General Conclusion - Geographical Information Retrieval in Textual Corpora - page 89

As illustrated in Figure 1.6 of Chapter 1, these propositions have been mainly
implementedintheGEOSEM2,GEOTOPIA,GEOCIMEandGEONTOprojects,and
have been tested in numerous versions of the PIV prototype. These experiments have
enabled us to validate the following working hypotheses:
- A devoted spatial IRS gives better results than a classic thematic IRS for IR
composed of only spatial criteria.
- A devoted temporal IRS gives better results than a classic thematic IRS for IR
composed of only temporal criteria.
- A “rough” coupling of spatial, temporal and thematic IRSs gives better results
than a classic thematic IRS for multicriteria IR despite the numerous possible biases.
4.1.3. Multicriteria IR in texts
These pieces of work have the objective of defining a meta-engine of geographic
IR, which federates spatial, temporal and thematic IRSs. They are therefore based on
the previous results relative to the spatial and temporal IR and are divided into two
main contributions.
Our first contribution targets the generalization of data representations. To avoid
possible biases, we have chosen to standardize the data representations and the
approaches of processing data relative to the different dimensions. Thus, from the
indexed representations of spatial and temporal information, we build higher level
indexes, suitable for the implementation of IR models such as the vectorial model.
Our second contribution targets the partially compensatory aggregation of results
in a context of spatial, temporal and thematic IRSs federation. The aggregation
model we developed offers modal operators allowing us to associate a level of
preference and requirement with each search criteria. This model integrates the
matching operators supported by the federated IRSs and extends their expressiveness
using modal operators.
As illustrated in Figure 1.6 in section 1.4.3, these propositions have been
implemented in the CIDRI project, and have been tested in numerous versions of
the PIV 2 and PIV 3 prototypes. These experiments have allowed us to validate the
following working hypotheses:
- The generalization of data representation is adapted to the spatial and temporal
information.
- The generalization associated with the vectorial model does not imply a loss of
precision with respect to the initially proposed devoted IRSs.