815停電事件對總統滿意度的影響：以文字探勘為途徑

The Influence of 815 Power Failure on The President Approval Rating: A Text-mining Approach

林玿弘(Shao-Hong Lin)

指導教授 ： 張佑宗

國立臺灣大學/社會科學院/政治學研究所/碩士(2019年)

https://doi.org/10.6342/NTU201901625

摘要

總統滿意度作為總統施政表現的重要指標，在現行以民意調查為主的測量方式中，受限於經費、時間等因素，執行的時間點較為分散，難以回答總統滿意度受事件影響多大、持續多久等問題。因此本文利用網路資料，嘗試發展新的測量方式，改善資料點不足而難以評估事件對總統滿意度影響的問題。
本文利用機器學習中的半監督式學習，標註2017年PTT八卦版中20餘萬個帳號的發言是否滿意蔡英文總統的施政，進而獲得較密集時間點的總統滿意度。藉此，我們可以得知不同時間、不同議題中，使用者的態度形成與變化。研究結果顯示，本文之測量方式與既有文獻結果相符，加上時間設定上較為自由，本測量方式將有助於短時間、即時性的事件影響評估。
本文以729限電事件與815停電事件為例，利用自然實驗法將本島與外島使用者分離，並討論受影響大小對總統滿意度的影響。此類突發事件，在過去的研究中，缺乏前測資料，難以對事件歸因，也缺乏資源對事件持續追蹤，進而難以得知事件的影響大小、持續時間等問題。然而本文明確指出，受影響較大的本島使用者較未受影響的外島使用者更傾向給予總統不滿意的評價，更可精準地評估事件影響的持續時間。

關鍵字

總統滿意度 ； 文字探勘 ； 機器學習 ； 課責

並列摘要

The presidential approval rating is an important indicator to judge the performance of the president. We have measured the president approval rating by survey. But limited by some factors such as budget or time, the data points in time are dispersion. Due to this dispersion, we are difficult to answer how the event affects the president approval rating. To solve the above trouble, this paper tries to advance a new measurement by internet data.
This paper will label whether user approval the president or not by its post or comment in PTT Gossiping Board during 2017. And this measurement is based on a semi-supervised machine-learning approach. Through this measurement, we can obtain more intensive data on time. That is why this paper can answer the user’s attitude changing and forming in a different time and on a different issue. This research points out that the performance of this measurement is fitting the past researches. More advantageous is that this measurement is free from the time setting, so is powerful to evaluate the short-term and the immediacy event affection.
This paper pick 729 power limited and 815 power failure in 2017 as examples. We distinguished the user about inner and outer Taiwan island. Through the natural experiment, this paper discusses whether the user changes its attitude after it is effected. In the past, because of lacking the pre-test data, we had difficult to estimate the event influence. We also had difficulty to tracking event affection. So we are difficult to know the sphere and the duration of the event influence in the past. This paper points out that the inner user which be effected more will more tend to give disapproval attitude than outer and without influence. Moreover, this paper can estimate the influence of event duration more exactly.

並列關鍵字

The President Approval Rating ； Text-mining ； Machine-Learning ； Accountability

參考文獻

Barry C. Burden and Anthony Mughan. 2003. “The International Economy and Presidential Approval”. The Public Opinion Quarterly. 67(4):555-578.

Brandice Canes-Wrone. 2004. “The Public Presidency, Personal Approval Ratings, and Policy Making”. Presidential Studies Quarterly. 34(3): 472-492.

Brian J. Gaines. 2002. “Where's the Rally? Approval and Trust of the President, Cabinet, Congress, and Government Since September 11”. Political Science and Politics. 35(3): 530-536.

Christopher Achen and Larry Bartels. 2016. Democracy for Realists. NJ, Princeton: Princeton University Press.

David W. Nickerson and Todd Rogers. 2014. “Political Campaigns and Big Data”. The Journal of Economic Perspectives. 28(2):51-73.