Natural Language Query Hybrid Architecture for Cross-Sectoral Analytics Dashboards on Public Statistical Data

Abstract

Background: Public statistical data are essential for evidence-based policymaking, yet the complexity of cross-sectoral indicators often limits accessibility for nontechnical users. This challenge is particularly relevant in East Java, where regional development disparities require integrated analytical tools. Objective: This study aims to develop a cross-sectoral analytics dashboard integrating Natural Language Query (NLQ), Apriori association-rule mining, and a supporting local large language model (LLM). Methods: The system was developed using the SDLC Waterfall model and utilized more than 50 Statistics Indonesia (BPS) indicators covering 38 districts/cities in East Java from 2016 to 2026. A deterministic, rule- and regular-expression (regex)-based Text-to-SQL mechanism processed natural-language queries, while Apriori identified cross-sectoral association patterns. A local LLM (Ollama, llama3.2:1b) generated supporting narratives. Evaluation included black-box testing, 50-query intent classification, association-rule metrics, and an exploratory LLM assessment. Results: The NLQ module achieved 96.00% accuracy (48/50 queries). Apriori generated 55,480 high-level rules with an average confidence of 0.909 and lift of 15.44. The exploratory LLM evaluation achieved a composite score of 4.35/5.00, a 2.5% hallucination rate, a 26.07-second P95 latency, and a 100% fallback success rate. Conclusion: The hybrid architecture effectively supports accessible statistical-data retrieval and cross-sectoral pattern discovery, while the LLM remains a supporting narrative component that requires further formal validation.