主題人物立場分析研究

Abstract

隨著網路的爆炸性成長，人們能夠輕易地從網路獲得龐大的資訊，而且人們可能會被網路上的多媒體所提供的資訊所掩沒，像是新聞、網路評論、論壇文章或是從社群媒體來的資訊。為了協助人們消化這些資訊，在此博士論文中，我們探究一個新穎的主題，稱為主題人物立場辨識，這個主題的目地是辨識主題文件中人物的立場。我們提出了兩套方法來解決這個問題。首先，我們提出一套叫做模式基礎EM的方法，利用人物名字共同出現在文件中的模式來辨識主題人物的立場。此外，文件中人名共同出現與不共同出現的程度被考量用以加權人名共同出現的模式。甚至，我們發展一個初始化演算法來穩定辨識人物立場社群，這是因為模式基礎的EM方法對於初始化是頗敏感的。第二套方法稱做使用友誼網路分析的主題人物立場社群辨識，這套方法考量文件的友善(敵對)傾向自動地從主題文件建構友誼網路。此外，我們提出立場擴展與立場修正演算法基於友誼網路來辨識立場社群。實驗結果驗證這兩套方法都比過去知名的分群演算法效能來得好。

With the explosive growth of the Internet, people can easily receive astronomical information from the Web, and could be overwhelming by the online medium, e.g. news, review comments, forum posts or information from the social medium. For facilitating the people digest the enormous information, we investigate a novel problem named “topic person stance identification,” which is to identify the stances of the topic persons from topic documents, in this dissertation. We proposed two methodologies to copy with the problem. First, we proposed a methodology named model-based EM method to identify the stances of the topic persons by leveraging the pattern of person name co-occurrence in the documents. In addition, the level of co-occurrence and non-co-occurrence of the person names in the documents are considered to weight the pattern of the person name co-occurrence. Moreover, we developed an initialization algorithm to stable the results of identifying the stance communities because the EM method is sensitive to the initialization. The second methodology is called stance community identification of topic persons using friendship network analysis. This method is to take the friendly (opposing) orientation of the documents into consideration to construct the friendship network automatically from the topic documents. For identifying the stance community, we proposed stance community expansion and stance community refinement algorithms to identify the stance communities based on the network. The experimental results of two methodologies demonstrated our methods are outperformed other well-known clustering approach, and can effectively identify the stances of the topic persons.