Multicriteria Information Retrieval in Textual Corpora - Geographical Information Retrieval in Textual Corpora - page 82

foragivendocument,thesumofthescoresobtainedforthedifferentsearchcriteriais
weighed by the number of IRSs having retrieved this document [PAL 10a, SAL 12].
Thus, PIV 3 , configured according to the CombMNZ model, improves the
precision of the Terrier , PIV 2 _spatial or PIV 2 _temporal IRS, alone
[PAL 10a, SAL 12]. As shown in Figure 3.11, the improvement with respect to the
bag of words-type ( Terrier ) thematic IR is equal to 66%. Let us note that the spatial,
temporal and thematic criteria are of the same importance here. Moreover, the
detailed analysis of the results shows a small number of documents returned
simultaneously for the three dimensions, consequently showing that the three
dimensions are complementary. A detailed description of these experiments is
presented in [PAL 10a, PAL 10c, CAB 11] and [SAL 12].
Figure 3.11. Comparison of the Terrier and PIV 3 CombMNZ IRSs
3.5.3.2. Evaluation of the partially compensatory multicriteria IR
No test collection exists which covers all the three geographic dimensions and
which, in addition, proposes topics associating requirements or levels of preference
with the search criteria. So, this experiment has been carried out on the MIDR_2010
test collection for 10 expressive geographic topics targeting territories, periods and
subjects invoked in this corpus (e.g. “Pyrénées mountains but not those of Gavarnie,
in the 19 th Century, if possible, unrelated to ascents”). We have experimented the
PIV 3 GIRS endowed with the CombMNZ aggregation models, on the one hand, and
CMRP, on the other. In this latter case, the different scores of a document are affected
by coefficients, which are the levels of preference expressed in the query.
The PIV 3 CMRP GIRS, based on a partially compensatory aggregation, improves
the precision of the PIV 3 CombMNZ GIRS, based on a totally compensatory
aggregation. As shown in Figure 3.12, this improvement is equal to 54%. Let us note
that here the set of queries associates levels of preference and requirements with the
criteria. The observed difference illustrates well the fact that CMRP allows users to
refine the roles for each criterion, contrary to CombMNZ. Indeed, CombMNZ
supports a totally compensatory aggregation while CMRP offers an increased power
of expression to the users. A detailed description of this experiment is presented in
[PAL 12a] and [SAL 12]. We have to continue our tests and evaluate more than 25
topics in order to confirm this trend.