In Blanchard et al. [5], the authors introduce a collection of intrinsic IC models based on different hypotheses to estimate the concept probabilities. Two of their methods are related to our CondProbUniform and CondProbLeaves IC models, however, like other models, they do not consider the structure axioms defined in our family of well-founded IC models. In order to evaluate their models, the authors compute the correlation between each IC model and one corpus-based IC model. The paper only reports the correlation values for two intrinsic IC models based on the node depths and subsumed BMN673 ratio, reporting respectively an approximated correlation values of 0.3 and 0.5 in WordNet. The comparison method introduced by the authors is very appropriate for evaluating directly the fitting quality of the intrinsic IC models as regards the corpus-based ones, in order to estimate how well an intrinsic IC model approximates a corpus-based model. Indeed, in the light of the findings in Lastra-Díaz and García-Serrano [26] and chorionic villi sampling (CVS) work, we have considered folowing a similar approach into a future deeper study between intrinsic and corpus-based IC models. Despite the proposed IC models deserving to be evaluated, the work does not report any result based on any standard similarity benchmark on the family of IC-based measures, such as we do herein. Therefore, we have discarded to evaluate them in our experiments due to the impossibility of replicating and comparing their results.
↧