Application of text topic and relationship mining for multi-document summarization: using movie reviews as an example

由於網際網路的普及造成資訊量愈來愈大，在資訊的搜尋、整理與閱讀上會耗費許多時間，因此本研究提出一應用文本主題及關係探勘的方法，將多份文件自動生成一篇摘要，以幫助使用者能降低資訊的閱讀時間，並能快速理解文件所欲表達之意涵。 本研究以電影評論文章為例，結合文章結構的概念，將影評摘要分為「電影資訊」、「電影劇情介紹」及「心得結論」三部分，其中「電影資訊」及「心得結論」為透過本研究建置之電影領域相關詞庫比對得出。接著將餘下之段落歸屬於「電影劇情介紹」，並透過LDA主題模型將段落分群，再運用主題關係地圖的概念挑選各群之代表段落並排序，最後將各段落去除連接詞及將代名詞還原為其所指之主詞，以形成一篇列點式影評摘要。 研究結果顯示，本研究所實驗之三部電影，產生之摘要能涵蓋較多的資訊內容，提升了摘要之多樣性，在與最佳範本摘要的相似度比對上，分別提升了10.8228%、14.0123%及25.8142%，可知本研究方法能有效掌握文件之重點內容，生成之摘要更為全面，藉由此方法讓使用者自動彙整電影評論文章，以生成一精簡之摘要，幫助使用者節省其在資訊的搜尋及閱讀的時間，以便能快速了解相關電影之資訊及評論。The rapid development of information technology over the past decades has dramatically increased the amount of online information. Because of the time-wasting on absorbing large amounts of information for users, we would like to present a method in this thesis by using text topic and relationship mining for multi-document summarization to help users grasp the theme of multiple documents quickly and easily by reading the accurate summary without reading the whole documents. We use movie reviews as an example of multi-document summarization and apply the concept of article structures to categorize summary into film data, film orientation and conclusion by comparing the thesaurus of movie review field built by this thesis. Then we cluster the paragraphs in the structure of film orientation into different topics by Latent Dirichlet Allocation (LDA). Next, we apply the concept of text relationship map, a network of paragraphs and the node in the network referring to a paragraph and an edge indicating that the corresponding paragraphs are related to each other, to extract the most important paragraph in each topic and order them. Finally, we remove conjunctions and replace pronouns with the name it indicates in each extracted paragraph s and generate a bullet-point summary. From the result, the summary produced by this thesis can cover different topics of contents and improve the diversity of the summary. The similarities compared with the produced summaries and the best-sample summaries raise of 10.8228%, 14.0123% and 25.8142% respectively. The method presented in this thesis grasps the key contents effectively and generates a comprehensive summary. By providing this method, we try to let users aggregate the movie reviews automatically and generate a simplified summary to help them reduce the time in searching and reading articles