運用搜尋關鍵字、社會化書籤與標籤以實現考慮信任度的個人化知識分享

Abstract

資訊超載早已是惡名昭彰的問題。尤其在目前以搜尋服務為主力的資訊取得模式下，許多真正被需要的線上資源常常因為受限於個人對於專業領域的掌握程度不足﹝導致想不出合適的搜尋關鍵字﹞而久不見天日。面對這樣的情況，如果系統能夠主動推薦資訊、同時考慮個人化的需求，相信能降低不少前述資訊超載所帶來的衝擊與挑戰。 本研究提出了運用從社會性標籤(Social Tags)與搜尋關鍵字(Search Keywords)搜集而來的詞彙，為使用者、線上內容/文件以及前述的「詞彙(terms)」(包含社會性標籤或搜尋關鍵字)建構以向量空間模式(VSM)為基礎的識別標誌(Profile)；同時一併保存由搜尋關鍵字與從搜尋結果中去蕪存菁地選為線上書籤(Online Bookmark)的文章，並將這樣的組合視為一種珍貴且值得推薦的資源。而透過使用者與使用者、使用者與文章、使用者詞彙的相關度計算，系統將透過上述的推薦以提供一種個人化知識分享的服務，嘗試更貼近使用者的需求。 在考慮識別標誌間的「相關度」時，我們除了引用常見的相似度(Similarity)計算，亦延伸了一種對人下標籤的「聯繫人管理(Contact Management)」機制作為具體化使用者間信任之平台以及個人表達需求的管道。對某人下標籤將使該人識別標記中與所被下之標籤相對應的特徵因此被加權，進而影響相關度/相似度的計算使結果更能反應個人的偏好與需求。 最後，我們亦延伸了常見的「加為好友」社群功能以加權來自於「朋友」的標籤(記錄)，並透過非對襯式的「標籤共存(tag co-occurrence)」分析以提供客製化的標籤關聯以提供更好的個人化服務。

Information overload has been notoriously posing tremendous obstacles to more efficient and effective (online) resource utilization. Under the dominance of information pull, in which users have to actively find what they need, services being able to push what we would be interested in or in demand of are often expected. In this paper, we propose an approach, with which social tags and keywords extracted from search inputs are coordinated in terms of VSM to profile users, online contents/documents, and the textual terms themselves. The system stores pairs of users’ search inputs and (Internet) bookmarks selected from the search results, and treats the resulting pairs as valuable resources that are worthy of recommendation. By computing user-user, user-document, and user-tag relatedness values from VSM-based profiles, the system aims to achieve personalized knowledge sharing by recommending the previously-mentioned search input-output pairs that are expected to be related to and catering for users’ needs. In addition to (profile) similarity, we also take interpersonal “trust” into consideration while defining “relateness.” By adopting an innovative contact management mechanism—People Tagging, we allow users to express their preferences for recommendations and their willingness to trust others in specific domains, therefore making recommendations more relevant. Lastly, based on commonly-seen social network function, for each individual we weight tagging records from people accepted as friends/buddies and provide customized tag relatedness based on an asymmetric tag co-occurrence measure. With these features we expect to achieve higher level of personalization.