas nasional melalui website pemetaan riset berdampak. Namun, proses pencocokan judul penelitian dengan rumusan masalah masih dilakukan secara manual dan kurang efektif, terutama jika hanya mengandalkan pencarian berbasis kata kunci. Penelitian ini bertujuan mengembangkan fitur rekomendasi rumusan masalah berbasis pencarian semantik agar hasil rekomendasi lebih relevan dan kontekstual. Sistem menggunakan model paraphrase-multilingual-MiniLM-L12-v2 untuk menghasilkan vector embedding dan ChromaDB sebagai basis data vektor. Data penelitian terdiri dari 10 statements pada setiap kategori rumusan masalah. Sistem juga menyediakan fitur pengelolaan kategori dan rumusan masalah pada dashboard admin serta fitur rekomendasi berdasarkan kesamaan makna. Hasil pengujian Black Box pada 33 skenario menunjukkan keberhasilan 100%, sementara evaluasi on topic rate meningkat dari 63,8% menjadi 83,3% setelah penambahan konteks kategori dan mekanisme reranking menggunakan model jinaai/jina-reranker-v2-base-multilingual.

The government has established strategic problem statements across eight national priority sectors through the impactful research mapping website. However, the process of matching research titles with problem statements is still performed manually and is considered less effective, particularly when relying solely on keyword-based search. This study aims to develop a research problem statement recommendation feature based on semantic search to provide more relevant and contextual recommendations. The system utilizes the paraphrase-multilingual-MiniLM-L12-v2 model to generate vector embeddings and ChromaDB as the vector database. The research data consist of 10 statements for each problem statement category. The system also provides features for managing categories and problem statements through an admin dashboard, as well as recommendation features based on semantic similarity. Black box testing on 33 scenarios achieved a 100% success rate, while evaluation using the on-topic rate increased from 63.8% to 83.3% after optimization through category context enhancement and a reranking mechanism using the jinaai/jina-reranker-v2-base-multilingual model.