Semantic Similarity based on Weighted Ontology - Geographical Information Systems: Trends and Technologies - page 312

S n }, n for K E . Assume h w k . The geographic class similarity , GSim ( m 1 , m 2 ),
is given as follows:

ܩܵ݅݉ሺ݉ ଵ ǡ݉ ଶ ሻൌߣܵ݅݉ሺܩ ଵ ǡܩ ଶ ሻ

כቀݓ ௣௔௥௧ כ݅ܿݏሺ݉ ଵ ǡ݉ ଶ ሻ൅ݓ ௔௧௧ כݐܵ݅݉ሺ݉ ଵ ǡ݉ ଶ ሻቁ

where,N Sim is the geometric type similarity of class names m 1 , m 2 ics is the
information content similarity, tSim is the tuple similarity, and w part , w att are
weights, which express the relevance to be given to the hierarchy ( ics ) and
attributes ( tSim ), respectively, such that w part + w att = 1.
In the above combined approach, the weights w part , w att are defi ned in line
with both the variability and commonality approaches of MDSM recalled
in the previous subsection. Accordingly, we obtain two measures for GSim ,
namely GSim v , GSim c . Note that, in the defi nition above if w att = 0, then GSim
is given by the shared information content between m 1 , and m 2 , as derived
from the partition hierarchy of the geographic classes. On the contrary, if
w part = 0, then GSim corresponds to the tuple similarity.
In the following, an example of such measures is given.

Example 10: Let us consider ics ( Municipality, County ), and tSim ( Municipality,
County ) given in Example 8, and Example 9 and weights (i.e., w part , w att ) of
Example 7. The measures of GSim for Municipality and County, according
to the variability and commonality approaches, are given below:

GSim v ( Municipality, County ) = 1 * (0.530 * 0.5993 + 0.470 * 0.4)
= 0.5056
GSim c ( Municipality, County ) = 1 * (0.401 * 0.5993 + 0.599 * 0.4)

= 0.4799

Experimental analysis

The experimental analysis is focused on the approaches shown in the
previous sections, i.e., Lin, Dice, MDSM and GSim. For the last two
approaches, the variability and commonality measures are given. Note that,
we refer to the weighted Part-of hierarchy shown in Fig. 3, and in particular,
we use the weights computed according to the frequency-based approach
indicated by w f .
In order to evaluate the approaches mentioned above, we refer to the
“right” values, that are established according to Human Judgment (HJ). In
the literature, studies on human judgments was performed by Rubenstein
and Goodenough (Rubenstein and Goodenough 1965) in 1965, and repeated
on a subset consisting of 30 pairs, in 1991 by Miller and Charles (Miller and
Charles 1991). These experiments were not addressed to evaluate similarity