Lexical semantics enhanced neural word embeddings

Current breakthroughs in natural language processing have benefited dramatically from neural models, through which distributional semantics can leverage data representations to facilitate downstream applications. Since embeddings use context prediction on word co-occurrences yield dense vectors, they are inevitably prone capture more semantic association than similarity. To improve vector space models deriving similarity, we post-process deep metric learning, inject lexical-semantic relations, including syn/antonymy and hypo/hypernymy, into a space. We introduce hierarchy-fitting, novel specialization approach modelling similarity nuances inherently stored the IS-A hierarchies. Hierarchy-fitting attains state-of-the-art results common- rare-word benchmark datasets for embeddings. It also incorporates an asymmetric distance function specialize hypernymy's directionality explicitly, it significantly improves vanilla multiple evaluation tasks of detecting hypernymy without negative impacts judgement. The demonstrate efficacy hierarchy-fitting specializing with relations late fusion, potentially expanding its applicability aggregating heterogeneous various knowledge resources learning multimodal spaces.