THE PARAMETRIC DESCRIPTION OF RETRIEVAL TESTS

Abstract

Two general requirements for overall measures of retrieval effectiveness are proposed, namely that the measure should be as far as possible independent of generality (this is interpreted to mean that it can be described in terms of recall and fallout), and that it should be able to measure the effectiveness of a performance curve (it should not be restricted to a simple 2×2 table). Several measures that have been proposed are examined with these conditions in mind. It turns out that most of the satisfactory ones are directly or indirectly related to Swets' measure A, the area under the recall‐fallout curve. In particular, Brookes' measure S and Rocchio's normalized recall are versions of A.