It is well-known fact that the combination of the retrieval outputs of different search systems in response to a query, known as metasearch, improves performance on average, provided that these combined systems (1) have compatible outputs, (2) produce accurate probability of relevance estimates of documents, and (3) be independent of each other. The objective of a normalization technique is to target the first requirement, i.e., document scores of different retrieval outputs are brought into a common scale so that document scores can be comparable across combined retrieval outputs. This has been a recent subject of researches in metasearch and information filtering fields. In this paper, we present a different perspective on multiple evidence combination and investigate various normalization techniques, mostly ad-hoc in nature, with a special focus on the SUM, which shifts minimum scores to zero and then scales their summation to one. This formal approach is equivalent to normalize the distribution of scores of all documents in a retrieval output by dividing them by their sample mean. We have made extensive experiments using ad hoc tracks of third and fifth TREC collections and CLEF'OO database. We argue that (1) the normalization method SUM is consistently better than the other traditionally proposed ones when combining outputs of search systems operating on a single database; (2) the SUM for combination of outputs of search systems operating on mutually exclusive databases is still valuable alternative to the one weighting score distributions of documents by their databases' size.
展开▼
