Geoscience Reference

In-Depth Information

and Grabisch [LAB 03], it is necessary to identify such links and measure them,

which is generally difficult for a non-specialist.

3.3.2.3. Filtering and GIR

In Chapter 2, we have defined geographic information composed of three

dimensions: spatial, temporal and thematic. We have presented various existing

GIRSs and noted that many systems were only interested in two of these dimensions

(generally spatial and thematic). Palacio [PAL 10a] proposes regrouping the

approaches to result combination implemented by these GIRSs in three categories:

filtering (sequential or parallel), linear combination and projection. Let us note that

the filtering approaches are the most widely employed. A presentation of these

categories, illustrated with examples, is proposed in [PAL 10a].

The approach corresponding to sequential filtering consists of performing a first

retrieval based on criteria of a given dimension: the result is a set of ranked relevant

documents. Then, a new retrieval, based on criteria relative to another dimension, is

carried out on this set of results. Finally, a last retrieval concerns the third geographic

dimension. The first two retrievals allow us to reduce the set of resulting documents

(filtering approach), and the last retrieval allows us to obtain a set of ranked results

with respect to the criteria relative to the last processed dimension. Thus, sequential

filtering allows us not to combine the three dimensions directly. The

STEWARD [LIE 07] and SINAI [GAR 09] systems implement such an approach for

a thematic and spatial filtering. Lieberman et al. [LIE 07] argue in favor of a

contextual configuration of the order of use of dimensions in the process of

sequential filtering.

The approach corresponding to parallel filtering consists of querying the

document repository with respect to each of the dimensions of the query separately,

then performing the intersection of the results. Thus, the query is subjected to three

processes proper to each geographic dimension. For every dimension, a set of results

is retrieved. Only the documents present in the three sets are kept. No ranking is

performed. However, in the set of results obtained, some documents are evidently

morerelevantthan others and should appear on top of the list. This approach does not

favor any dimension as does sequential filtering, for example. The SPIRIT [VAI 05],

GRID [VAL 06], GEOTRACKER [CHE 07], WatWasWaar [LIB 08],

CITER [PFO 09] and Document Trajectory [STR 10] systems implement parallel

filtering.

Other approaches to combination are applied to GIR. Linear combination allows

us to combine scores from several IR systems [VOG 99], the objective being to

obtain a unique score. For example, the arithmetic average corresponds to a linear

combination associating an equal weight with each dimension: GEOSEM [BIL 07]

Search WWH ::

Custom Search