誰在說話？兩岸三地媒體在反送中事件中之消息來源比較研究

Abstract

本研究以2019年香港反送中運動為例，探討兩岸三地媒體在社會運動事件中的消息來源使用模式。研究蒐集中國大陸、香港、臺灣共八家媒體計24,916篇報導，結合大型語言模型自動化標註、內容分析、相似度分析、社會網絡分析等方法，系統性地檢視媒體消息來源的選擇策略與結構特徵。 研究發現，兩岸三地媒體的消息來源使用呈現高度分化格局。《人民日報》對建制派來源偏好強度達+123.0%，《自由時報》對泛民來源則為+79.7%，兩者形成鏡像式的兩極化。值得注意的是，《中國時報》展現出獨特的來源策略，其對中港官方來源偏好強度為+51.1%，且在「中美社群」的使用率上與中國大陸媒體趨近，與其他臺灣媒體形成顯著差異。引用相似度分析亦顯示，《中國時報》與香港親建制媒體《文匯網》、《香港經濟日報》的相似度約為0.56，為臺灣媒體中唯一與香港親建制媒體具有明顯資訊流動特徵之媒體。 本研究最核心的發現在於匿名與模糊來源在新聞網絡中的結構性主導地位。網絡中心性分析顯示，「消息人士」在加權度數、中介中心性、緊密中心性三項指標均居首位，「網友」與「民眾」亦位居前列。這意味著新聞報導的核心樞紐，都為難以驗證身份、缺乏問責性的來源。進一步分析發現，各地媒體發展出差異化的模糊來源策略，香港媒體偏好「消息人士」、《自由時報》大量援引「網友」、中國大陸媒體則偏好「民眾」一詞。此種匿名依賴，將可能性導致新聞可信度與可追溯性構成潛在風險。本文更認為，不同媒體使用的模糊來源差異並非偶然，顯示出其對於報導議題的立場有著策略上的意義。 方法論上，本研究調整ProTeGi文字梯度框架（Pryzant et al., 2023），設計加權相似度配對機制與雙模型架構，使大型語言模型標註達到實證水準，為中文新聞內容的大規模自動化分析提供技術支持。

This study uses the 2019 Hong Kong Anti–Extradition Bill movement to examine how media in Mainland China, Hong Kong, and Taiwan employ news sources when covering social movements. The dataset contains 24,916 articles from eight outlets across the three regions. Combining large language model (LLM)–assisted automated coding with content analysis, similarity analysis, and social network analysis, the study systematically maps source-selection strategies and the structural features of source networks. Findings show a highly polarized source landscape. People’s Daily exhibits a strong preference for pro-establishment sources (+123.0%), while Taiwan’s Liberty Times strongly favors pro-democracy sources (+79.7%), forming a mirrored two-pole pattern. Notably, China Times demonstrates a distinctive strategy: it prefers China–Hong Kong official sources (+51.1%) and resembles Mainland outlets in its use of a “China–US community.” Citation-similarity results further indicate that China Times is unusually close to Hong Kong pro-establishment outlets Wen Wei Po and Hong Kong Economic Journal (similarity ≈ 0.56), making it the only Taiwanese outlet showing clear information-flow characteristics aligned with Hong Kong pro-establishment media. The most central structural finding is the dominance of anonymous or vague sources in the news network. Centrality measures identify “sources familiar with the matter/insiders” as the top hub across weighted degree, betweenness, and closeness; “netizens” and “the public” also rank highly. This implies that core reporting nodes are often difficult to verify and weakly accountable, potentially undermining credibility and traceability. Regions also display differentiated vague-source strategies: Hong Kong outlets favor “insiders,” Liberty Times heavily cites “netizens,” and Mainland outlets prefer the term “the public,” suggesting strategic alignment with editorial stances rather than random variation. Methodologically, the study adapts the ProTeGi text-gradient framework (Pryzant et al., 2023) by introducing a weighted similarity-matching mechanism and a dual-model design, enabling LLM annotations to reach empirical research standards and supporting large-scale automated analysis of Chinese-language news.