Correlation of Ontology-Based Semantic Similarity and Human Judgement for a Domain Specific Fashion Ontology

Abstract

Evaluation of semantic similarity is difficult because semantic similarity values are highly subjective. There are several approaches that compare automatically computed similarities with values assigned by humans for general purpose terms and ontologies that contain general purpose terms. However, ontologies should be as domain specific as possible to capture the maximal amount of semantic knowledge about a domain. To evaluate the semantic knowledge captured by a custom fashion ontology we conducted a survey and crowdsourced similarity values for fashion terms. In this article we compare the manually assigned similarities to those computed automatically with several ontology-based similarity measures. We show that our proposed feature-based measure achieves the highest correlation with human judgement and give some insight into why this kind of similarity measure most resembles human similarity assessments. To evaluate the influence of the ontology on similarities we compare the results achieved with our fashion ontology to similarity values computed using a fragment of DBpedia.