Efficient graph-based document similarity

Abstract

Assessing the relatedness of documents is at the core of many applications such as document retrieval and recommendation. Most similarity approaches operate on word-distribution-based document representations - fast to compute, but problematic when documents differ in language, vocabulary or type, and neglecting the rich relational knowledge available in Knowledge Graphs. In contrast, graph-based document models can leverage valuable knowledge about relations between entities - however, due to expensive graph operations, similarity assessments tend to become infeasible in many applications. This paper presents an efficient semantic similarity approach exploiting explicit hierarchical and transversal relations. We show in our experiments that (i) our similarity measure provides a significantly higher correlation with human notions of document similarity than comparable measures, (ii) this also holds for short documents with few annotations, (iii) document similarity can be calculated efficiently compared to other graph-traversal based approaches.