Study of Adaptive Relevance Feedback - UIUC TREC-2008 Relevance Feedback Experiments

摘要

In this paper, we report our experiments in the TREC 2008 Relevance Feedback Track. Our main goal is to study a novel problem in feedback, i.e., optimization of the balance of the query and feedback information. Intuitively, if we over-trust the feedback information, we may be biased to favor a particular subset of relevant documents, but under-trusting it would not take advantage of feedback. In the current feedback methods, the balance is usually controlled by some parameter, which is often set to a fixed value across all the queries and collections. However, due to the difference in queries and feedback documents, this balance parameter should be optimized for each query and each set of feedback documents. To address this problem, we present a learning approach to adaptively predict the balance coefficient (i.e., feedback coefficient). First, three heuristics are proposed to characterize the relationships between feedback coefficient and other measures, including discrimination of query, discrimination of feedback documents, and divergence between the query and the feedback documents. Then, taking these three heuristics as a road map, we explore a number of features and combine them using a logistic regression model to predict the feedback coefficient. Experiments show that our adaptive relevance feedback is more robust and effective than the regular fixed-coefficient relevance feedback.