Distinguish Antonyms in Multiple Viewpoints on Distributed Word Representations

Abstract

Distributed representations of words (word embeddings) are proven effective for measuring word similarity and additive compositionality. However, it is challenging to distinguish between synonyms and antonyms because their proximity in the word embedding space is generally close due to their interchangeability in the context. In this paper, we aim to build a model that can discriminate antonyms in the embedding space without affecting important properties of original word embeddings. Our method is designed to learn multiple hyperplanes for capturing antonymy relations in the embedding space in a supervised manner. Our empirical evaluation demonstrates that we can reasonably distinguish between synonyms and antonyms and reveals that several intriguing issues still remain.