財金新聞情緒分類實證研究──以鉅亨網財金新聞為例

Empirical Study of Financial News Sentiment Classification: Evidence from Anue Financial News

現今網路新聞媒體崛起，網路財經新聞成為投資人執行投資決策的判斷依據之一，然而財金新聞網站每天所產生的大量新聞資訊，投資人如何從大量財金新聞中逐一判斷每篇新聞所反映之市場情緒，將成為一大挑戰。因此，本研究以Anue鉅亨網(www.cnyes.com)的臺灣股市新聞進行中文新聞文本情緒分類，使用機器學習、文字探勘與文本分類的技術，探討文本的資料前處理(包含N-gram、Term Frequency–Inverse Document Frequency [TF-IDF]、卡方與互資訊)搭配文本分類模型(樸素貝葉斯分類器、fastText與深度學習)之情緒分類效果。研究結果指出：(1) N-gram字詞特徵提取在三種分類模型中皆能提升準確率，特別是應用在樸素貝葉斯分類器，可以有效地改善模型假設字詞特徵獨立的缺點。(2)深度學習之多層感知器模型搭配卡方檢定與互資訊的關鍵特徵挑選下，二元分類與三元分類準確度均可分別突破90%與80%以上，但訓練時間較長。(3) fastText搭配卡方檢定與互資訊特徵挑選，相對於深度學習模型，可以用更短的訓練時間達到與深度學習相似的分類準確率。

"Online news has become one of the major sources for investors to make investment decisions. However, it has become a great challenge for investors to judge and verify the current market sentiment from a huge amount of financial news generated by financial news websites. Consequently, using financial news collected from Taiwanese stock market news of Anue website (www.cnyes.com), this paper employs three text classification models (naive Bayes classifier, fastText, and deep learning model) with four feature extraction and selection techniques of text mining (N-gram, term frequency–inverse document frequency [TF-IDF], Chi-square test, and mutual information) to classify news sentiments. The empirical results demonstrate that (1) N-gram feature extraction can improve the accuracy of all classification models, especially for the naive Bayes classifier which can effectively overcome the shortcomings of independence assumptions of text features. (2) In spite of the long training time, multi-layer perceptron with Chi-square test and mutual information feature selection techniques can provide the best performance for two and three classification with the accuracy up to 90% and 80%, respectively. (3) Compared to multi-layer perceptron, fastText model with Chi-square test and mutual information feature selection techniques can reduce the training speed without loss of the accuracy of classification."