Multicriteria Information Retrieval in Textual Corpora - Geographical Information Retrieval in Textual Corpora - page 63

a function f as the weighted sum, for example. A weak score for one criterion can
perhaps be balanced by one or more higher scores obtained for other criteria. So this
kind of approach does not handle any mandatory criterion whose satisfaction would
be required for any result. Prominent compensatory aggregation functions are Borda
Count [DEB 81], CombMAX, CombMNZ and CombSUM [FOX 93].
- Non-compensatory aggregation: Inthiscase,aggregationismainlybasedonthe
value of one criterion, such us the weakest score rsv ji assigned to the document d i .
The remaining criteria are only used to discriminate between documents of similar
scores. The main disadvantage of this class of aggregation functions is that a large
part of the criteria (and the corresponding scores) are ignored or play a minor role.
CombMIN [FOX 93] and Prioritized And [DA 09, DA 12] are non-compensatory
aggregation functions.
- Partially compensatory aggregation: FarahandVanderpooten[FAR 08]propose
an approach to result combination that takes into account the underlying vagueness
of the query formulation process. Similarly, Pereira et al. [DA 09, DA 12] propose
the “Prioritized Scoring Model” (PSM) aggregation function that assigns the highest
weight to the first criterion and decreasing weights to the following criteria.
Figure 3.1. The principles of multicriteria IR [PAL 12a]
Consequently, we rank GIRSs based on approaches of linear combination in the
category of totally compensatory aggregation functions. However, let us note that the
resultsproposedbythesesystemscouldbebiased: indeed,theyaggregateresultsfrom
dedicated IRSs characterized by specific models of data representation and relevance
score calculation.
Similarly, we associate the GIRSs based on filtering approaches with the category
of non-compensatory aggregation functions. These systems process sequentially, or
in parallel, each dimension and calculate the intersection of the corresponding result
sets. The final result is unranked or ranked following only one dimension.