Nonlinear semantic space based on lexical graph

Distributional semantics approaches such as LSA/LSI, HAL, or Eigenwords provide low-dimensional vector space representations of words, so that similarity between words can be measured. However, they come from the statistics of words derived from natural texts where both lexical and contextual similarities are mixed together-synonyms, antonyms, or context-related word pairs are all close to each other. We present a semantic representation based purely on the lexical relations of synonymy and antonymy. Laplacian Embedding transforms a synonymy/antonymy graph into a space where lexically similar words form distinct semantic branches with antonyms placed at opposite sides from the vector space origin. This approach also discovers highly related words that are either novel or missing from the thesaurus. The structures become more distinct after applying Independent Component Analysis (ICA); each dimension clearly represents distinct semantic content, with a word sense as its emergent property. Finally, a comparison of several semantic vector space representations is provided against human evaluated word similarity scores.