Edge dependent pathway scoring for calculating semantic similarity in ConceptNet

Most techniques that calculate the relatedness between two concepts use a semantic network, such asWikipedia, WordNet, or ConceptNet, to find the shortest intermediate pathway between two nodes. These techniques assume that a low number of edges on the shortest pathway indicates conceptual similarity. Although this technique has proven valid in conforming to psychological data, we test the usefulness of additional pathway variables in ConceptNet, such as edge type and user-rated score. Our results show strong evidence for the application of additional pathway variables in calculating semantic similarity. 1 ConceptNet Pathways ConceptNet 3 is one of the largest commonsense semantic networks in existence, relying on its users to make conceptual assertions and collectively vote on the legitimacy of other users’ assertions. ConceptNet is valuable as a semantic resource because it suggests transitive inference between ideas, enabling dissimilar concepts to share a semantic, indirect relationship. Unlike Wikipedia and WordNet, the edges in ConceptNet contain additional semantic information between two concepts. Each edge is assigned a relation type (such as ”Is A” or ”Located At”) and a score that correlates to how well ConceptNet users believe in the validity of the relation [Havasi and Alonso (2007)]. Previous work on calculating semantic relatedness between two concepts ignores these extra edge features, using only the inverse of number of edges on a short path [Rada and Blettner (1989); Wubben and A. (2009)]. Instead of only looking for the shortest pathway from one concept to another we assess all pathways in measuring the semantic similarity of an association. A simple example is shown in Figure 1: two nodes (cat and dog) with two pathways containing varying intermediary nodes and edge types. Figure 1: Transitive links between nodes in ConceptNet may occur through a variety of pathways, some being more appropriate than others.