A Learning Personal Agent for Text Filtering and Notification

Abstract

The WWW is increasingly being used for announcements of important events and solicitations such as conference announcements and requests for proposals. This information is accessed by users using direct manipulation tools. The volume of this information is increasing daily and users currently must sift through large amounts of text to access relevant information. We describe a reusable agent that learns a model of the user's preferences, scouts appropriate information sources, fi lters the information, and notifi es the user when relevant information becomes available. The agent is a Personal Assistant which operates autonomously with minimal user intervention. The agent's task is to identify conferences and request for proposals that fi ta user's research interests. For this task, there is a large volume of irrelevant documents and the proportion of relevant documents is very small. It is also critical that the agent not misclassify relevant documents, if need be, at the cost of misclassifying a few irrelevant documents. Information Retrieval and Neural Network techniques were utilized to learn the model of user's preferences. Readily available textual information was used for training, so that the agent's performance at startup is quite high. The agent has been evaluated through extensive experimentation under a variety of conditions. The results are analyzed and the comparative performance of the learning techniques used are discussed. An interesting result observed is that though Neural Network techniques are inferior in performance to Information Retrieval techniques for learning from text in more general text fi ltering tasks, in our task they had comparable performance.