Finding analogies in semantic networks using the singular value decomposition

Abstract

We present CROSSBRIDGE, an algorithm for finding analogies in large, sparse semantic networks. We treat analogies as comparisons between domains of knowledge. A domain is a small semantic network, i.e., a set of concepts and binary relations between concepts. We treat our knowledge base (the large semantic network) as if it contained many domains of knowledge, then apply dimensionality reduction to find the most salient relation structures among the domains. Relation structures are systems of relations similar to the structures mapped between domains in structure mapping[6]. These structures are effectively n-ary relations formed by combining multiple pairwise relations. The most salient relation structures form the basis of domain space, a space containing all domains of knowledge from the large semantic network. The construction of domain space places analogous domains near each other in domain space. CROSSBRIDGE finds analogies using similarity information from domain space and a heuristic search process. We evaluate our method on ConceptNet[10], a large semantic network of common sense knowledge. We compare our approach with an implementation of structure mapping and show that our algorithm is more efficient and has superior analogy recall.