Generalizing Representations of Lexical Semantic Relations

Abstract

We propose a new method forunsupervised learning of embeddings forlexical relations in word pairs. The modelis trained on predicting the contexts inwhich a word pair appears together in cor-pora, then generalized to account for newand unseen word pairs. This allows us toovercome the data sparsity issues inherentin existing relation embedding learning se-tups without the need to go back to thecorpora to collect additional data for newpairs.