IRE
Information Retrieval Experiment
The Cranfield tests
chapter
Karen Sparck Jones
Butterworth & Company
Karen Sparck Jones
All rights reserved. No part of this publication may be reproduced
or transmitted in any form or by any means, including photocopying
and recording, without the written permission of the copyright holder,
application for which should be addressed to the Publishers. Such
written permission must also be obtained before any part of this
publication is stored in a retrieval system of any nature.
Criticisms of Cranfield 1 267
Electric test supported the Cranfield results in relation to success rates and
reasons for failures, while the WRU test was notable first for searching
without reference to wanted document numbers, and second, for the use of
both recall and precision; the results for recall for the facet system were like
those of the main test, while the WRU failures were attributable chiefly to
poor searching.
According to Lancaster and Mills, the Aslib Cranfield Project as a whole
served to establish
`reasonably reliable figures. . . for the degree of recall and relevance [i.e.
precision] likely to be achieved by a good index employing fairly exhaustive
indexing' (p.8),
i.e. performance in terms of operating efficiency. Further, the broad
implication is that
`as one factor amongst several determining the operating efficiency of
indexes, the indexing system, per se, is of less importance than has been
assumed generally.' (p.9)
However in their view
`the first Cranfield investigation was perhaps less important for the light it
shed on four actual indexes than for its contribution to the development of
techniques for the testing of information retrieval systems and for exposing
the basic parameters in the operation of an indexing system.... What
was, to begin with, a rather blunt tool is being developed into a sharp
analytical instrument.' (p.9)
Lancaster and Mills conclude that the inverse relationship between recall
and precision is inescapable, and that the recall/precision character of a
system is determined by indexing exhaustivity and specificity, determining
recall and precision respectively. Moreover, since search relaxation can
improve recall,
`we are left with the specificity of the index as its key characteristic.
Hospitality to specific indexing, by which we mean the ability to describe
precisely the concepts chosen as retrieval tags, is the most important single
criterion of an index language' (p. 11);
or
`to summarise, for every index language, depending upon its hospitality for
specific indexing, there is a maximum relevance ratio which cannot be
exceeded. However it is always possible to improve the recall ratio along
the fixed performance curve to its maximum by means of variations in
search programmes. But the ability to do this is dependent upon the
exhaustivity of the indexing.' (p. 11)
13.2 Criticisms of Cranfield 1
The aims and character of the Aslib-Cranfield Project meant that the
Cranfield 1 results attracted much attention. Those with indexing language
positions to defend tended to attack the results without offering any concrete
i