Data Analytics for Policy Proposal Support

Overview

Researchers unfamiliar with programming languages often face difficulties in performing in-depth analysis for future forecasting or policy support. To improve the utility and efficiency of analyzing vast amounts of raw research reports, there is a need for a service that can autonomously analyze and learn from data to generate optimized results, such as statistics and visualizations. This project aims to demonstrate a service by implementing a GPT-based chatbot capable of providing natural language answers based on the research outputs stored in the NKIS.

Project description

For model training, raw research report data was sourced from the NRC Data Information System (NDIS), the National Knowledge Information System (NKIS), and the Policy Research Information Service Management (PRISM). The technical procedure involved establishing a retrieval engine based on training data and performing fine-tuning on the MaumGPT model. Subsequently, the retrieval engine and the fine-tuned MaumGPT model were integrated using proprietary middleware (Orchestra). This service is specifically designed to support researchers in future forecasting and policy-making by enabling the large language model to autonomously process vast datasets and present optimized analysis results, including statistical data and visual materials.

Key technological features

• LLM Training Dataset: Collection and pre-processing of training data, including the conversion of 10 sets of Q&A pairs using a PDF Annotation Tool
• Retrieval Engine Implementation: Implementation of a search engine using research reports within the NKIS (National Knowledge Information System) as metadata
• GPT Fine-tuning: Fine-tuning of MaumGPT (13B parameters) using research reports from NKIS focused on disaster and safety themes

Expected outcomes

This project successfully demonstrated the creation of a GPT-based chatbot capable of generating answers directly from the original texts of research reports. By utilizing the developer's specialized tools to support workflow creation and UI integration, the project proved the feasibility of a full commercial service launch. Furthermore, the implementation of a retrieval engine effectively addresses the "hallucination" issues common in standard chatbots. By overcoming these existing limitations, the service is expected to become a highly practical and reliable tool for researchers, providing accurate information based on verified policy and research data.

Launch year: 2023

Attribution

Case study courtesy of the National Information Society Agency, Republic of Korea, Apolitical's Content Partner. Sourced from Use Cases of Public AI Service, Vol.2, https://eng.nia.or.kr/site/nia_eng/ex/bbs/View.do?cbIdx=31975&bcIdx=29182&parentSeq=29221