Improving the Filtering Quality of Selective Dissemination of Information by Observing User Task Behavior
A selective dissemination of information (SDI) service alerts users to latest documents in their field of interest. SDI helps users cope better with streams of incoming massive amounts of documents by filtering out uninteresting documents. Existing SDI filtering mechanisms typically use user feedback based on a binary 'interesting/not interesting' decision per document. However, in some situations the user will use information from the document in a structured way to complete a task, for example to fill out a form. In this work we describe a method of using user task behavior to improve the quality of SDI filtering. Our method creates, via machine learning, information extraction models of the information that the user is interested in. We then use these models on incoming documents to extract parts of the document that are likely to be of interest to the user. The results of the extraction are added as features into the filtering algorithm. We describe some experimental results of this method that demonstrate improved filtering performance of the SDI system (9%). We then describe the conditions where this method may be applied.