Efficient and density adaptive edge weight model for measuring semantic similarity

Abstract

The measurement of semantic similarity between concepts is an important research topic in natural language processing. However, previous efforts suffered from the mismatch of the accuracy and efficiency. In this paper, we propose an edge weight model for improving the accuracy of edge-based measures that have an inherent high efficiency. It combines the edge counting model with the information theory and deduces a function of edge weight based on the number of direct hyponyms of the subsumer in the edge. This model doesn't require any additional parameter and can adapt the effect of different densities to edges. Extensive experiments on four test datasets for WordNet and SNOMED-CT demonstrate that the proposed edge weight model can significantly improve the accuracy of various edge-based similarity measures and has a wide coverage over different ontologies. Compared with IC-based measures, our model has a remarkable advantage in efficiency and is comparable to it in accuracy.