提升地緣政治風險評估：以台灣為應用案例的創新大型語言模型方法

Abstract

傳統以固定、手工編製字典為基礎的地緣政治風險（GPR）指數，往往忽略在地脈絡與混合戰手法的演變。本論文提出並驗證一套由大型語言模型（LLM）輔助的「提示—篩選—計數」工作流程：首先生成 40 份以臺灣為焦點的華語 GPR 字典，將每份字典轉換為週頻指數，再取其中位數作為主指標。依據 2017–2024 年共 330 萬則新聞報導，這種指數組合設計可降低 LLM 輸出隨機性（平均配對相關係數 = 0.70），並維持可解釋性。相較於 Caldara 的全球字典與 Lau 的臺灣改編版本，我們的指數涵蓋更廣泛的風險面向---軍事升級、外交孤立、經濟脅迫、網路／資訊攻擊與國內政治。事件研究顯示，本指數能捕捉傳統指數遺漏的臺灣特定衝擊：經濟排斥（臺灣被排除於 IPEF）、北京主導的邦交國轉向（宏都拉斯與諾魯）、區域安全震盪（南韓 2024 戒嚴風波），以及領導人更迭不確定性（習近平三連任與 2024 年美國總統大選）。來源層級分析發現《自由時報》波動劇烈，《中國時報》輸出較穩定卻呈現內容農場現象，提醒研究者在計算新聞占比型指數時須慎防分母偏誤。七變數季頻結構向量自迴歸（SVAR）結果顯示，臺灣經濟韌性十足：一個兩倍標準差的 GPR 衝擊後，投資增加、工時延長、股價上升，而隱含波動率僅短暫下跌。此方法成本低、人工校對需求少、可解釋度高，可輕鬆擴充至其他議題，例如跨國或台積電專屬的風險指數。

Geopolitical risk (GPR) indices based on fixed, handcrafted dictionaries often miss local context and evolving hybrid-warfare tactics. This thesis introduces an LLM-assisted “prompt–filter–count” workflow that generates 40 Mandarin, Taiwan-specific GPR dictionaries, converts each into a weekly series, and takes their median as the headline index. Using 3.3 million news articles (2017–2024), the ensemble mitigates LLM output randomness (mean pair-wise correlation = 0.70) while preserving transparency. The median index captures a wider range of risks—military escalation, diplomatic isolation, economic coercion, cyber/information attacks, and internal politics---than Caldara's global list and is more balanced than Lau's Taiwan adaptation. Event-study tests show the index flags Taiwan-specific shocks that legacy lists miss: economic exclusion (Taiwan's IPEF omission), Beijing-driven ally switching (Honduras \& Nauru), regional-security shake-ups (South Korea's 2024 martial-law scare), and leadership-transition uncertainty around Xi's third term and the 2024 U.S. election. Source-level analysis reveals sharp contrasts: Liberty Times shows episodic surges, China Times steadier output but signs of content farming, cautioning against naïve denominator use in news-share indices. A seven-variable quarterly SVAR indicates Taiwan's economy is resilient: a two-s.d. GPR shock precedes higher investment, longer work hours, rising equity prices, and a short-lived dip in implied volatility. The method's low cost, minimal curation, and interpretability make it adaptable to other contexts, such as TSMC-specific or cross-country risk indices.