Extending Relevance Model for Relevance Feedback

Relevance feedback is the retrieval task where the system is given not only an information need, but also some relevance judgement information, usually from users’ feedback for an initial result list by the system. With different amount of feedback information available, the optimal feedback strategy might be very different. In TREC Relevance Feedback task, the system is given different sets of feedback information from 1 relevant document to over 40 judgements with at least 3 relevant. Thus, in this work, we try to develop a feedback algorithm that works well on all levels of feedback by extending the relevance model for pseudo relevance feedback to include judged relevant documents when scoring feedback terms. Within these different levels of feedback, it is more difficult for the feedback algorithm to perform well when given minimal amount of feedback. Experiments show that our algorithm performs robustly in those difficult cases.