Exploiting Synonymy And Hypernymy To Learn Efficient Meaning Representations

Word representation learning methods such as word2vec usually associate one vector per word; however, in order to face polysemy problems, it's important to produce distributed representations for each meaning, not for each surface form of a word. In this paper, we propose an extension for the existing AutoExtend model, an auto-encoder architecture that utilises synonymy relations to learn sense representations. We introduce a new layer in the architecture to exploit hypernymy relations predominantly present in existing ontologies. We evaluate the quality of the obtained vectors on word-sense disambiguation tasks and show that the use of the hypernymy relation leads to improvements of 1.2% accuracy on Senseval-3 and 0.8% on Semeval-2007 English lexical sample tasks, compared to the original model.