運用意見探勘於企業聲譽分析之研究──以企業社會責任為主題

A Study of Corporate Reputation Analysis by Opinion Mining - A Case of Corporate Social Responsibility

網路的發達及社群媒體的發展，讓許多企業高層開始重視社群媒體的影響力，希望能瞭解在網路上的企業聲譽及形象。近年來大眾亦相當看重企業對於整體經濟、公益活動、社會參與以及環境永續等社會責任相關議題，使得「企業社會責任」逐漸成為企業建立聲譽及形象的重要指標。因此，本研究以企業社會責任為主題，透過網路資料爬取、意見探勘技術，分析網路上對企業在社會責任上的評價，並透過兩個網路資料源比較、驗證各式機器學習分類方法之成效。本研究蒐集PTT與FB網路社群平台中與企業相關之新聞及評論留言，進行文本分類及情緒分析；使用SVM、CNN及LSTM三種方法進行新聞文本分類，以CNN、LSTM及Bi-LSTM方法將對應之評論留言進行情緒分類，並以語意分析計算出企業在企業社會責任四個構面與整體之情緒得分，藉以呈現企業在企業社會責任之表現與評價。本研究主要結論如下：（1）本研究提出之情緒分析模型，經驗證可有效分析出網路大眾對特定企業在「企業社會責任」四個構面與整體之評價；（2）SVM用於辨識「企業社會責任」新聞文本之成效相對最為穩定，CNN及LSTM之分類成效則較不穩定；（3）Bi-LSTM用於辨識「企業社會責任」新聞文本評論留言之正負情緒傾向成效最佳，其次為CNN及LSTM；（4）不同資料來源會呈現出不同評論留言情緒傾向。

The development of the Internet and social media have led many corporate executives pay more attention to the influence of social media, hope to understand their reputation and image on the Internet. In recent years, the public has paid great attention to corporate social responsibility issues, such as overall economy, charitable event, social participation, and environmental sustainability, making the "Corporate Social Responsibility" dimensions become an important indicator of corporate reputation and image. Thus, this study takes Corporate Social Responsibility (CSR) as an example, through text mining and deep learning technology, proposes an index and analysis module for measuring CSR. This study collects news text and comment messages about corporations on the social media (PTT、FB), marks the news texts according to the CSR dimensions, and uses SVM, CNN and LSTM three classification methods to find out the better classification method. At last, uses CNN, LSTM and Bi-LSTM to sentiment classify the comment about the corporation, calculate sentiment scores on all the dimensions to show the CSR performance and evaluation. The main conclusions of this study are as follows: (1) The sentiment analysis model proposed in this study can be verified to effectively analyze the Internet public's evaluation of specific enterprises in the four aspects of CSR and overall; (2) SVM is relatively stable for the effectiveness of identifying the CSR news text, and the classification effectiveness of CNN and LSTM is relatively unstable; (3) Bi-LSTM is best to identify the positive and negative sentiment tendencies of the CSR news text comments, followed by CNN and LSTM; (4) Different sources of data pool will show different sentiment tendencies to comments.