應用文件探勘技術進行媒體框架自動化分析

Automatic Content Analysis of Media Framing by Text Mining Techniques

郝德(De Hao)

指導教授 ： 林福仁

國立清華大學/科技管理學院/服務科學研究所/碩士(2014年)

摘要

台灣的政治新聞在各平台都具有相當大的影響力，然而新聞在產生的過程中
會因問記者和媒體公司個人的政治傾向和個人意見而進行某程度的撰稿選擇或偏頗，
重要的訊息會因此而消失。
台灣的立法院圖書館有提供完整而詳細的會議記錄，包含逐字稿與影片，雖
說這樣詳細的資料有被記錄著，這樣的開放資料有助於一般民眾與學者去研究與追
蹤立法院內部的詳細情況，也有助於減少因為僅接收大眾媒體而失去重義的議題，
但其資訊量過於龐大而無法輕易的以人力的方式去追蹤與研究，為了縮減民眾與立
法院內部的距離，此研究提出應用文件探勘技術於立法院文本與新聞文本進行媒體
框架之分析。
此研究嘗試設計一個以最低限度的人為參予的媒體框架系統，過去多數的研
究需要專家因應特定的議題而預先定義每個框架的代表詞彙，這樣的過程本身也是
一種偏見，所以我們在過程中除了少量的停用詞需要專家預先定義，剩餘的SVD、
分群與LDA 皆無人為的參予。
此系統的數據結果被專家與一般民眾評估後得知確實具備著協助專家計算媒
體框架的實證與發現不曾預料的媒體框架。一般民眾與專家可使用此系統於追蹤及
監控各新聞媒體對於事件的媒體框架現象。

關鍵字

文字探勘 ； 媒體框架 ； 立法 ； 分群 ； 奇異值分解 (SVD) ； 隱含狄利克雷分 配 (LDA)

並列摘要

Political-related news is one of the most popular topics in different media platforms,
such as TV channels, Internet news and newspapers in Taiwan. When news are produced
through a process of selection and rephrase by reporters and media firms, reporters’
personal political leaning and personal opinions may influence the process, important
messages may inevitably loss.
In Taiwan, the Parliamentary Library of Legislative Yuan website provides detailed
contents about activities happening inside Legislative Yuan, including contents, such as
transcripts and video recordings of interpellation, conference speech, interim proposal
and legislation proposal. Although there is a complete record of information regarding
Legislative Yuan provided online, but the quantity of Legislative documents are far too
much for citizens to make sense of. It is imperative that better organized information
released to the public would facilitate readers to ease the cognitive loads in understanding
what issues have been discussed by legislators and reported by the media. To minimize
the gap between legislative documents and the general public, this study proposes a text
mining mechanism to automatically cluster legislative and news documents by media
frames, and then represents the proportion of each frame on certain sources.
This study aims to develop an automatic clustering system to determine media
frames with the minimum amount of human interference. Domain expert labeling was not
considered to increase classification accuracy, except the stopping words and few minor
variables in the feature selection stage requires human adjustment according to input files,
and the rest of the system are able to process input documents with zero expert labeling
or adjustments.
Experimental results evaluated by domain experts shows the system proposed in
this study is able to assist and provide political domain experts hard evidences of media
framing. People can monitor and discover media framing phenomenon using accurate
numbers and data.

並列關鍵字

Text mining ； media framing ； legislative ； clustering ； Singular Value Decomposition (SVD) ； Latent Dirichlet Allocation (LDA)

參考文獻

Huijboom, N., & Van den Broek, T. (2011). Open data: an international comparison of strategies. European journal of ePractice, 12(1), 1-13.

Moira, W., & Hsieh, Y.-Y. (2014, 2014/03/27). 324: Dispatches from Taipei. from https://nplusonemag.com/online-only/foreign-affairs/324-dispatches-from-taipei/

Abbott, J. P. (2011). Electoral Authoritarianism and the Print Media in Malaysia: Measuring Political Bias and Analyzing Its Cause. Asian Affairs: An American Review, 38(1), 1-38. doi: 10.1080/00927678.2010.520575

Allan, J. (2002). Introduction to topic detection and tracking Topic detection and tracking (pp. 1-16): Springer.

Allen, B., O'Loughlin, P., Jasperson, A., & Sullivan, J. L. (1994). The media and the Gulf War: Framing, priming, and the spiral of silence. Polity, 255-284.