<![CDATA[The preparation of descriptive narratives for BPS statistical publications is still performed manually, requiring time and potentially causing errors in numerical interpretation. This study aims to develop PRANATA (Statistical Narrative Automation Portal) to automate statistical narrative generation and integrate it with data visualization in a single dashboard using a Large Language Model (LLM) with Retrieval-Augmented Generation (RAG) and Chain-of-Thought (CoT). The study adopts the CRISP-ML(Q) framework to evaluate Gemini 3 Flash, GPT-OSS (120B), and Llama 3.3 (70B) using 11 statistical indicators from BPS Pematangsiantar Municipality. Gemini 3 Flash achieved the best factuality performance, with Macro Precision of 94.96%, Macro Recall of 96.44%, Macro F1-Score of 95.47%, and Macro Hallucination Rate of 4.88%. It also achieved Fluency and Completeness scores of 5.69 and 5.60 out of 7. PSSUQ with five BPS employees yielded an overall score of 6.00 out of 7. These results indicate that PRANATA supports integrated statistical information dissemination.]]>
Copyrights © 2026
