Antonymy-Synonymy Discrimination through Repelling Parasiamese Neural Networks

Discriminating antonyms and synonyms is a challenging NLP task for which a number of methods have been proposed. A main challenge of this task is that both relations involve semantically similar words, despite the contrariness of antonyms, often occurring nearby in word embedding spaces. In this work, we introduce the repelling parasiamese neural network, a model that considers a parasiamese network for antonyms and a siamese network for synonyms relying on the same encoder and repels their outputs. The resulting model is a fully differentiable neural network for supervised antonymy-synonymy discrimination on pretrained word vectors. We show that this model achieves state-of-the-art results on this task.