Tuning a semantic relatedness algorithm using a multiscale approach

The research presented in this paper builds on previous work that lead to the definition of a family of semantic relatedness algorithms. These algorithms depend on a semantic graph and on a set of weights assigned to each type of arcs in the graph. The current objective of this research is to automatically tune the weights for a given graph in order to increase the proximity quality. The quality of a semantic relatedness method is usually measured against a benchmark data set. The results produced by a method are compared with those on the benchmark using a nonparametric measure of statistical dependence, such as the Spearman’s rank correlation coefficient. The presented methodology works the other way round and uses this correlation coefficient to tune the proximity weights. The tuning process is controlled by a genetic algorithm using the Spearman’s rank correlation coefficient as fitness function. This algorithm has its own set of parameters which also need to be tuned. Bootstrapping is a statistical method for generating samples that is used in this methodology to enable a large number of repetitions of a genetic algorithm, exploring the results of alternative parameter settings. This approach raises several technical challenges due to its computational complexity. This paper provides details on techniques used to speedup the process. The proposed approach was validated with the WordNet 2.1 and the WordSim-353 data set. Several ranges of parameter values were tested and the obtained results are better than the state of the art methods for computing semantic relatedness using the WordNet 2.1, with the advantage of not requiring any domain knowledge of the semantic graph.