The Impact of Selectional Preference Agreement on Semantic Relational Similarity
Bryan Rink, Sanda M. Harabagiu · 2013
Relational similarity is essential to analogical reasoning. Automatically determining the degree to which a pair of words belongs to a semantic relation (relational similarity) is greatly improved by considering the selectional preferences of the relation. To determine selectional preferences, we induced semantic classes through a Latent Dirichlet Allocation (LDA) method that operates on dependency parse contexts of single words. When assigning relational similarities to pairs of words, if the agreement of selectional preferences is considered alone, a correlation of 0.334 is obtained against the manual ranking outperforming the previously best reported score of 0.229. 1