以臺灣央行及美國聯邦銀行會議記錄預測台灣總體經濟指標：LDA主題模型及情緒分析的應用

Abstract

本文使用LDA主題模型，將中華民國中央銀行和美國聯邦公開市場委員會(Federal Open Market Committee，FOMC)所發布之官方文件，以每句話為單位，加入先驗知識分為不同經濟主題，再用情緒分析方法分為正面或負面情緒，並以情緒數量建立指標來預測台灣總體經濟變數。實證結果發現，在區分主題後，迴歸結果普遍比未區分主題來的好；而情緒分析方法，使用BERT來預測情緒的結果相較財務情緒字典方法來的合理。

This paper uses the Latent Dirichlet Allocation topic modeling to divide the official documents issued by the Central Bank of Taiwan and the Federal Open Market Committee(FOMC) into different topics with the prior economic knowledge, and then use sentiment analysis to divide them into positive or negative sentiment, and use the amount of sentiment to establish indicators to predict Taiwan's economic variable. The empirical results that after distinguishing topics, the regression results are generally better than those without distinguishing topics. The method of using BERT to predict sentiment is better than tranditional sentiment analysis.