Antonymy-Synonymy Discrimination through the Repelling Parasiamese Neural Network
Antonymic and synonymic pairs may both occur nearby in word embeddings spaces because they have similar distributional information. Different methods have been used in order to distinguish antonyms from synonyms, making the antonymy-synonymy discrimination a popular NLP task. In this work, we propose the repelling parasiamese neural network, a model which considers a siamese network for synonymy and a parasiamese network for antonymy, both sharing the same base network. Relying in the antagonism between synoymy and antonymy, the model attempts to repell siamese and parasiamese outputs making use of the contrastive loss functions. We experimentally show that the repelling parasiamese network achieves state-of-the-art results on this task.
PDF Abstract