Evaluation of Score Standardization Methods for Web Search in Support of Results Diversification

Zhongmin Zhang, Chunlin Xu, Shengli Wu

Abstract: Score standardization is a necessary step for many different types of Web search tasks in which results from multiple components need to be combined or re-ranked. Some recent studies suggest that score standardization may have impact on the performance of some typical explicit search result diversification methods such as XQuAD. In this paper, we evaluate the performance of six score standardization methods. Experiments with TREC data are carried out with two typical explicit result diversification methods XQuAD and PM2. We find that the reciprocal standardization method performs better than other score standardization methods in all the cases. Furthermore, we improve the reciprocal standardization method by scaling those scores up so as to better satisfy the requirement of probability scores and obtain better results with XQuAD. We confirm that score standardization has significant impact on the performance of explicit search result diversification methods and such a fact can be used to obtain more profitable score standardization methods and result diversification methods.