A Component-level Analysis of an Academic Search Test Collection

Abstract

This study analyzes search performance in an academic search test collection. In a component-level evaluation setting, 3,276 configurations over 100 topics were tested involving variations in queries, documents and system components resulting in 327,600 data points. Additional analyses of the recall base and the semantic heterogeneity of queries and documents are presented in a parallel paper. The study finds that the structure of the documents and topics as well as IR components significantly impact the general performance, while more content in either documents or topics does not necessarily improve a search. While achieving overall performance improvements, the component-level analysis did not find a component that would identify or improve badly performing queries.