The Effect of Normalization when Recall Really Matters.
April Kontostathis, Scott Kulp · 2008
Abstract: In this article we describe a series of search and retrieval experiments to compare the impact of two different types of weighting and normalization schemes on recall. We first describe both the state-of-the-art BM25 algorithm and the newly introduced power normalization algorithm in detail. We then determine the optimal parameter settings for both systems. Finally we compare their retrieval performance. Our results show that BM25 does not maximize recall for two different collections (TREC Legal and TREC Disks 1 & 2) and three different query sets. Designers of applications which require maximal recall, such as legal discovery, should consider using power normalization instead of BM25.