Relevance distributions across Bradford Zones: Can Bradfordizing improve search?

The purpose of this paper is to describe the evaluation of the effectiveness of the bibliometric technique Bradfordizing in an information retrieval (IR) scenario. Bradfordizing is used to re-rank topical document sets from conventional abstracting & indexing (A&I) databases into core and more peripheral document zones. Bradfordized lists of journal articles and monographs will be tested in a controlled scenario consisting of different A&I databases from social and political sciences, economics, psychology and medical science, 164 standardized IR topics and intellectual assessments of the listed documents. Does Bradfordizing improve the ratio of relevant documents in the first third (core) compared to the second and last third (zone 2 and zone 3, respectively)? The IR tests show that relevance distributions after re-ranking improve at a significant level if documents in the core are compared with documents in the succeeding zones. After Bradfordizing of document pools, the core has a significant better average precision than zone 2, zone 3 and baseline. This paper should be seen as an argument in favour of alternative non-textual (bibliometric) re-ranking methods which can be simply applied in text-based retrieval systems and in particular in A&I databases. Introduction The perceived expectations of users searching the web are that retrieval systems should list the most relevant or valuable documents in the result list first (so-called relevance ranking). More approaches appear that draw on advanced methods to produce relevant results and alternative views on document spaces. Google PageRank and its derivations (see e.g. Lin, 2008) or Google Scholar’s citation count are just two popular examples for informetric-based rankings applied in Internet search engines. Distributed search across multiple A&I databases will also generate large and heterogeneous document sets with the effect that users are confronted with a massive load of results from different scientific domains, even for specific research topics. Furthermore, empirical tests with typical A&I databases like Medline show that conventional term frequency inverse document frequency (tf-idf) best match models and especially recent web-based ranking methods implemented in search engines (originally for web pages) are not always appropriate for search in heterogeneously collected scholarly metadata documents. In this paper we want to apply and evaluate a non-textual ranking technique, called Bradfordizing. Introduced by H.D. White (1981), Bradfordizing is a bibliometric method to reorganize a search result for a topic. Bradfordizing is set up by applying the following procedure: “... that is sorting hits (1) by the journal in which they appear, and then sorting these journals not alphabetically by title but (2) numerically, high to low, by number of hits each journal contains. In effect, this two-step sorting ranks the search output in the classic Bradford manner, so that the most productive, in terms of its yield of hits, is placed first; the secondmost productive journal is second; and so on, down through the last rank of journals yielding only one hit apiece.” (White, 1981: p. 47). Bradford Law Journals play an important role in the scientific communication process. They appear periodically, they are topically focused, they have established standards of quality control and