TWClimateChat：結合大型語言模型的社會感知對話系統，用於台灣即時在地化天氣資訊傳遞;TWClimateChat: A Socially-Aware Dialogue System for Real-Time, Localized Weather Intelligence in Taiwan with Large Language Models

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/98516

請使用永久網址來引用或連結此文件: https://ir.lib.ncu.edu.tw/handle/987654321/98516

題名:	TWClimateChat：結合大型語言模型的社會感知對話系統，用於台灣即時在地化天氣資訊傳遞;TWClimateChat: A Socially-Aware Dialogue System for Real-Time, Localized Weather Intelligence in Taiwan with Large Language Models
作者:	許育瑄;Hsu, Yu-Hsuan
貢獻者:	資訊工程學系
關鍵詞:	大型語言模型;檢索增強生成;氣象對話系統;即時天氣預報;地理語意解析;臺灣;LLM;RAG;Meteorological Dialogue System;Real-Time Weather Forecast;GeographicSemantic Parsing;Taiwan
日期:	2025-08-05
上傳時間:	2025-10-17 12:52:36 (UTC+8)
出版者:	國立中央大學
摘要:	當地社會多元的社群中，要進行有效的氣候溝通，需要能結合在地知識、文化脈絡與動態環境資料的系統。我們提出 TWClimateChat，一個專為台灣設計、具有社會意識的氣象對話系統。它結合了大型語言模型（LLMs）與檢索增強生成（RAG）框架，能提供即時的鄉鎮層級天氣預報，並輔以在地諺語與歷史災害事件等文化相關資訊。本系統採用多階段語意解析，能準確理解自然語言查詢中常見的模糊地理與時間表達，進而生成精確的 SQL 查詢並從國家災害防救科技中心（NCDR）及其他權威來源擷取資料。在知識型問題的測試中（包括氣象知識、文化諺語、歷史事件），TWClimateChat 的表現超越多個先進基準模型，包括 GPT‑4o。在即時預報任務中，系統達到 92.59% 的地理解析準確率、90.12% 的時間理解率，以及 85.18% 的整體預報正確率，民眾需求上甚至超越具備網頁瀏覽功能的 GPT‑4o。透過將計算語言理解與社會及環境脈絡相結合，TWClimateChat 展現出智慧型對話系統如何提升公眾參與、情境感知，以及在面對氣候災害時的韌性。本研究為計算社會系統領域做出貢獻，將 AI 驅動的自然語言處理與在地、文化化的災害溝通相結合。 ;Effective climate communication in socially diverse communities requires systems that integrate local knowledge, cultural context, and dynamic environmental data. We present TWClimateChat, a socially-aware meteorological dialogue system designed specifically for Taiwan, which combines large language models (LLMs) with a Retrieval-Augmented Generation (RAG) framework to deliver real-time, township-level weather forecasts alongside culturally relevant information such as local proverbs and historical disaster events. The system employs multi-stage semantic parsing to accurately interpret ambiguous geographic and temporal expressions commonly found in natural language queries, enabling precise SQL query generation and data retrieval from the National Science and Technology Center for Disaster Reduction (NCDR) and other authoritative sources. In evaluations on knowledge‑based questions—including meteorological knowledge, cultural proverb, and historical events—TWClimateChat outperforms multiple advanced baselines, including GPT‑4o. In real‑time forecast tasks, it achieves 92.59% geographic parsing accuracy, 90.12% temporal understanding, and 85.18% overall forecast correctness, surpassing even GPT‑4o with web browsing capability. By integrating computational language understanding with social and environmental context, TWClimateChat exemplifies how intelligent dialogue systems can enhance public engagement, situational awareness, and resilience in the face of climate hazards. This work contributes to the field of computational social systems by bridging AI-driven natural language processing with localized, culturally grounded disaster communication.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	0	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....