Getting Creative with Semantic Similarity

This paper investigates how graph-based representations of entities and concepts can be used to infer semantic similarity and relatedness, and, more speculatively, how these can be used to infer novel associations as part of a creative process. We show how personalised PageRank on a co-occurrence graph can obtain competitive scores on a standard semantic similarity task, as well as being used to discover interesting and surprising links between entities. We hypothesise that such links could form the first stage in a creative ideation process.