Semantic Similarity based on Weighted Ontology - Geographical Information Systems: Trends and Technologies - page 322

than MDSM and Dice . We also observe that by increasing the occurrences of
parts, the correlation of MDSM v increases, and the one obtained by MDSM c
decreases. In particular, the values obtained by Lin and Dice are invariant,
since in both cases the focus is on attributes and are independent from parts.
However, we note that the correlations achieved by GSim c are higher than
the ones obtained by Lin as the number of occurrences of parts increases
from 11 to 31.

Analysis of weight effects

We analyze the effect of different weighting approaches, discussed
in the Weight Assignment Methods section, on the correlations
achieved by applying the similarity methods. To this end, we use
the weights computed according to the frequency-based approach
( w f , see Fig. 3), the uniform probabilistic approach ( w p , see
Fig. 3), the uniform probabilistic weighted arc approach ( w pa , see Fig. 4),
and the non-uniform probabilistic weighted arc approach ( w pa , see Fig. 5).
The results are shown in Table 8, where we observe that the correlations
of Dice and MDSM are invariant since in such methods the similarity of
concepts are computed on the basis of common and non-common features.
However, as we expect, the correlations of Lin and GSim v , and GSim c , which
are hierarchy-centered methods change and the results obtained by applying
the non-uniform probabilistic weighted arc approach in all three cases are
better than the other two probabilistic based approaches (i.e., uniform
probabilistic and uniform probabilistic weighted arc) and frequency-
based approach. However, the worse results are achieved by applying the
uniform probabilistic approach, indicating that it is not appropriate for
hierarchies similar to that shown in Fig. 3, where most concepts have only
one part. Overall, in the case of GSim , the results by applying the frequency
approach are close to the ones obtained by the uniform and non-uniform
probabilistic weighted arc and the differences of results obtained by these
last two approaches are very low.

Table 8. Correlation of methods by different weighting approaches.

Correlation

Approaches Lin Dice MDSM v MDSM c GSim v GSim c

Frequency 0.7127 0.2972 0.3774 0.3419 0.7214 0.8198

Uniform prob. 0.5469 0.2972 0.3774 0.3419 0.6556 0.7186

Uniform prob. weigh.arc 0.7473 0.2972 0.3774 0.3419 0.8061 0.8784

Non-uniform prob.
weigh.arc

0.7543 0.2972 0.3774 0.3419 0.8319 0.8863