隨著網際網路的發展，使用者可以輕易地經由網路取得大量的資訊，卻也同時必須面對著資訊過載(Information Overload)的問題。因此如何從大量的資訊中，提取出使用者感興趣的資訊，就成為資訊爆炸時代非常重要的議題，資訊過濾的研究議題也因此應運而生。然而，不同於傳統的分類問題著重於對靜態的資料進行分類，資訊過濾系統必須時常面對使用者興趣隨著時間逐漸甚至快速改變的情況，我們將這種資料分布會隨著時間遞移而有所改變的問題稱為概念漂移(Concept drift)。當使用者的興趣發生概念漂移，資訊過濾系統必須要有足夠的能力去偵測概念漂移的發生，並即時的調整更新使用者的興趣模型。傳統的資訊過濾系統通常必須透過蒐集大量的使用者回饋資訊，反應使用者的改變，才能維持穩定的過濾效果。本研究運用NGD能夠即時計算字詞間語意關係的特性，提出個一個能以極少量訓練文件為基礎建立使用者興趣模型的動態文件過濾系統，改進突發性概念飄移發生後，文件過濾系統的反應效率不足的問題。 Due to the development of the Internet, people can access mass information easily from a variety of search engines and portal; however, in the meantime, people also have to face the problem of “information overload”. Therefore, how to extract useful information for the users from the mass information has become a vital issue in the information explosion era, and the research of information filtering has been caused. Nevertheless, different from the traditional classification which focused on the classification of static information, information filtering system has to face the situation that the interests of users would change dynamically. The phenomenon that the distribution of data changes over time is called “Concept drift”. When concept drift happens to the interests of users, the information filtering system has to have sufficient ability to detect the happening of concept drift; furthermore, it has to adjust and update the interest models of users in time. Traditionally, the information filtering system has to collect a lot of feedback information to reflect the interest change of user, so that the filer could be stable and effective. In order to improve the inefficiency of information filtering system when concept drift happens, this research applied the characteristic of NGD, which could recognize the relationships between the meanings of different terms, to propose a dynamic information filtering system which could establish the interest models of users by limited training documents.
