Data Extraction, Transformation and Integration Guided by an Ontology - Data Warehousing Design and Advanced Engineering Applications - page 32

Table 2. L2R results on Cora data sets

RDFS+ RDFS+ & DP

Recall (REC)
Recall (NREC)

52.7%
50.6%

52.7%
94.9%

Recall 50.7% 94.4%

Precision 100% 100%

al., 2005). The point is that L2R is not supervised
and guarantees a 100% precision.

no reconciliations inferred by L2R. Moreover,
the reconciliations inferred by L2R are not re-
computed in N2R.
These experimentations show that good results
can be obtained by an automatic and unsupervised
method if it exploits knowledge declared in the
schema. Furthermore, the method is able to obtain
F-Measure which is better than some supervised
methods such that (Parag & Domingos, 2005).

N2R Results
In the following we presents the results (see
Figure 6) obtained by N2R after the application
of L2R.
For T rec =1 , N2R do not obtain more results than
L2R. The evolution of the recall and precision
values in function of T rec is interesting. Indeed,
when the threshold is decreased to 0.85, the recall
increases by 33% while the precision only falls
by 6%. The best results are obtained when T rec
=0.85 . The F-measure is then at its maximum
value of 88%. Besides, when the recall value is
almost of 100%, for T rec =0.5 , the precision value
is still about 40%.
The exploitation of the non-reconciliation
inferred by L2R allows an important reduction
of the reconciliation space handled in N2R. For
the Cora data set the size of the reconciliation
space is about 37 millions of reference pairs. It
has been reduced of 32.8% thanks to the correct

FUTURE RESEARCH DIRECTIONS

Work on integration has evolved in recent years
to heterogeneous information sources available
from the Web or through the intranets and the
heterogeneity is becoming more and more impor-
tant. In the future, we will have to deal with a still
larger variety of data structures types (structured,
semi-structured or unstructured sources), but also
with sources containing both semi-structured and
unstructured (or textual) parts. Current wrapper
approaches could not be applied with the last kind

Figure 6. N2R results obtained on Cora data set